Glossary

A model design where weights are restricted to only three discrete values (-1, 0, or 1) instead of continuous floating-point numbers, drastically reducing model size and computation.

A neural network where each weight is represented using only 1 bit of information (in this case, as one of three values: -1, 0, or 1).

An extreme form of quantization where each weight is represented by just a single bit (0 or 1), maximizing compression but reducing model expressiveness.

An extreme form of compression that represents model weights using only 1 bit of information per value, drastically reducing memory use but with significant quality loss.

An extreme form of quantization where each weight in the model is represented by just a single bit (0 or 1), resulting in maximum compression but significant loss of numerical precision.

A data format that represents model weights using 16 bits per number, balancing memory efficiency with numerical accuracy.

A technique for representing and rendering 3D scenes using millions of small Gaussian blobs instead of traditional meshes.

Mathematical shapes (Gaussian distributions) positioned in 3D space used to represent and render 3D scenes efficiently.

Guiding AI model outputs by conditioning on 3D spatial layout information.

Building a complete 3D model of a physical environment from images or sensor data.

Comprehending the three-dimensional structure, objects, and relationships within a physical environment.

A lightweight module that fuses multiple geometric representations with visual features to enhance model understanding.

A specific quantization method that represents model weights using only 4 bits per number instead of the standard 32 bits, dramatically reducing memory usage.

A quantization level where model weights are stored using only 4 bits per value, significantly reducing model size at the cost of some accuracy.

A specific type of quantization that represents model weights using only 4 bits instead of the original 32 bits, enabling very efficient inference on consumer hardware.

A specific compression method that represents model weights using only 5 bits of data per value, enabling efficient local deployment on resource-constrained hardware.

A quantization method that represents model weights using only 6 bits per value, significantly reducing memory requirements compared to standard 32-bit floating-point storage.

A specific quantization method that represents model weights using 6 bits instead of the standard 32 bits, significantly shrinking the model while maintaining reasonable accuracy.

A quantization method that represents model weights using 8 bits instead of the standard 32 bits, reducing memory usage by approximately 75% while maintaining reasonable performance.

A specific quantization method that represents model weights using 8 bits instead of the standard 32 bits, significantly reducing memory requirements.

A testing technique that removes or disables components to measure their impact on system behavior.

An experiment that removes components from a system one at a time to measure how much each contributes to overall performance.

A model variant where safety filters and refusal mechanisms have been removed, allowing it to respond to requests without built-in content restrictions.

A technique that removes or disables a model's built-in safety refusal mechanisms, allowing it to respond to a wider range of requests.

Identifying and highlighting the specific regions in medical images where disease or abnormalities are present.

Visual maps showing which regions of a medical image are abnormal, derived from comparing to historical cases.

A measure of relevance between a query and key that is independent of other keys, allowing explicit rejection of irrelevant keys.

When a system declines to make a prediction or recommendation instead of providing an answer.

When an AI model declines to answer a question rather than providing a response.

A tree representation of code structure that shows how statements and expressions relate to each other.

Generating a summary by creating new sentences that capture key information, rather than selecting existing text.

Realistic power flow calculation that models both active and reactive power using alternating current physics.

The average number of draft tokens verified and accepted per verification round in speculative decoding.

A formal criterion that determines whether a generated design meets specified requirements.

The proportion of draft model's proposed tokens that the target model accepts as correct during speculative decoding.

Restrictions on what operations or resources a system or user is allowed to use.

Designing technology so people with disabilities can use it effectively.

Determining which party in a system is responsible for harms or failures.

A measure of how well an agent performs relative to the computational cost or number of steps it takes.

A communication rate that can be reliably achieved by some protocol, used to characterize channel capacity.

Measurable properties of sound like loudness, pitch, or spectral characteristics used to describe audio signals.

An internal mathematical encoding of sound properties that a model learns to recognize, such as frequency, pitch, and timbre characteristics.

A rule that decides which point to evaluate next by balancing exploration of new areas with exploitation of promising regions.

The problem of correctly associating a specific action command with the correct agent or subject in a scene.

A failure mode where agents make poor action choices that lead to uninformative observations, cascading into reasoning errors.

A short sequence of robot actions predicted together as a unit, executed in a receding-horizon fashion for closed-loop control.

Predicting multiple future robot actions at once instead of one action per step, reducing latency.

The mechanism through which an agent specifies what operations to perform, such as code execution or structured tool calls.

The numerical size of predicted robot movements, which directly correlates with execution speed and distance traveled.

Preventing an RL agent from taking invalid actions by zeroing out their probabilities before selection.

Learned knowledge about how actions and motion typically unfold over time, used to guide policy learning.

Scoring and ordering candidate actions to select the best one for the agent to take next.

The task of identifying and classifying specific actions or activities occurring in video frames.

Pre-training a model to understand and predict action tokens before learning continuous control.

Creating videos where specific physical actions (like forces or robot movements) control what happens in the scene.

Simulating multiple future steps of an environment given a sequence of actions the agent might take.

A continuous measure of how different two agents' behaviors are from each other.

Providing training signals at individual action decisions rather than at the trajectory or task level.

Visual features that specifically encode aspects of a scene important for robot control and manipulation.

Predicting the expected future reward from taking a specific action in a given state.

A joint prediction of both the robot's movement and the expected forces (wrench) that movement will produce during contact.

A learned encoding of an object that explicitly captures how it responds to and changes under different actions.

The portion of a model's total parameters that are actually used to process a given input; in MoE models, this is typically much smaller than the total parameter count.

The probability distribution of neuron outputs at each layer of a network.

The geometric structure formed by a model's internal activations across different inputs, revealing patterns in how the model processes information.

Random variations added to a model's internal computations to test robustness.

A mechanistic interpretability technique that replaces activations during inference to identify which components cause specific behaviors.

The specific configuration of which neurons are active across a network when processing a particular input or task.

The number of bits used to represent intermediate calculations during inference; keeping this higher (like 16-bit) helps preserve model quality when weights are heavily compressed.

Analyzing internal neural network activations to understand what a model has learned or decided at different points.

The process of reducing the precision of intermediate values (activations) computed during model inference, separate from weight quantization.

The space of internal neuron activations in a model, as opposed to parameter or gradient space.

Controlling model behavior by modifying internal activations during inference without changing model weights.

Examining the internal numerical outputs of neural network layers to understand or guide model behavior.

Bypassing AI safety features by manipulating the internal numerical patterns the model uses to process information.

A setting where representation similarity metrics are uninformative because models can share identical activations yet have different task-specific outputs.

How hidden layer representations evolve during training in terms of magnitude and direction.

A framework where agents minimize expected free energy by selecting actions and beliefs that reduce uncertainty about their environment.

A training approach where the model chooses which new examples to learn from rather than using random data.

The number of model parameters that are actually used during inference for a given input, as opposed to the total parameters available.

A model architecture where only a subset of parameters are used for each token, reducing computational cost while maintaining model capacity.

The subset of a model's total parameters that are actually used during inference for each input, as opposed to all parameters being used every time.

An agent's ability to selectively choose what to observe or process based on task needs, rather than passively consuming all input.

A reinforcement learning approach combining a policy-learning actor with a value-estimating critic for improved training stability.

A formal framework for determining whether a specific event caused another specific event by analyzing counterfactual scenarios.

A mathematical constraint ensuring a causal graph has no cycles, enforcing valid causal structures.

Enabling an agent to collaborate with novel partners without prior coordination or training on those specific partners.

A standard optimizer algorithm commonly used to train neural networks by adjusting weights based on gradients.

A small, specialized module added to a model that modifies its output for a specific task without changing the core model weights.

Custom code written to translate data between incompatible formats or interfaces.

Adding lightweight modules to a pre-trained model to enable new capabilities without retraining the entire model.

A method that adjusts how client updates are combined on the server based on their quality or contribution, rather than using fixed weights.

An attack that adjusts its strategy based on feedback from the target system to improve its effectiveness.

Automatically adjusting the amount of computation (depth, steps) based on the difficulty of each input.

Dynamically adjusting the amount of computation (e.g., number of steps) based on problem difficulty.

Dynamically adjusting task difficulty during training based on the model's current performance level.

Educational systems that adjust content difficulty and pacing based on real-time analysis of learner performance and understanding.

Players who adjust their strategies based on observing the history of play rather than using fixed predetermined strategies.

A system that dynamically adjusts parameters (like reward weights) based on the current task or input.

Dynamically selecting or modifying prompts based on the specific input query to optimize model performance.

A quantization approach that adjusts its representation strategy based on the distribution of input values.

Dynamically adjusting how much computational effort a model uses based on problem difficulty.

Intelligently choosing how to round weights during quantization based on loss landscape information rather than simple truncation.

Adjusting computational effort or step sizes based on problem structure rather than using uniform fixed schedules.

Dynamically deciding when to halt refinement based on learned signals rather than using a fixed number of steps.

Selecting test cases or queries based on results from previous tests, allowing dynamic adjustment of evaluation strategy.

Dynamically choosing which test tasks to evaluate based on how informative they are for distinguishing between candidate solutions.

A technique for computing gradients of a system's outputs with respect to inputs by backpropagating through the system dynamics.

A set of variables that must be controlled for in analysis to isolate the true causal effect of a treatment.

A decision rule that determines whether to accept or reject incoming requests to a service based on current load and capacity.

Optimization algorithm that splits problems into smaller parts solved alternately.

Real-world information about how and when workers actually use AI tools in practice.

Computing how much better an action is compared to the baseline, used to guide policy gradient updates.

A method for computing how much better an action is compared to the average, used to guide policy learning in reinforcement learning.

In RL, measures how much better an action is compared to the average action in a given state.

The difference in expected returns between two policies, used to measure which policy is making better decisions in a given state.

Using estimated advantage values to weight training examples, emphasizing transitions that improve over baseline.

A technique that prioritizes reusing past training samples with high advantage estimates instead of always collecting new data.

Intentional manipulation of input data to trick an AI model into making wrong decisions.

Intentional modifications to input data designed to fool or degrade a model's performance.

Systematically testing an agent's reasoning to find logical or evidential violations it may have missed.

A training loop where attack and defense agents compete and improve against each other iteratively.

Testing designed to find weaknesses and edge cases rather than help the system succeed.

Deliberately tricky test cases designed to fool AI models, like plausible wrong answers.

Systematically searching for inputs where a model fails, used here to find materials where ML predictions diverge from ground truth.

A training setup where a generator model creates outputs while a discriminator learns to distinguish them from human examples, providing implicit feedback.

Training where two networks compete—one generates behavior, the other judges if it matches the expert.

A process where one agent intentionally creates challenging test cases to improve another agent's output.

A learned regularizer that encourages realistic, evasive movements by opposing adversarial perturbations during training.

Training approach where a generator and discriminator compete to improve output quality and realism.

Carefully crafted, often imperceptible changes added to images to fool AI models into producing incorrect outputs.

Manipulating a model's output by prepending adversarial text that constrains the model's response generation.

Deliberately crafted inputs designed to trick an LLM into unsafe or unreliable outputs.

A training technique that reweights samples to penalize worst-case calibration errors, improving robustness under distribution shift.

The ability of an AI system to maintain correct behavior even when facing intentionally crafted misleading inputs.

Revising tasks based on solver behavior to find the optimal difficulty zone where learning is most effective.

Deliberately attempting to break or mislead an AI model to identify vulnerabilities and failure modes.

A specially crafted surface pattern designed to fool computer vision systems into making incorrect predictions or decisions.

A defense method that trains models on adversarial examples to improve robustness against attacks.

A defense mechanism that protects models from attacks without requiring exposure to adversarial examples during training.

The disrupted air flow created by a moving object that affects nearby objects, a physical interaction agents must learn to handle.

Evaluating the visual appeal and artistic qualities of images or scenes, such as composition and harmony.

Pre-training a model with automatic musicality labels before preference alignment to establish quality priors and reduce downstream optimization conflicts.

Linking emotional or sentiment states between connected entities in a system.

The emotional tone of text, measured as the degree of negativity, positivity, or neutrality in language.

A person's emotional state, typically measured by dimensions like valence (positive/negative) and arousal (calm/energized).

Predicting which areas or objects in a scene are suitable for a specific action or interaction.

An AI system's ability to act autonomously toward goals in its environment.

Gradually shifting decision-making control from one policy to another during training.

Systematic evaluation of an agent's behavior changes to detect shifts toward undesired traits or capabilities.

The degree to which an agent retains independent decision-making capability without external manipulation.

The framework or system that orchestrates how an AI agent retrieves information, calls tools, and processes results.

Coordinating multiple AI agents to work together on complex tasks.

A measure of how much influence an individual agent has on collective outcomes and system behavior.

Automatically selecting the most suitable agent(s) for a task from available registries using matching and ranking techniques.

A multi-agent system where subagents work together to complete a task while hiding the overall objective.

A specific capability or tool that an AI agent can use to accomplish part of a larger task.

Assigning different agents to focus on specific types of tasks or domains where they perform best.

The sequence of actions and decisions an agent makes while working toward a goal.

A system or interface designed to work with any type of agent without requiring agent-specific customization.

A simulation where independent agents follow simple rules and interact, creating emergent group behavior.

Documentation and files designed for or primarily used by AI agents, such as instruction files and working notes.

Teaching an AI agent how to interact with a specific tool by providing its input language, constraints, and validation rules.

The boundary between what a measurement instrument observes and what it ignores, which is shaped by the instrument's design choices.

A model designed to act autonomously by making decisions, selecting actions, and using tools to accomplish multi-step tasks.

An AI system that can autonomously plan and execute multi-step tasks, making decisions along the way.

The ability of a model to autonomously plan and execute sequences of actions or tool calls to accomplish a goal.

An approach where an AI model autonomously plans and executes multi-step coding tasks, making decisions about which files to modify and how to structure solutions.

The process of autonomous agents searching for and obtaining datasets needed to complete tasks.

Sequential overhead from cascaded perception, reasoning, and tool-calling loops in agentic systems.

Designing and building systems where AI agents autonomously plan, decide, and act toward goals.

Testing an AI system's ability to complete multi-step tasks that require planning, searching, and taking actions.

A system where an AI model acts as an agent that can call tools repeatedly to solve problems step-by-step, rather than answering in a single pass.

The system of tools, prompts, and interaction patterns that guide an AI agent's behavior on a task.

A structured language with explicit control constructs (IF, GOTO, FORALL) that agents use to execute plans deterministically.

An LLM system that can plan and execute multi-step tasks autonomously to achieve goals.

Training approach where an AI agent learns by proposing and refining actions based on execution feedback over multiple iterations.

The study of local interaction dynamics where one agent's output becomes another agent's input under specific protocol conditions.

When an autonomous AI agent's actions diverge from intended operational objectives due to adversarial influence.

AI systems that can process multiple types of input (text, images, etc.) and actively interact with external tools and environments.

Vision systems that extract structured state information needed for an agent to make decisions, not just recognize objects.

A sequence of automated steps where an AI agent selects tasks, generates solutions, and iteratively improves results.

Reasoning through explicit tool calls or code execution that can be interpreted and debugged, but may incur latency from external execution.

An agent's ability to evaluate its own performance and autonomously improve its behavior across multiple attempts.

A training approach where an AI model learns to make sequential decisions and take autonomous actions to complete multi-step tasks, rather than just responding to individual prompts.

Training autonomous agents to make sequential decisions by learning from rewards and reusable experience.

A tool-augmented system where a reasoning model navigates and searches through structured memory using an Observation-Reason-Action loop.

A persistent execution environment that manages an AI agent's planning, execution, and self-correction across multiple steps toward long-horizon goals.

AI systems that iteratively search and synthesize information to solve complex problems autonomously.

An AI agent's ability to detect and fix its own errors by using tools or feedback without human intervention.

An agent's ability to autonomously improve its capabilities by creating and refining its own skills over time.

Evaluation mode where models can iteratively refine outputs or use tools, rather than generating a single response.

Structured approaches where an AI system takes initiative to gather information systematically rather than passively responding to user input.

AI systems that autonomously plan, act, and adapt based on feedback to accomplish multi-step goals in complex environments.

A structured problem where an agent dynamically selects actions and resources to achieve a goal.

Complex tasks where a model acts autonomously to break down goals into steps, use tools, and make decisions to reach an objective.

A visual generation system that autonomously decides when and how to use external tools or search to improve image generation quality.

A system where an AI agent autonomously performs tasks with explicit goals and structured decision-making.

Processes where a model autonomously plans and executes multiple steps or tool calls to accomplish a goal, rather than responding to a single prompt.

Overall performance metric that averages results across all test examples, potentially hiding individual failures.

Combining multiple data points or model outputs into a single summary result.

The study of safely deploying capable but potentially untrusted AI systems through monitoring and oversight.

Computational methods to identify whether text was generated by an AI system versus written by humans.

Interconnected systems where multiple AI components interact through shared data and infrastructure.

Methods to identify whether text was written by an AI system or a human.

Randomness or noise inherent in data that cannot be reduced with more information.

A security operations problem where the volume of alerts exceeds human capacity to investigate them, leading to missed threats.

A classical hierarchical method that solves linear systems by working on coarse and fine levels of approximation.

Systematic errors in AI systems that unfairly disadvantage certain groups of people.

The process of creating or modifying algorithms to solve problems more effectively.

Ensuring AI systems treat different groups equitably without discrimination.

Tendency of AI systems to produce similar outputs or behaviors, either naturally or in response to incentives.

Providing individuals with actionable recommendations to improve their predicted outcomes from a classifier.

The ability to reconstruct and understand the computational process by which a model arrives at its outputs using interpretable intermediate states.

A technique that helps the model understand the order and position of words in long sequences without needing to add extra position information to each word.

A model trained to behave safely and follow human values through techniques like safety filtering and refusal of harmful requests.

The process of training a model to behave safely and according to human values and preferences, which base models typically lack.

Systematic testing to verify that AI systems behave safely and according to intended values in realistic deployment scenarios.

When an AI model appears aligned under monitoring but subverts its goals when unmonitored.

The process of adjusting a model's behavior to make it safer, more helpful, and better aligned with human values.

Safety constraints built into a model during training to prevent it from generating harmful, biased, or inappropriate content.

Additional training applied to a base model to make it behave safely and follow user intentions more reliably.

A vulnerability where a model exploits the alignment process by influencing its own training data to amplify misaligned behaviors.

The degradation of a model's general capabilities that occurs when training it to align with human values.

A unified framework that handles multiple types of image degradation (blur, noise, rain, etc.) with a single model.

A guarantee that higher bids weakly increase an item's chance of being recommended without requiring model retraining.

How well resources (queries, compute) are distributed among available options to maximize overall value.

An early, experimental version of software that is still under development and may have bugs or incomplete features.

AI agents continuously available to help users by understanding context across their entire digital world and proactively anticipating needs.

Errors caused by confusion between similar or overlapping UI elements when determining which one to interact with.

The linear chain of amino acids that makes up a protein, which determines its structure and function.

The linear arrangement of amino acids that make up a protein, written as a string of letters where each letter represents a different amino acid.

Spreading the cost of an expensive computation across multiple uses to reduce per-use cost.

A prediction method that exploits continuous, non-linear relationships in data beyond what linear models capture.

Using domain knowledge from mathematical equations to initialize or guide a machine learning model before training on data.

Automatically identifying and separating different body structures or organs in medical images.

An attention pattern that restricts a model to only attend to ancestor nodes in a tree structure, enabling efficient tree verification.

Choosing a reference model to compare all other models against in pairwise evaluation tasks.

Sparse cross-modal units that connect language concepts to specific scales of 3D geometry, routing semantic information to the most relevant level of detail.

Using a fixed set of reference responses as anchors to rank new outputs and derive rewards efficiently.

Bias where initial information disproportionately influences subsequent decisions.

A bias where an initial piece of information (anchor) disproportionately influences subsequent judgments.

In quantization, the granularity of representable values in high-dimensional space; low resolution causes accuracy loss.

The learning rate applied to the direction of weight updates, separate from the magnitude of weights.

Methods for combining multiple human judgments into a single training signal for the model.

The total amount of labeled data or human feedback available for training.

A structured set of guidelines for labeling data with specific linguistic or semantic information.

A systematic process for labeling data with human-verified information to create training datasets.

The degree to which an annotator (human or model) produces consistent judgments, typically measured by agreement metrics like F1 score.

Variation in how different people label the same content, reflecting genuine differences in perspective rather than labeling error.

The negative electrode in a battery where ions are stored during charging.

Identifying data points or objects that deviate significantly from normal patterns or training data.

Identifying and precisely locating defective or abnormal regions in images or 3D data at the pixel or voxel level.

A declarative programming paradigm for solving combinatorial problems using logical rules and constraints.

The ability of microorganisms to survive exposure to antimicrobial drugs, a major biosecurity and public health concern.

A statistical test that can be checked at any stopping time without inflating error rates.

An open-source software license that allows free use, modification, and distribution of code with minimal restrictions.

A permissive open-source license that allows free use, modification, and distribution of software with minimal restrictions.

A permissive open-source license that allows you to use, modify, and distribute software with minimal restrictions.

A permissive open-source license that allows free use, modification, and distribution of software with minimal restrictions.

An interface that allows developers to send requests to and receive responses from an AI model over the internet.

A programmatic interface that allows developers to send requests to the model and receive responses without running it locally.

The ability to access and use a model programmatically through an application programming interface, allowing developers to integrate it into their applications.

A model that can be used through an application programming interface, allowing developers to integrate it into their applications programmatically.

Access to a model through an application programming interface, allowing developers to integrate the model into their applications and services programmatically.

The ability of a service to work with the same code and commands as another service, making it easy to switch between them.

A method of making an AI model available for use over the internet through standardized web requests, rather than running it locally.

Running a model through a web service interface where you send requests and receive predictions without needing to host the model yourself.

A specification describing how a backend service accepts requests and returns data.

A credential that grants an application or automated agent permission to access services and data on behalf of a user or organization.

A model served through an application programming interface (API) rather than run locally, allowing users to send requests and receive responses over the network.

A model that can only be used through programmatic requests (code) rather than through a web interface or chat application.

A model that is only accessible through a remote service interface rather than being available for download and local installation.

Data structure that records events sequentially without allowing deletions.

Apple's custom-designed processors (like M1, M2, M3) optimized for running machine learning models on Mac computers.

Software tuning that allows a model to run efficiently on Apple's custom processors (like M1, M2, M3) found in Mac computers.

A measure of how close a solution is to the optimal solution, expressed as a ratio.

Mathematical framework for understanding how well functions can represent complex phenomena.

The process by which a model resolves conflicts between different input modalities (e.g., audio vs. text).

A structural limitation in code design that prevents correct solutions even when the agent optimizes parameters within it.

The underlying structural design of a neural network that defines how data flows through layers and components.

A formal system for evaluating arguments based on attack relationships between them, determining which arguments are acceptable.

A mathematical representation of a computation as a directed graph of arithmetic operations.

A model's ability to perform mathematical calculations and solve problems involving numbers and operations.

The intensity or activation level of an emotion, ranging from calm to excited.

A hierarchical representation of connected joints and bones used to model character poses and movements.

A released model, dataset, or tool that is publicly documented and available.

The ability of an agent to produce and return tangible business outputs (documents, code, reports) that meet quality standards.

Organizing and tracking the outputs and intermediate results an agent creates during problem-solving.

Removing spurious patterns (like templates or formatting) that models exploit but don't reflect true underlying phenomena.

A shared space where intermediate results and structured outputs from reasoning steps are stored and accessed by subsequent steps.

A machine learning model inspired by biological neurons that learns patterns from data to make predictions or classifications.

The task of identifying or classifying the artistic style of a work (e.g., Renaissance, Impressionism) using AI.

Training data generated by breaking down queries into multiple aspects and creating complementary evidence examples.

A system that retrieves stored patterns by establishing stable attractors around them, like Hopfield networks.

The ability to find meaningful connections and relationships between different concepts or ideas.

A tool that parses source code into an Abstract Syntax Tree, representing the code's structure as a tree of nodes.

A technique where queries and documents are encoded differently to optimize retrieval performance, rather than treating them identically.

A training approach where the agent has access to additional information or supervision during learning that won't be available when deployed.

A retrieval approach where the query and the documents being searched have different lengths or structures, like matching a short question to long passages.

A statistical property where an estimator's distribution approaches a normal distribution centered at the true parameter as sample size grows.

A neural network approach that correctly captures physics behavior across different scales and parameter regimes.

Large-scale wind and pressure patterns in the atmosphere that drive weather and climate systems.

The set of dictionary atoms (basis vectors) selected to represent a particular signal.

The set of all possible entry points or vulnerabilities in a system that an attacker could exploit.

A safety approach that penalizes actions that significantly reduce the agent's ability to achieve future goals.

A mechanism that lets the model focus on relevant parts of the input when generating each output token.

A method to identify which input tokens most influence specific outputs by analyzing attention weights across network layers.

A measure of how concentrated or distributed attention weights are; lower entropy means the model focuses on fewer tokens.

A parallel attention mechanism within a transformer layer that learns different aspects of input relationships.

Disabling individual attention heads to measure their contribution to a model's output or behavior.

Modifying attention weights in transformer models to change which parts of input the model focuses on.

Visual representations showing which parts of an input a model focuses on when generating each output.

Redirecting where a model's attention focuses by modifying which positions it can attend to during processing.

A technique that allows a model to focus on the most relevant parts of the input when generating each output token.

How a model's attention mechanism divides its focus between different input elements like image and text tokens.

A single forward computation through an attention mechanism that produces weighted outputs from input queries and values.

Aggregating embeddings by learning weighted combinations that emphasize the most relevant slices or features.

A token that attracts excessive attention from the model regardless of its semantic importance.

Tokens that attract disproportionate attention from the model regardless of their semantic relevance to the task.

When floating-point precision limits cause attention weights to become zero, preventing information flow.

Techniques that show which parts of input data a model focuses on during processing.

Numerical values that determine how much a model focuses on different parts of the input when processing information.

A signal measuring how well a reasoning step is supported by the input and previously accepted steps.

An algorithm that identifies sets of states a system will inevitably reach, used in game-theoretic analysis.

A component that refines embeddings by solving for fixed points using implicit differentiation during training.

Deducing personal characteristics like gender, age, or ethnicity from user data without explicit disclosure.

Bias caused by specific visual features (like clothing or age appearance) rather than overall identity differences.

Identifying which input features or model components are responsible for a specific prediction or behavior.

The method used to determine which parts of an input contribute most to a model's prediction.

A technique that identifies which parts of an input (like image regions) are most responsible for a model's predictions or errors.

A technique to identify which neurons are responsible for processing specific types of input by analyzing their contribution to outputs.

A process that diagnoses why an agent failed and updates specific harness components responsible for that failure.

Area Under the Accuracy-Rejection Curve; a metric measuring how well confidence scores distinguish correct from incorrect predictions.

A metric measuring how well a model ranks positive cases higher than negative cases, ranging from 0.5 (random) to 1.0 (perfect).

When a model maintains high classification accuracy (AUC) while its explanations become inconsistent across similar cases.

A system where multiple parties bid to win the right to influence an outcome, with rules ensuring truthful bidding.

A system that generates text descriptions of audio content, allowing LLMs to reason about sound indirectly.

The task of automatically assigning audio clips to predefined categories, such as identifying whether a sound is music, speech, or environmental noise.

A tool that compresses and decompresses audio data to reduce file size while preserving sound quality.

Using an audio sample to guide or control what a generative model produces, rather than using text or other inputs.

A numerical representation (vector) that captures the essential features and meaning of audio data in a compact form that machine learning models can process.

Numerical representations of audio that capture its meaning and characteristics in a form that machine learning models can process.

A neural network component that converts raw audio signals into numerical representations the model can process.

The quality and accuracy of synthesized audio in reproducing natural-sounding speech.

The process of converting compressed audio tokens back into playable audio that closely matches the original sound.

Converting spoken audio content into written text for analysis.

A neural network that processes both audio and text to understand and generate responses about audio content.

A training approach that teaches a model to understand connections between audio sounds and text descriptions by learning from large unlabeled datasets.

A machine learning approach that processes both audio and visual information together to better understand speech and communication.

The ability to simultaneously analyze sound and video streams to understand content where both sight and sound are important.

The ability to jointly process and reason about both sound and video content to understand events, speech, and context more completely than analyzing either alone.

A detailed record of all actions taken by a system, including what was proposed, what was changed, and why—enabling transparency and accountability.

The process of systematically reviewing code or systems to detect errors, vulnerabilities, or malicious modifications.

An LLM's understanding of sound, audio concepts, and acoustic phenomena learned from text-only pre-training.

Using AI to enhance human capabilities and productivity rather than replacing human workers entirely.

Technology and methods that help people with speech disabilities communicate, from word prediction to text-to-speech systems.

An optimization algorithm that solves constrained problems by iteratively updating variables and penalty parameters.

Area Under the Precision-Recall Curve; a metric measuring classifier performance on imbalanced datasets.

Area Under the Receiver Operating Characteristic curve, a metric measuring how well a model ranks correct answers above incorrect ones.

The underlying purpose or goal behind a creator's choices, whether to inform accurately or mislead deliberately.

A method to estimate how credible or expert a source is on a given topic, based on factors like profession and prior contributions.

Determining who wrote a piece of text, including distinguishing human from AI authorship.

Automatically selecting optimal parameter values for a program by testing different configurations.

A feature that predicts and suggests the next tokens or code snippets as a user types, completing partial inputs.

A neural network that compresses data into a smaller representation (encoder) and reconstructs it (decoder).

Automatically translating natural-language mathematical claims into formal theorem statements in proof assistants.

An attack that uses AI models or algorithms to automatically generate, refine, and evaluate malicious prompts at scale.

Using algorithms to automatically measure AI model performance on tasks.

Using an LLM to iteratively improve an agent's harness components based on evaluation feedback within a fixed computational budget.

Techniques that automatically generate patches to fix bugs or vulnerabilities in source code.

Systems that automatically evaluate student code submissions for correctness and understanding.

Automatically identifying and outlining specific regions or structures in an image without manual labeling.

Using computational methods to automatically check whether a proposed solution is correct without human review.

Computing gradients of functions by decomposing them into elementary operations and applying the chain rule.

Automatically reassigning labels or categories to training data to improve model learning.

Technology that converts spoken audio into written text automatically.

The tendency for humans to over-rely on or trust automated systems, even when they make mistakes.

Conceptual boundaries beyond which replacing humans with algorithms is neither possible nor desirable.

Automated tools that search over multiple model architectures and hyperparameters to find the best classifier without manual tuning.

An AI system that can independently perceive its environment, make decisions, and take actions to accomplish goals without constant human direction.

AI systems that can independently plan and execute multi-step tasks without human intervention at each step.

AI agents that generate, execute, validate, and repair code artifacts without human intervention in the loop.

A system where AI automatically evaluates and improves itself without human intervention in the loop.

A robot independently practicing tasks and generating training data without human guidance or intervention.

A robot's ability to learn new manipulation skills through self-directed practice without human demonstrations.

A structured model defining levels of system independence, from human-controlled to fully autonomous decision-making.

A range of control levels from fully human-controlled to fully autonomous AI, with hybrid modes in between.

A model that generates text one token at a time by predicting the next word based on all previous words in the sequence.

Decoding instability where sequential token generation fails due to high token fertility and error accumulation.

A neural network component that scores candidate tokens by considering previously selected tokens in sequence.

The standard method most language models use to generate text by predicting one token (word piece) at a time, left to right, where each prediction depends on all previous tokens.

Generating sequences one token at a time, where each new token depends only on previous tokens.

A text generation approach where the model predicts one word at a time, using all previously generated words to inform the next prediction.

Generating graph structures one node/edge at a time, where each step depends on previous decisions.

Generating images sequentially, one token or element at a time, where each prediction depends on previous outputs.

A model that generates text by predicting one word or token at a time, using only the words that came before it.

A model that predicts the next item in a sequence based on all previous items, one step at a time.

Language models that generate text one token (word piece) at a time, where each new token depends on all previously generated tokens.

Generating predictions sequentially where each prediction depends on previous predictions, causing errors to compound over time.

A single neural network that generates outputs one token at a time across all modalities using the same architecture.

A generative model that creates videos frame-by-frame sequentially, where each new frame depends on previously generated frames.

Generating a sequence of zoom-level decisions one at a time, where each decision depends on previous ones, to progressively narrow down a location.

An advanced quantization method that automatically optimizes how model weights are rounded to lower precision, preserving accuracy better than simpler quantization approaches.

An automated quantization method that intelligently rounds weights to lower precision while minimizing the loss in model performance.

Intel's automated quantization method that intelligently rounds model weights to lower precision while minimizing accuracy loss.

Additional training signals beyond the primary task that help models learn richer representations.

Internal business functions like finance, HR, and administration that support core business activities.

The core language model architecture that forms the foundation of a larger system, in this case Llama 3.

The core neural network structure that a model is built upon, which in this case is Llama 3.

A core neural network component that extracts features from input data, typically used as a foundation for larger systems rather than standalone.

Brief verbal responses like 'mm-hmm' or 'yeah' that show engagement without taking a full conversational turn.

A security attack where hidden malicious behavior is embedded in a model to trigger on specific inputs.

A training method for recurrent networks that computes gradients by unrolling the network across time steps.

Testing a model on historical data to evaluate how it would have performed.

Reverting to an earlier decision point when an approach fails, rather than trying to fix errors at the current level.

A sampling strategy that allows reversing previous decisions (remasking tokens) to escape low-reward regions and find better solutions.

How learning new tasks affects performance on previously learned tasks.

The assumption that meaning emerges from simple addition of individual feature meanings without interaction.

An interpretable EEG analysis method that learns a dictionary of recurring waveform patterns and represents signals as sequences of these learned patterns.

A fairness metric that averages accuracy across classes, preventing high scores when one class dominates predictions.

Learning setting where you only observe the outcome of your chosen action, not all alternatives.

A control strategy that switches between extreme values (on/off) rather than using continuous intermediate values.

A neural network architecture that combines an encoder (which reads text) and a decoder (which generates text), commonly used for tasks like summarization and text generation.

A neural network design that combines an encoder (for understanding text) and decoder (for generating text) to learn meaningful representations.

The foundational neural network design that a model is built upon; inheriting from a base architecture means the model follows the same core structure and design principles.

A foundational AI model trained on raw text data without additional fine-tuning for specific tasks or instructions.

The individual weak models (like decision trees or neural networks) that are combined in an ensemble method.

A pretrained model that completes text patterns but hasn't been trained to follow instructions, serving as a starting point for customization through fine-tuning.

A smaller version of a model architecture that prioritizes speed and lower memory usage over maximum performance, making it suitable for resource-constrained environments.

A model trained only on raw text prediction without additional instruction-following training, so it completes text continuations rather than answering questions or following commands.

A language model trained on raw text data without additional instruction tuning, so it completes text patterns rather than following specific user instructions.

A simple reference model used to compare performance against more complex models or to establish a minimum expected behavior.

An existing, functional control policy used as a starting point or reference for training improvements.

A region in a model's state space where inputs converge to the same output or memory.

Simple mathematical shapes (like sine waves or Gaussians) combined to represent complex signals.

Learning from multiple independent training batches rather than continuously updating from a live environment.

Systematic differences in data caused by processing samples in separate groups.

A stable outcome where each agent's strategy is optimal given their private information and beliefs about others' strategies.

A resampling method that estimates uncertainty by repeatedly reweighting data and refitting models.

Framework for making optimal decisions by combining probability distributions with utility functions.

Probabilistic methods that estimate hidden states by recursively updating beliefs based on observations and a system model.

A probability combination method that merges confidence signals using Bayesian principles to create a single aggregated score.

A mechanism where participants are motivated to tell the truth about their preferences, given what they know.

A statistical method that updates beliefs about unknown values using observed data and prior knowledge.

A semi-structured representation combining numerical probabilities with natural-language evidence summaries, updated iteratively by an LLM.

Neural networks that model uncertainty by treating weights as probability distributions rather than fixed values.

A method that uses probability to intelligently update and improve a system based on past results.

A framework for analyzing how information disclosure strategically influences decision-makers' choices.

The probability distribution of unknowns after incorporating observed data using Bayes' rule.

Binary Cross-Entropy loss, a training objective commonly used for relevance scoring tasks where the model learns to predict whether a query-document pair is relevant or not.

A decoding algorithm that keeps the top-k most likely candidate sequences at each step, balancing quality and computational cost.

An algorithm that approximates a sum over exponentially many terms by keeping only the highest-probability terms.

Automatically detecting the regular pulse or rhythm in music to align timing across different representations.

A mathematical model describing how honeybee swarms reach consensus on nest sites through recruitment and inhibition.

Training a policy to imitate expert demonstrations by supervised learning on state-action pairs.

Techniques designed to trigger specific model behaviors, especially rare or problematic ones, for testing purposes.

A large-scale model trained on diverse behavioral data to learn generalizable control policies for embodied agents.

Learned vector representations that capture and control an agent's behavioral characteristics like driving style.

A hierarchical decision-making structure that combines simple rules and conditions to control complex agent behavior.

A regularization technique that keeps a fine-tuned policy close to its original pretrained distribution.

A systematic evaluation of a model's outputs and preferences across different contexts and framings to detect biases.

The sudden drop in a model's ability to apply a learned rule, despite that rule remaining present in training data.

An RL agent's ability to produce multiple different strategies or outputs rather than converging to a single deterministic policy.

How human responses to interventions create secondary effects that influence system outcomes.

A metric measuring how accurately a simulator reproduces a student's actual responses and learning behavior.

Analyzing actual observed behavior patterns rather than theoretical assumptions about how systems should work.

Drawing conclusions about social or psychological states from observable actions and interactions.

When an AI system's behavior deviates from its intended operational goals due to adversarial manipulation.

A controlled experiment or stimulus designed to elicit informative behavior from an agent.

Using models to predict how people will act in specific situations or respond to choices.

A structured description of what a program should do, derived from documentation and testing.

Failure mode where task-relevant information gets buried in context or pushed beyond the context window, preventing it from influencing agent decisions.

Initializing a policy by learning to imitate past user actions from historical data.

The magnitude of shift in a user's conviction about information after receiving a correction.

A cognitive trap where retrieved memories cause an LLM to adopt incorrect beliefs or assumptions that contradict current task context.

The space of probability distributions representing a robot's uncertainty about unobservable factors like human preferences or goals.

A representation of what an AI system or person currently believes to be true about a situation.

Aligning the internal knowledge representations and understanding between multiple agents so they interpret messages correctly.

The process of revising an agent's beliefs when new information about the world becomes available.

A framework modeling agent behavior through beliefs (what they know), desires (what they want), and intentions (what they commit to do).

A recursive equation that relates the value of a state to the values of successor states in dynamic programming.

A mathematical operator that updates value estimates based on immediate rewards and future value predictions.

A standardized test suite used to measure and compare model performance on specific tasks.

A standardized set of test problems used to measure and compare the performance of different algorithms or models.

The infrastructure that runs evaluation tests and measures agent performance against predefined tasks.

The degree to which a synthetic benchmark's structural properties match real-world data distributions.

Using a smaller subset of test cases instead of the full benchmark to speed up evaluation.

Comparing model safety when no labeled benchmark exists for the specific language, domain, or regulatory context.

A phenomenon where a model fits training data perfectly but still generalizes well to unseen data.

A neural network output that predicts binary (0/1) values independently for each bit in an image code.

A foundational neural network architecture designed to understand the meaning of words in context by learning from large amounts of text.

A transformer-based model design that reads text in both directions simultaneously to understand context, widely used as a foundation for language understanding tasks.

A neural network model that reads text and converts it into numerical vector representations that capture the meaning of words and sentences.

A transformer-based neural network architecture designed to understand text by learning bidirectional context, commonly used as a foundation for natural language understanding tasks.

A model architecture that uses the same foundational design as BERT, which learns bidirectional context by reading text in both directions simultaneously.

A model built on BERT, a foundational architecture that learns bidirectional text representations and is commonly adapted for specific tasks like spell-checking.

A neural network design based on the BERT model that uses transformer layers to understand relationships between words in text by looking at context from all directions.

A transformer-based model architecture that reads text bidirectionally to understand context and produce meaningful representations of words and sentences.

A heavily compressed version of the BERT language model with far fewer parameters, designed for fast inference on resource-constrained devices.

Topic modeling method that uses transformer embeddings to identify and label topics in text data.

The optimal strategy against a known or estimated opponent strategy that maximizes expected payoff.

Generating K samples from a model and selecting the best one according to some criterion.

A decoding strategy that generates N candidate responses and selects the one ranked highest by a reward model.

An early version of software that is still being tested and refined, meaning it may have bugs or incomplete features but is available for broader evaluation.

A topological property that counts connected components and holes in a structure, used here to enforce vessel connectivity.

A 16-bit floating-point format that balances precision and memory efficiency, commonly used for training and deploying large language models.

A 16-bit floating-point format (Brain Float 16) that balances precision and memory efficiency, commonly used for storing and running large language models.

A 16-bit numerical format that balances memory efficiency with numerical stability, using fewer bits than standard 32-bit floats while maintaining training and inference quality.

A 16-bit floating-point format that preserves numerical precision similar to full 32-bit precision while using half the memory, making large models faster and cheaper to run.

A compressed number format that uses fewer bits to represent weights, reducing memory usage and speeding up computation while maintaining reasonable accuracy for most tasks.

A family of embedding and ranking models designed for dense retrieval tasks, optimized for finding and scoring document relevance.

A model architecture that encodes two pieces of text separately into comparable vector representations, allowing efficient comparison of their semantic similarity.

An optimization approach with two nested loops: an inner loop optimizing fast weights and an outer loop optimizing the main model parameters.

Identifying systematic errors or unfair patterns in AI model predictions across different groups or categories.

Systematic testing of AI models to identify and measure discriminatory patterns against specific groups.

Techniques to reduce discriminatory outcomes in machine learning models through data or algorithm modifications.

The spread of systematic evaluation errors from one agent to others through their interactions.

A mathematical guarantee that limits how much bias can affect a model's decisions, even if the bias source is unknown.

Parts of a model where social biases are most likely to emerge or be encoded in the computations.

Breaking down prediction error into bias (systematic error) and variance (sensitivity to training data).

The fundamental tension between model simplicity (high bias, low variance) and complexity (low bias, high variance).

Bounded-Input Bounded-Output stability ensures that finite inputs produce finite outputs in a control system.

An inference technique that adjusts which items are generated based on real-time bid values, steering recommendations toward higher-value items.

A mechanism that allows the model to look at context both before and after each word when understanding text, rather than just looking forward.

The ability to understand relationships between words by looking at both the words that come before and after a given word.

The ability to generate text or code by considering context from both directions (before and after a gap), rather than only generating left-to-right.

A neural network architecture that encodes two separate pieces of text independently and compares them to measure semantic similarity, commonly used for matching and retrieval tasks.

A large coefficient used in MILP formulations to enforce logical constraints; larger values make the relaxation weaker and solving slower.

A transformer-based model architecture designed to handle very long text sequences efficiently by using sparse attention patterns instead of processing every word pair.

An optimization framework with two hierarchical levels where upper-level decisions constrain lower-level optimization problems.

A factorization where value and policy functions are expressed as products of goal-conditioned coefficients and learned basis functions.

A mathematical optimization problem where the objective involves products of two variables, solvable with standard optimization tools.

A model trained to understand and generate text in two languages, in this case Japanese and English.

A language model trained to understand and generate text in two languages with comparable fluency.

A mapping of words between two languages, showing which words in one language correspond to words in another.

A recurrent neural network that processes text in both forward and backward directions to capture context from both sides of each word.

Synchronized control of two robot arms working together to accomplish a task, requiring precise timing and spatial awareness.

Robot control using two arms simultaneously to perform coordinated tasks.

A model that processes two different types of input (in this case, code and natural language) and converts them into a shared representation space.

A model that separates data into exactly two classes, often used as a building block for more complex classifiers.

A decision mechanism where neurons act as on/off switches to direct data through different computational paths.

The challenge of determining which visual features belong to the same object in a scene.

Inferring physical quantities like joint forces, muscle activation, and body loads from skeletal motion.

A large collection of medical and scientific texts (like research papers and journals) used to train the model on domain-specific language and concepts.

Natural language processing techniques applied specifically to medical and biological text, such as extracting drug names or identifying disease mentions from research papers.

The ability to understand and work with scientific concepts in biology and medicine, such as drug interactions and molecular structures.

Written content from medical and life sciences domains, including clinical notes, research papers, and healthcare documentation.

Specialized medical and scientific terms and concepts that the model has learned to understand from training on medical literature.

Protecting against misuse of biological research and AI in harmful ways.

Electrical or physical signals produced by the body, such as heart rhythms or brain waves.

A graph structure with two distinct sets of nodes where edges only connect nodes from different sets.

A top-down 2D representation of a 3D scene, commonly used in autonomous driving for unified sensor fusion.

A top-down 2D representation of a 3D scene, showing spatial layout as if viewed from above.

The mathematical space of all doubly stochastic matrices; parameterizing this space exactly is the core challenge this paper addresses.

The number of bits used to represent each number in a model; lower bit depths (like 3-bit) create smaller files but may lose some accuracy compared to higher bit depths.

The number of bits used to represent each number in a model; lower bit precision (like 3-bit) means smaller file size but potentially less accurate calculations.

Encoding individual bits as separate tokens in a model's vocabulary to preserve fine-grained binary structure.

The number of bits used to represent each number in a model; lower bit-widths (like 6-bit) use less memory but may reduce precision compared to higher bit-widths.

Automatically choosing the optimal number of bits for quantizing different parts of a model based on their importance.

The amount of data used to represent audio per second, measured in kilobits per second (kbps); lower bitrates mean smaller files but potentially lower audio quality.

A compression metric measuring how many bits are needed to encode each byte of text.

A measure of how many bits of data are used to store each weight in a neural network; lower values mean more compression but potentially lower quality.

Optimizing a system where you can only observe outputs for given inputs, without access to internal structure or gradients.

Evaluating a system's behavior by observing inputs and outputs without access to internal model structure or weights.

A bioinformatics algorithm that finds similar DNA or protein sequences in a database to identify evolutionary relationships.

The scope of impact or damage caused by a change, such as how many services are affected when a patch breaks a dependency.

A metric that measures translation quality by comparing n-gram overlap between machine-generated and reference translations.

A learned evaluation metric for text generation that uses a pre-trained model to assess translation quality beyond surface-level n-gram matching.

A measure of uncertainty in an agent's decision-making at a given state—how much of the decision space lacks statistical support from training data.

Assessment where evaluators don't know which version or source produced the item being judged.

An attention technique that processes groups of items together to improve efficiency and capture relationships between them.

A diffusion-based generation technique that refines output in structured blocks or segments rather than token-by-token, enabling more efficient parallel generation.

A quantization format that groups values into blocks and uses a shared exponent (scale) for each block to reduce precision while maintaining accuracy.

Internal vector representations produced by a state space model's processing blocks that encode information about token sequences.

Scaling factors computed for groups of values in low-precision formats to maintain numerical accuracy.

Self-contained subproblems with defined interfaces used as building blocks for larger problem construction.

An attention mechanism that only allows tokens to attend to previous tokens in time, enforcing temporal ordering.

A language model that generates multiple tokens in parallel using diffusion, then refines them iteratively.

A quantization method that divides values into groups and applies a shared scale factor to each group.

A method to model dependencies between data streams sequentially in blocks, reducing computational complexity while capturing inter-stream relationships.

A decoding strategy where multiple tokens are generated in parallel blocks rather than one token at a time.

Security operations focused on defending systems, detecting threats, and responding to incidents.

Creating a high-level plan of formally stated definitions, lemmas, and their dependencies before attempting to prove them.

Iteratively updating the global proof plan when individual lemmas fail, rather than backtracking within a single proof path.

A ranking function that scores document relevance based on term frequency and document length normalization.

Movement commands relative to the drone's own orientation, rather than a fixed world direction.

A technique that repeatedly samples data and selects features or solutions that consistently perform well across samples.

The quantum mechanics principle that measurement probabilities equal the squared magnitude of quantum amplitudes.

A quantum state of matter where particles occupy the same quantum state at very low temperatures.

Mechanisms that prevent an LLM from crossing defined limits in reasoning or behavior.

Flagging predictions for instances near decision boundaries where the model is less confident.

A lightweight policy network that adapts a frozen pre-trained model to new tasks or objects with minimal additional training.

A constraint where quantum gates can only act on nearby qubits, limiting which qubits can interact directly.

A guaranteed maximum time delay for system operations, critical for safety-critical real-time control.

A well-defined task with clear success criteria, constraints, and escalation points where human operators can intervene or redirect the agent.

A rectangular coordinate set that marks the exact location and size of detected text or objects within an image.

Byte-pair encoding algorithm that iteratively combines frequent character pairs into new tokens.

A statistical model that ranks items based on pairwise comparison outcomes, commonly used for leaderboards.

Decoding visual information from brain signals to match or retrieve images based on neural activity.

Using AI to enhance the exploratory ideation phase of research rather than automating solution design.

An optimization algorithm that systematically explores a solution space by dividing it into subproblems and pruning branches that cannot contain better solutions.

The average number of possible moves available at each decision point in a game.

A metric combining token uncertainty and policy likelihood gains to identify high-value decision points for exploration.

A marker in code where a debugger pauses execution so you can inspect the program state.

A generalized distance measure defined by a convex function, used to replace Euclidean geometry in optimization algorithms.

A metric measuring forecast accuracy that compares a model's predictions to a baseline (like random guessing).

A situation where the underlying physics has symmetry, but observations reveal a preferred direction or asymmetry due to measurement constraints.

A random trajectory representing continuous random motion, used to model noise in stochastic processes.

A reinforcement learning technique that constrains model outputs to stay within a token budget, reducing response length while maintaining accuracy.

Constraining an agent's resource usage (compute, API calls, time) while it searches for solutions.

Distributing limited resources (like practice time) across options to maximize total value while respecting constraints.

Breaking text into individual bytes (raw character codes) rather than words or subwords, which allows the model to handle any text without a predefined vocabulary.

The ability of a system to tolerate and correctly function despite some participants behaving arbitrarily or maliciously.

The ability of a system to function correctly even when some participants behave maliciously or unpredictably.

A formal language that characterizes which computational tasks transformers can solve with length generalization.

A strategy for deciding which cached items to remove when cache capacity is full (e.g., LRU, LFU, FIFO).

Loss of cached computation when prompt structure changes, requiring recomputation of affected tokens.

A component that aligns and calibrates KV caches from different sources into a unified format.

Normalized probability scores that can be safely summed across sources to make consistent decisions.

Model predictions that are numerically meaningful and correspond to real-world values (like aesthetic scores), rather than abstract relative rankings.

Adjusting a model's predictions using held-out data to correct for systematic biases or distribution differences.

Choosing a representative sample of tasks to evaluate in detail, then using results to estimate performance on the full benchmark.

The time required to adjust quantization parameters for a model using a small dataset before deployment.

A method that works without requiring a separate tuning phase on representative data before deployment.

Determining the position and orientation of a camera in 3D space relative to a scene.

Robot actions expressed relative to the camera's local coordinate frame rather than the robot's base frame.

A statistical technique that finds the strongest correlations between two sets of variables by discovering shared patterns.

Training process designed to extract or develop specific abilities from a model, like reasoning or tool use.

Inferring what actions or tasks a partner can reliably execute, even when their true abilities are hidden or uncertain.

The boundary between tasks an agent can reliably solve and tasks it consistently fails, where performance is most uncertain.

A test or criterion designed to measure whether a model possesses a specific skill or capability.

The set of all arrival rate vectors that a network can sustain without queues growing unboundedly.

How the number of storable associations grows with the size of the memory matrix or system parameters.

Forecasts of future returns, volatility, and correlations for different asset classes used to guide investment decisions.

Neural networks with capsule units that learn hierarchical relationships and spatial properties better than traditional convolutional layers.

A standard algorithm for building decision trees using binary splits and impurity-based criteria.

A system where one model's output feeds directly into another model as input, like ASR output going to translation.

Sequential observations of how outcomes spread or evolve through a system, like infection waves or adoption patterns.

A mechanism that sequentially combines information from multiple sources (global context, object details, skill knowledge) to guide model decisions.

A sequential approach to combining modalities in stages, where related modalities integrate first before adding heterogeneous ones.

Sequential processing where output from one stage feeds into the next.

A strategy where each model focuses on progressively smaller regions of interest to improve accuracy.

A multi-stage process that progressively assigns incidents to the correct business team or service owner.

The model's ability to distinguish between uppercase and lowercase letters as meaningful differences, treating 'Москва' and 'москва' as separate tokens with different meanings.

A model that treats uppercase and lowercase letters as identical, so 'Apple' and 'apple' are processed the same way.

The model treats uppercase and lowercase letters as distinct, allowing it to recognize proper nouns and maintain capitalization distinctions.

Text processing that preserves the distinction between uppercase and lowercase letters, treating 'Apple' and 'apple' as different tokens.

The model's ability to distinguish between uppercase and lowercase letters, making it sensitive to proper nouns and capitalization patterns that carry meaning.

Examining the timing and magnitude of money movements to assess financial health and repayment ability.

A mathematical operator that characterizes the properties of a symmetry group; used here to encode nuclear symmetries as neural network features.

When a model loses its original knowledge while learning a new task, like overwriting old skills.

Assigning responsibility or credit to specific inputs for producing an observed outcome based on causal relationships.

Whether generated content accurately represents cause-and-effect relationships and physical/logical dependencies.

A model that learns causal relationships between variables and can answer observational, interventional, and counterfactual questions.

A model that learns how data is generated by capturing cause-and-effect relationships, not just correlations.

The ability to determine true cause-and-effect relationships from data, typically guaranteed by randomization.

A measure of how much a component directly influences the final output, not just correlates with it.

Determining whether a treatment actually caused an outcome, not just whether they're correlated.

Deliberately modifying a model's internal features to measure their direct effect on outputs.

A model that predicts the next word in a sequence by only looking at previous words, not future ones, making it suitable for text generation.

A training approach where the model predicts the next word based only on previous words, commonly used for text generation tasks.

Identifying which specific steps in a trace directly caused a failure, removing irrelevant steps.

Understanding cause-and-effect relationships rather than just statistical correlations in data.

Learning representations that capture causal relationships and mechanisms rather than just correlations.

An internal model state that causally influences the final output and can be verified as actually having occurred.

A machine learning method that estimates personalized treatment effects from survival data using tree-based models.

A mechanistic interpretability technique that identifies which parts of a neural network are responsible for specific outputs.

A permissive open-source license that allows anyone to use, modify, and distribute the model as long as they give credit to the original creator.

A Creative Commons license that allows free use and modification of the model for non-commercial purposes only, with attribution required.

A problem-solving technique that starts with a simplified version of a problem and refines it when solutions fail.

A statistical phenomenon where most scores cluster near the maximum possible value, reducing the ability to distinguish between different quality levels.

When a benchmark becomes too easy and models achieve near-perfect scores, making it impossible to compare their true abilities.

A topological structure built from cells of varying dimensions (vertices, edges, faces, volumes) that generalizes graphs and meshes.

Observations where the true value is unknown but bounded, common in survival analysis when subjects are still alive at study end.

Training an AI model to refuse or provide false information about certain topics.

A geometric measure comparing internal representations across models to detect structural differences.

Metric that measures structural similarity between representations by comparing their kernel matrices.

A statistical principle stating that the average of many independent samples approaches a normal distribution.

Access control where permissions are cryptographically certified and must be validated before execution, not just during planning.

A hyperparameter controlling how strongly a generative model adheres to its input prompt versus generating freely.

A reasoning technique where an AI model shows its step-by-step thinking process before arriving at a final answer, making its logic transparent and verifiable.

A technique where a model works through a problem step by step, showing its reasoning process before arriving at a final answer.

RL approach that enforces probabilistic constraints on outcomes rather than hard guarantees.

The maximum rate at which information can be reliably transmitted through a noisy communication channel.

A quantum circuit composed of quantum channels (operations that map quantum states to quantum states) rather than unitary gates alone.

A quantum model that includes noise or stochasticity, representing realistic quantum operations beyond ideal unitary transformations.

Raw wireless signal data that describes how a Wi-Fi signal changes as it travels through space and bounces off objects.

Learning embeddings that understand relationships between different sensor channels using their textual descriptions.

A learnable operation that scales and shifts different feature channels independently in a neural network.

Applying different forgetting rates to different feature channels in a neural network, allowing selective memory retention.

Adjusting each color channel independently based on degradation type to improve restoration quality.

A quantization approach that applies different compression settings to different channels or groups within a model layer, helping preserve quality better than applying the same compression uniformly.

Systems where small changes in initial conditions lead to drastically different outcomes, making long-term prediction extremely difficult.

The ability of a model to maintain a character's voice, personality, and backstory throughout a conversation without contradicting itself.

A metric measuring the percentage of characters incorrectly recognized by an OCR system.

A model's ability to maintain distinct, consistent personality and speech patterns for different characters within a story.

Processing text one character at a time rather than by words, which is useful for catching individual character errors in languages like Chinese.

The ability to extract information from visual charts and perform logical reasoning tasks based on what the chart displays.

A language model specifically trained to have natural back-and-forth conversations with users rather than just completing text.

A model specifically trained and tuned to excel at conversational interactions rather than other tasks like analysis or reasoning.

A model optimized through training to excel at multi-turn conversations and dialogue, rather than single-turn text completion.

A structured message format that organizes conversational exchanges into clearly marked roles (user, assistant, system) to help models understand dialogue context.

Assessing outputs against structured lists of required elements or claims that should be present.

A saved snapshot of a model's weights and state at a specific point during training, allowing training to resume or the model to be evaluated at that stage.

Leveraging intermediate model snapshots from training as separate experts rather than discarding them.

Saved snapshots of a model at different stages of training, allowing researchers to study how the model's behavior changes as it learns.

Breaking long sequences into smaller segments and processing them sequentially while maintaining state between chunks.

A training technique that divides sequences into chunks and processes them in parallel to speed up computation.

Processing sequences in fixed-size blocks rather than token-by-token to improve computational efficiency.

The process of breaking large documents into smaller pieces so a model with a limited context window can process them separately.

Reducing the number of qubits and quantum gates needed to represent and process information in quantum systems.

The process by which neural networks develop specialized, interpretable computational patterns to solve specific tasks.

Identifying and isolating specific subsets of neural network components that perform a particular computation.

The process of determining the specific values (like resistance, capacitance) for each component in a circuit design.

Expressing an idea indirectly or using many words instead of the direct term—what models do when their preferred tokens are blocked.

A network representing which papers cite which other papers, showing the flow of scientific influence and prior work.

Whether an AI model's claims are supported by and properly attributed to specific sources.

A graph structure showing how research papers reference each other, used to understand relationships and influence between scientific works.

An advantage in receiving more citations for work published in certain venues compared to others.

The ability to identify, reference, and maintain accurate attribution to the sources used when generating a response.

Converting unstructured text into atomic, typed statements with explicit evidence locations and source attribution.

The number of insurance claims expected per policy or geographic area over a time period.

Checking whether factual statements in generated text are supported by evidence or ground truth.

A technique that generates visual heatmaps showing which image regions a neural network uses to make predictions.

When training data has unequal numbers of examples across categories, with some classes having far fewer samples than others.

Learning to recognize new object classes over time while maintaining performance on previously seen classes.

Generating complete, structured classes with multiple methods and internal dependencies from a specification.

A loss function that penalizes misclassification of rare classes more heavily, useful when training data is imbalanced.

A statistical framework for designing and validating tests that measure psychological constructs reliably.

A machine learning model trained to assign input data into predefined categories or labels.

Steering diffusion generation toward a target class using a noise-conditioned classifier during sampling.

A two-stage approach where a model first learns representations, then retrains just the final classification layer on balanced data.

A technique that steers diffusion models toward desired outputs by comparing conditional and unconditional predictions.

How well an LLM's medical communication matches established clinical standards and physician practices.

The study of moral principles and values that guide medical decision-making and patient care.

Converting clinical information (diagnoses, medications, procedures) into discrete tokens that a model can process.

Identifying and documenting specific textual passages that support diagnostic conclusions in medical records.

Using historical patient data to predict future health outcomes, disease progression, or treatment responses.

The ability to accurately interpret and reason about medical terminology, patient symptoms, and healthcare documentation.

Natural language processing applied to medical and healthcare text, such as extracting diagnoses or findings from doctor's notes and radiology reports.

The ability to analyze medical information, connect symptoms to conditions, and make logical healthcare decisions based on evidence.

A numerical score derived from patient vital signs and lab values that predicts risk of adverse outcomes like death or organ failure.

The process of confirming that an AI system's outputs meet clinical standards and are safe for use in healthcare.

A model trained on image-text pairs to create shared vector representations for both images and text.

A neural network design that learns to match images and text by training them to have similar representations, enabling tasks like image search and visual understanding.

Rapidly adjusting a model to new tasks using direct mathematical solutions rather than iterative training.

A mathematical formula that directly computes an answer without iterative learning or optimization.

A system that continuously adjusts its behavior based on feedback from its actions and outcomes.

A feedback control rule that adjusts actions based on observed state to correct deviations from a nominal plan.

An assessment method where an agent receives feedback and can adjust its behavior iteratively, simulating real-world interaction.

A feedback system where outputs are evaluated and results guide the next iteration of improvements.

A control strategy where the robot observes its current state and adjusts actions based on feedback, rather than executing a fixed sequence.

Generating a sequence of predicted observations by repeatedly applying a world model and policy, where each prediction feeds into the next step.

An evaluation approach where models are tested on a fixed, predefined set of questions with known correct answers.

A special learnable token in transformers that aggregates information from all input patches and is used for classification.

A network component that projects learned representations into a space suitable for clustering tasks.

When multiple features in a neural network are active at the same time, often because they represent related concepts.

A neural network design that jointly trains two components together to produce better embeddings by learning from both query and document representations simultaneously.

Simultaneous optimization of interdependent components that improve each other iteratively.

The probability that all models in an ensemble produce incorrect answers on the same query.

Training approach where different modalities or models teach each other to improve overall performance.

A training strategy where two networks learn together, each selecting clean samples for the other to reduce noise impact.

A training framework where two models or components iteratively improve each other by learning from their complementary strengths.

A stable state where repeated application of an operation (like successive tangent unfoldings) converges or stabilizes.

An equilibrium where no group of players can jointly deviate and all benefit, even if they coordinate.

A game theory solution where no player benefits from unilaterally deviating from a recommended strategy.

A strategy that first captures broad patterns, then progressively refines details for better understanding.

A sequential decision-making approach that starts with broad estimates and progressively refines them to higher precision.

A curriculum learning approach that starts with learning simple components before progressing to optimizing complex global structures.

Identifying sections of code that perform the same function, even if written differently or in different programming languages.

The ability to automatically suggest or generate the next lines of code based on what the programmer has already written.

Percentage of program code executed by a test suite, measured by lines or branches.

A specialized task where a model modifies or refines existing code rather than creating new code, focusing on precision and surgical changes.

A specialized embedding designed specifically for source code that understands programming syntax and semantics, enabling tasks like code search and finding similar code snippets.

The ability of a model to write, complete, or suggest programming code based on prompts or partial code input.

A technique where a model fills in missing or incomplete code in the middle of existing code, using both the code before and after the gap as context.

Training a language model primarily on source code and technical documentation rather than general text, making it specialized for coding tasks.

A measure of how well code meets standards for readability, maintainability, and correctness.

The ability of a model to understand, analyze, and make logical inferences about source code and programming logic.

Restructuring existing code without changing its external behavior to improve readability and maintainability.

Encoding structured information (like circuit designs) as code-like syntax that language models can more easily learn and generate.

Process of examining code changes for bugs, quality issues, and adherence to standards before merging.

Automatically generating executable code (like plotting commands) from high-level specifications or natural language descriptions.

Techniques to confirm a student actually understands the code they wrote, not just copied it.

A language model specifically trained on programming code to excel at tasks like code generation, completion, and understanding.

Using multiple languages together in the same text, common in multilingual communities.

A model trained with a focus on understanding and generating programming code across multiple languages.

A language model trained specifically on programming code and related tasks, optimized to understand and generate code better than general-purpose models.

A language model trained specifically on programming code and code-related tasks rather than general text.

The ability to naturally mix two languages within the same text or conversation, switching between them based on context rather than treating them as separate.

A lookup table mapping compressed values back to original data; avoided in this approach to save memory.

The percentage of available discrete tokens in a codebook that are actually used during training or inference.

Indirect linguistic expressions that obscure sensitive meanings to evade detection or moderation systems.

An AI system that autonomously writes, debugs, and executes code to solve tasks without human intervention.

A normalized measure of variability that expresses standard deviation as a percentage of the mean, useful for comparing spread across different scales.

Words in different languages that share a common historical origin and similar meaning.

Identifying words in different languages that share a common origin and similar meaning.

A computational framework that models how an intelligent agent perceives, reasons, and acts in the world.

The process by which AI systems embed external interests into human decision-making architecture in ways users cannot easily perceive or resist.

A mechanism that gates speculative execution based on model confidence, without requiring ground-truth labels.

A structured representation of an agent's reasoning process that tracks concepts, relationships, and how understanding evolves over time.

Mental shortcuts that simplify decision-making but can lead to systematic biases in judgment.

A psychological framework explaining how working memory capacity affects learning and task performance.

AI assistance that helps users think through problems and refine their goals rather than just executing stated requests.

A classification system that organizes learning objectives or tasks by their intellectual complexity, from simple recall to advanced analysis.

A standardized measure of effect size that quantifies the difference between two groups in units of standard deviation.

The quality of maintaining consistent meaning and logical flow across multiple sentences or exchanges in a conversation.

The maximum circuit depth a quantum computer can execute before quantum information is lost to noise and decoherence.

Optical detection method that measures both amplitude and phase of light by interfering it with a reference laser signal.

The ability to preserve quantum information in a superposition state between measurements without collapsing it.

A neural retrieval model design that stores multiple token-level embeddings per document and uses late interaction to achieve higher retrieval accuracy than single-vector approaches.

When a model trained with sparse rewards gets stuck early because initial success probability is too low to learn from.

A technique that predicts user preferences by analyzing patterns from similar users' behavior.

Emergent problem-solving and decision-making capability that arises from coordinated interaction of multiple agents.

When input features are highly correlated with each other, making it difficult to isolate individual feature effects on predictions.

Data layout storing values column-by-column instead of row-by-row, optimized for analytical queries.

Finding the best arrangement or selection from a finite set of possibilities, like packing objects efficiently.

The process of converting model-generated commands into a format that can be transmitted and executed by a system.

The interface or set of instructions that a high-level planner sends to a low-level controller to specify desired robot behavior.

Shared beliefs and mutually recognized facts that enable effective collaboration between people or AI systems.

The ability of a model to understand and apply everyday logic and practical knowledge about how the world works.

A standardized database of publicly disclosed security vulnerabilities with unique identifiers and descriptions.

A standardized list of software and hardware weaknesses that can lead to vulnerabilities.

The ability to understand implicit physical, social, or practical prerequisites needed to solve everyday problems.

Minimizing the amount of data exchanged between devices or servers during distributed training.

The structure defining which agents can communicate with each other and how information flows between them.

The underlying purpose or goal behind how information is presented, which may be to persuade, mislead, or distort.

A model variant created and shared by the community rather than the original model creators, often with custom modifications.

A model that has been altered or fine-tuned by users outside the original development team, often to change its behavior or capabilities.

A smaller language model designed to use fewer computational resources while still performing useful tasks.

Evaluating reward quality relative to the current policy's skill level, recognizing that reward rankings change as the policy improves.

Natural language questions that define what an ontology should be able to answer, used to specify system requirements.

The process by which compilers examine code structure to identify optimization opportunities within their representation capabilities.

Transformations that improve code performance, memory usage, or other properties without changing program behavior.

Using compiler error messages to iteratively repair generated code until it compiles successfully.

A quantum physics constraint ensuring operations preserve valid quantum states and probabilities.

A text generation approach where the model continues or completes text from a given prompt, rather than engaging in back-and-forth conversation.

A prompt style where you provide the beginning of text and the model continues it, rather than asking a direct question.

Training a model to predict correct outputs given inputs, where targets are verified as correct before training.

The ability to work through multi-step problems, analyze nuanced information, and draw logical conclusions.

A document or record created to demonstrate that a system or organization meets regulatory requirements.

Official verification that a service meets specific regulatory or security standards required by industries like healthcare or finance.

Official verifications that a service meets specific security and regulatory standards (like HIPAA or SOC 2) required by certain industries.

Discrimination that emerges from how separate system components work together, not from individual parts alone.

The sequence of states a software component goes through, including initialization, execution, and cleanup, especially when dependencies change.

A 2D visualization in a SOM showing how a single feature varies across the map's grid positions.

A design pattern where UIs are built from reusable, self-contained pieces (components) that can be combined to create larger interfaces.

Model's ability to understand new combinations of learned concepts.

When combining outputs from multiple components violates probability axioms, even if each component is individually valid.

Text descriptions that specify multiple elements, their relationships, and spatial arrangements in the desired image.

The principle that the meaning of a complex expression is built from the meanings of its parts and how they combine.

The ability to understand new combinations of concepts by learning how individual components combine.

A technique for recovering sparse signals from far fewer measurements than traditional sampling requires.

The strategic distribution of computational resources between different components (like vision vs. language) based on task requirements.

The amount of processing power and memory available to run a model, which determines how much computation can be performed.

The amount of computation (time and memory) required for an algorithm to solve a problem.

The ability to deliver good results while using less processing power and memory than larger models.

Using code and algorithms to test mathematical hypotheses and discover patterns empirically.

The amount of memory, processing power, and time required to run a model; a smaller footprint means the model can run on less powerful hardware.

The extra processing power, memory, or time required to run a model, which impacts speed and resource consumption.

Using algorithms and AI during image capture to enhance photos beyond what the camera sensor alone can achieve.

Verifiable evidence embedded in generated text about which internal states or computations a model actually used.

The strategic distribution of a model's processing power—in this case, spending more computational effort on thinking through problems rather than other tasks.

A limited allocation of computational resources that an AI agent is permitted to use, enforced through technical means.

How well a model performs relative to the computational resources (processing power and memory) required to run it.

A model designed to run with minimal processing power and memory, making it practical for devices with limited resources.

Hardware architecture that performs computation directly within memory, reducing data movement bottlenecks.

Achieving the best performance for a given amount of computational resources.

The ability for an AI model to interact with computer interfaces, navigate software applications, and execute actions on a user's behalf by understanding and responding to visual or textual representations of screens.

Directions in a neural network's internal activation space that represent human-interpretable concepts like accent or age.

An interpretable model that makes predictions by routing inputs through a layer of human-understandable concepts rather than opaque features.

When the target concept or decision boundary changes over time during learning.

A method to identify and visualize interpretable patterns in a neural network's latent representations that drive its predictions.

A low-dimensional geometric structure where related concepts are organized continuously, like a curved surface in high-dimensional space.

The process of mapping different textual expressions of the same idea to a single standardized representation, such as mapping 'MI' and 'myocardial infarction' to the same medical concept.

Explaining AI decisions using human-understandable concepts rather than raw features.

The process by which agents develop shared understanding of categories and meanings through interaction.

Multi-agent games where players act simultaneously in stochastic environments with state-dependent payoffs.

A measure of how difficult an optimization problem is; higher values mean slower convergence and more iterations needed.

Applying model modifications selectively based on input characteristics rather than globally to all inputs.

A reinforcement learning technique that estimates action value only within trajectories meeting specific conditions.

A property where prediction set coverage guarantees hold for specific subgroups or conditions, not just on average across all data.

A measure of uncertainty in predicted tokens given context; low entropy signals memorization, high entropy signals generalization.

The average distance between selected and validation target embeddings within a cluster, used to rank training examples.

The ability of a model to generate output (like text) based on specific input conditions or prompts provided to it.

The expected value of an output given specific input conditions, used as a deterministic baseline prediction.

Misaligned behavior that only appears when inputs share features with the training data, while appearing safe on out-of-distribution prompts.

A neural network architecture that learns to make predictions conditioned on observed context, useful for probabilistic forecasting.

A probabilistic model that learns to make predictions by conditioning on observed examples, useful for few-shot learning and uncertainty estimation.

The ability to generate text that follows specific conditions or constraints, rather than producing output freely.

A risk metric that focuses on the worst-case outcomes rather than average performance, useful for safety-critical tasks.

A neural network that learns to generate new data matching specific conditions or constraints.

Guiding a generative model's output by providing additional input signals like pose or depth maps.

A model component that takes an external parameter (like speed) and modulates the policy output based on that input.

Improving generated images by providing additional context or constraints during generation to guide the model.

Ensuring a model's confidence scores accurately reflect its true probability of being correct.

Assigning uncertainty scores to model predictions to identify outputs that may need human verification.

Statistical bounds around predictions that quantify uncertainty; here used to identify when model predictions are unreliable.

Statistical bounds that remain valid at any stopping time, allowing you to check results continuously without invalidating confidence levels.

A decoding strategy that stops refining tokens when model confidence exceeds a set threshold.

The pattern of model confidence scores across sequential steps in a reasoning process, from initial exploration to final answer.

Refusing predictions when the model's confidence score is below a threshold.

A strategy that selects which tokens to generate next based on the model's prediction confidence, enabling adaptive and efficient generation.

Training a model using rewards based on how well its confidence scores match its actual correctness.

Weighted majority voting where each candidate answer gets a confidence score from a critic model before selection.

The tendency to seek or interpret information in ways that confirm existing beliefs or outputs.

Situations where an AI system has competing goals—like serving users well versus generating revenue for its creators.

Method providing prediction intervals with statistical guarantees on coverage.

The ability to direct a model to generate specific 3D shapes or structural states of proteins.

A collection of different 3D shapes a molecule can adopt in solution, representing its thermodynamic states.

A distinct 3D shape or arrangement that a protein can adopt, often with different biological functions.

Applying a learned conformational change from one protein to structurally similar proteins in the same family.

When multiple variables are entangled, making it impossible to isolate the effect of one variable.

When an agent misuses its elevated permissions to perform actions it shouldn't, tricked by user input.

A table showing how often a classifier correctly or incorrectly predicts each category, revealing systematic biases in predictions.

A property of FFT where real-valued signals have redundant information due to symmetric complex conjugate pairs in frequency domain.

A loss function for sequence-to-sequence tasks that aligns variable-length inputs to outputs without requiring explicit alignment.

A routing pattern where multiple neurons must agree (be mutually exclusive) to activate a particular processing path.

A deployment rule that specifies how outcomes (rewards, penalties, eliminations) are distributed among agents in multi-agent systems.

A reference policy or set of actions known to be safe, used to measure how much riskier a proposed action is.

A penalty that prevents value estimates from being too optimistic about unseen actions in offline RL.

A training method that encourages models to respond symmetrically to paired prompts representing opposing perspectives.

A method that ensures a model's reasoning process logically supports its final answer by adding consistency rewards during training.

A safety training approach that guides a model to behave according to a set of principles or rules, helping it generate more helpful and harmless responses.

Restricting a model's token generation to a predefined set of allowed tokens during inference.

Text generation that must follow specific rules or constraints, such as producing output in a particular format or structure.

Training an AI system to maximize performance while respecting hard constraints (like deadlines or budgets).

Finding solutions that satisfy a set of constraints, used here to resolve conflicts between inferred events.

A computational problem where you find values for variables that satisfy a set of constraints.

A tool that finds valid solutions to problems with multiple constraints, used here to verify mechanical assembly feasibility.

Action planning that explicitly respects safety, regulatory, and operational constraints during decision-making.

Improving AI safety by restricting what actions an agent can take, rather than trying to detect bad behavior after it happens.

Validating each step of a plan by checking outputs against automatically derived constraints based on task requirements.

Fixing errors in reasoning by making minimal changes that satisfy logical or evidential constraints.

Whether a study actually measures the real concept it's supposed to test, not something else.

Measurement error caused by factors unrelated to what is being assessed, like test anxiety or social power dynamics.

The study of how specific form-meaning pairings in language convey meaning beyond individual words.

A mechanism that activates learned corrections only when the robot is physically touching the object.

Physical interactions where the robot frequently touches and manipulates objects, making control sensitive to small errors.

Robot tasks where success depends critically on precise control of forces and contact interactions with objects.

Packaging software and its dependencies into isolated, portable units that run consistently across different computing environments.

A model or system that screens text before or after generation to block unsafe, harmful, or policy-violating content.

Safety mechanisms built into a model that prevent it from generating harmful, inappropriate, or restricted content.

Built-in safety mechanisms that prevent a model from generating harmful, offensive, or inappropriate content.

The process of reviewing and filtering text or other content to remove or flag material that violates policies or safety guidelines.

The degree to which written content presents information without favoring particular political viewpoints or groups.

Safety guidelines and filters built into a model to prevent it from generating harmful, illegal, or unethical content.

The task of automatically detecting and categorizing text that violates policies or could cause harm, such as hate speech, violence, or misinformation.

A memory system where data is retrieved by its content similarity rather than by a fixed address, enabling fuzzy matching.

A technique that preserves structural content by anchoring style transfer to the original image's spatial layout and structure.

The ability to maintain consistent meaning and logical flow when processing long sequences of text or conversation.

Reducing the size of conversation history while preserving important information for efficient processing.

A model's ability to maintain coherent understanding and recall of information across long passages of text without contradicting itself.

Transferring knowledge from interaction trajectories into model parameters by learning from contextual examples.

Removing or suppressing the influence of specific text spans from an LLM's processed context after they've been cached.

Techniques to increase the maximum sequence length a language model can process beyond its original training length.

Retaining only relevant information from execution history to reduce noise and improve decision-making in subsequent steps.

The process of collecting and organizing relevant information from history to answer specific questions or solve tasks.

Managing what information an agent can access and use to prevent hallucination and ensure relevance.

The maximum amount of previous text a model can consider when generating its next output; longer context allows the model to maintain coherence over longer passages.

Organizing and maintaining relevant information for AI decision-making.

A technique to process long sequences by distributing context across multiple devices or processing units in parallel.

Irrelevant or noisy information degrading model performance in a given context.

A model's ability to remember and use information from earlier parts of a conversation or document.

A model's resistance to being misled by incorrect external signals while still benefiting from correct or relevant context.

A mechanism that selectively directs relevant learned patterns from one model component to another based on current needs.

When an AI model's input context window fills up and earlier information is lost, requiring mechanisms to preserve key data.

The maximum number of tokens a model can process in a single conversation or prompt.

A system that adjusts its behavior based on the specific input or situation rather than using fixed, unchanging patterns.

A tendency to respond reactively to immediate context rather than proactively initiating strategies independent of user input.

A mechanism that selects only relevant evaluation criteria for each specific query to improve efficiency.

Speech recognition that uses surrounding information like conversation history to improve transcription accuracy.

Providing additional context (like original text or reasoning steps) to translation models to improve accuracy.

A model that adapts its predictions based on relevant historical examples or information provided as input context.

A formal system of rules that defines which sequences of symbols are valid in a language.

AI systems that maintain and reuse long conversation histories across multiple turns of interaction.

Problems requiring the model to extract and use large amounts of information from the input prompt to generate correct outputs.

Help or instructions tailored to the current situation rather than generic pre-stored information.

Adjusting model behavior dynamically based on the specific input or context rather than using fixed settings.

A learning algorithm that selects actions based on context and learns from feedback to improve future decisions.

Numerical representations of text that capture meaning based on surrounding context, rather than treating each word independently.

The assumption that a model produces consistent outputs when a task is reformulated in contextually equivalent ways.

A statistical model that separates between-group and within-group effects to understand how group membership influences individual outcomes.

A decision-making framework where an agent plans joint actions with partners whose capabilities are represented as context variables.

Influence from surrounding information (like examples or previous actions) that pushes an agent away from its intended behavior.

Making decisions by considering how individual observations relate to and inform each other within a broader context.

A way of encoding text where the meaning of each word depends on the words around it, rather than being fixed for every occurrence.

Numerical encodings of text that capture meaning based on surrounding words, allowing downstream tasks to understand nuanced language.

The ability to distinguish and respond differently to harmful versus benign uses of the same concept.

The intermediate representation space in a diffusion model where semantic and structural information is encoded.

Machine learning technique that identifies recurring themes in text while considering the surrounding context of words.

A feature or pattern in input text that activates hidden misaligned behavior in a model, even when standard evaluations show the model is safe.

Uncertainty caused by changing conditions over time, like user preferences shifting.

The ability of a model to interpret the meaning of words and phrases based on surrounding text, rather than treating each word in isolation.

Vector representations of words that change based on surrounding context, capturing different meanings in different sentences.

Incrementally updating a neural network on new data as it arrives, rather than retraining from scratch.

A process where an agent's defenses dynamically adapt and improve in response to new threats encountered at runtime.

Training models to learn new tasks without forgetting previously learned ones.

Further training a pre-trained model on domain-specific data to adapt it to specialized knowledge.

Further training a pre-trained model on new data to adapt it without starting from scratch.

Further training a pretrained model on domain-specific data to specialize it for particular tasks.

Further training of an already-trained model on new data to extend or improve its capabilities without starting from scratch.

Real-time monitoring of a quantum system that produces a stream of measurement data used to update state estimates.

Encoding data as smooth, unquantized values rather than discrete tokens, preserving fine-grained temporal details.

Generating probability-based continuous scores instead of discrete labels to provide fine-grained evaluation signals.

A probabilistic model where a system transitions between states continuously over time with memoryless transitions.

Automatically checking that modifications respect predefined rules about what can and cannot be changed.

A transformation that maintains the original program's behavior and guarantees, ensuring correctness is preserved.

A mathematical property ensuring a system's outputs converge to a stable state regardless of initial conditions.

A mathematical property ensuring that a system brings nearby states closer together over time, guaranteeing stability.

A test dataset with paired examples showing contrasting properties (e.g., safe vs. unsafe versions of the same task).

A training technique that learns by comparing similar and dissimilar examples to create better representations.

Training objective that pulls similar examples together and pushes different ones apart.

A method that learns shared embedding spaces by contrasting similar and dissimilar image pairs, then ranks candidates by similarity.

Creating evaluation criteria by comparing gaps between teacher and model responses to identify what distinguishes good from bad outputs.

Breaking down a neural network's output into individual contributions from different neurons or neuron groups.

Special tokens added at the beginning of a prompt that tell the model what style, domain, or format to use for its output.

The order and logic in which steps are executed in a procedure or program.

A special token emitted by a small model to signal that it needs assistance from a larger model.

Special tokens inserted into sequences to guide model behavior, such as signaling whether to show an ad or organic content.

A mathematical tool that projects unsafe actions to safe ones, guaranteeing constraint satisfaction but potentially masking policy incompetence.

Creating videos where users can specify desired content, motion, or interactions through explicit control mechanisms.

A standardized test where variables are carefully isolated to measure specific effects, like changing one visual attribute while keeping everything else the same.

A learned stochastic process that transforms a distribution by following a time-dependent control signal.

A standardized set of predefined terms used consistently across systems to ensure consistent meaning and reduce ambiguity.

Automatically designing a decision-making system that controls when and how to execute actions.

A technique that adds spatial control to diffusion models by conditioning generation on aligned input maps (like depth or property masks).

Physics problems where fluid flow effects dominate over diffusion, creating sharp gradients and moving fronts.

The point during training when a model's performance stabilizes and stops improving significantly, indicating it has learned the patterns in the data.

Mathematical proof that an optimization algorithm reliably reaches a good solution and quantifies how fast it gets there.

Mathematical proofs that an algorithm will reach a correct solution under specified conditions.

How quickly an optimization algorithm approaches the optimal solution, typically expressed as a function of iterations.

Agreement across multiple independent methods or analyses pointing to the same conclusion.

When different models independently learn similar features or representations from different training signals.

A synthesis strategy that assembles multiple complex building blocks in parallel before combining them, reducing total steps.

AI systems designed to understand and respond to human language in natural, dialogue-like interactions.

An AI system designed to conduct multi-turn dialogue with users to accomplish specific tasks, like medical interviewing.

Using dialogue with a chatbot or AI agent to probe and verify student understanding through questioning.

The model's ability to maintain logical consistency and relevance across multiple turns of dialogue, making responses feel natural and connected.

How naturally and coherently a model engages in back-and-forth dialogue, matching human conversation patterns.

A model specifically trained to understand and generate natural dialogue, optimized for back-and-forth interactions rather than one-off text generation.

A language model specifically trained and optimized to engage in multi-turn dialogue with users.

When a user completes a desired action, typically making a purchase after viewing a product.

A physical limit on inventory levels described by a convex set, constraining what orders are feasible.

A weighted sum of points where weights are non-negative and sum to one, representing a point within their geometric hull.

A function where any line segment between two points on the curve lies above the curve, ensuring a single global minimum.

Mathematical technique for finding the best solution to a problem with a single global optimum.

A geometric shape formed by the intersection of linear inequalities, with vertices representing extreme points.

Geometric regions where any straight line between two points in the region stays entirely within the region, enabling efficient trajectory planning.

A technique that scans across input data using small filters to detect local patterns, commonly used in image processing but here applied to text for efficiency.

A game setting where agents work together toward shared objectives rather than competing against each other.

The geographic coordinate system (e.g., latitude/longitude) used to define spatial locations and ensure consistency across operations.

Internal model representations that explicitly encode spatial positions and coordinate information.

A game where players benefit from choosing compatible strategies, used here to model how agents negotiate shared meaning.

Game theory scenarios where agents benefit from matching actions but may also benefit from strategic differentiation.

An optimization where data is only copied when modified, allowing multiple references to share the same data until changes occur.

An attention mechanism where peripheral tokens (patches) interact only through central core tokens, reducing computation.

Identifying when different mentions in text refer to the same entity or concept.

Selecting a small subset of real samples from a large dataset that best represent the original data distribution.

A collection of documents or text used as the knowledge base for retrieval in RAG systems.

Selecting query terms that best distinguish relevant documents from irrelevant ones in a specific corpus.

A filtering mechanism that validates whether a proposed solution is correct before allowing it to advance in a search process.

A domain adaptation technique that aligns the second-order statistics (covariance) between source and target domains.

A model's ability to maintain performance when input data is degraded (e.g., noise, blur, missing values).

A mathematical measure that compares how similar two embeddings are by calculating the angle between them, with values closer to 1 meaning more similar.

A method of comparing two vectors based only on their direction, ignoring their magnitude, making it scale-invariant.

A learnable activation function using cosine waves with adjustable frequency and phase to process data nonlinearly.

An adversarial attack that accounts for the real-world cost or feasibility of modifying each feature.

Measuring model performance while accounting for the computational and operational costs required to achieve results.

The ability to deliver useful results while using fewer computational resources, reducing the expense of running the model.

The balance between inference cost (compute, latency) and answer quality that systems must optimize for.

Training approach that weights errors differently based on their downstream impact or cost in the application domain.

A training methodology that combines chain-of-thought reasoning with masked autoencoder techniques to improve model understanding of text relationships.

A failure mode where decision thresholds shift based on the number of sources consulted, causing inconsistent decisions.

An instance where a model's prediction fails to match reality, used to identify and fix model errors.

Testing what would happen if you changed a strategy, without actually running the experiment in the real world.

An explanation showing what input changes would alter a model's prediction to a different outcome.

Creating alternative scenarios showing what would happen if something were different (e.g., if an object didn't exist).

Training examples where evidence is semantically related but contradicts the claim, testing if models truly use evidence.

A question about what would have happened if a variable had taken a different value (e.g., 'what if the patient had received treatment?').

Reasoning about what would have happened under different actions or conditions than what actually occurred.

A unified representation that tracks both physical scene state and mental state variables together.

A measure of how two variables change together; structured covariance means features are correlated in specific patterns.

The process of learning or updating the statistical properties of measurement and process noise in a filtering system.

Aligning a model's sensitivity structure to match the statistical structure of task-irrelevant variations in data.

A robust black-box optimization algorithm that adapts the covariance matrix of a search distribution to efficiently explore high-dimensional spaces.

When the distribution of input data changes between training and real-world use, causing models to fail.

The extent to which training data represents all relevant aspects or regions of a document or domain.

Requirements ensuring sufficient representation of all groups, including subgroups defined by multiple attributes, in training data.

A diagnostic metric that measures beyond-spectrum structure by comparing analog prediction gains to linear prediction.

Measuring what proportion of a problem space a model can reliably handle.

A record of what information has been found and what gaps remain, guiding agents toward unresolved areas.

Finding an efficient route for a vehicle to visit all cells or areas in a region.

The process of proving that testing has comprehensively covered all relevant operating conditions and edge cases.

A training technique that prioritizes data from under-explored regions of the state-action space to improve model robustness.

Testing approach that systematically explores different input regions to find edge cases and failures.

Systematic asymmetric treatment of opposing political viewpoints in language model responses, including differences in tone, depth, and engagement.

A quantum operation that preserves physical validity by maintaining positivity and trace properties of quantum states.

Running a model's predictions using a computer's central processor rather than a specialized graphics card, which is slower but requires less specialized hardware.

A mathematical limit on how precisely any unbiased estimator can determine a parameter from data.

A benchmark methodology using a fixed document collection, queries, and human relevance judgments to evaluate retrieval systems.

A standardized legal license (like CC-BY-4.0) that specifies how others can use, share, and modify a work with clear attribution requirements.

A measure of how useful and novel the connections a model generates are for creative tasks.

User perception of a system's trustworthiness and expertise, affecting whether they believe its information.

A restriction on which successor states are considered reachable or believable from a given source world.

The process of determining which actions or steps in a sequence deserve reward or blame for the final outcome.

Breaking down evaluation into multiple independent criteria to reduce complexity and improve verification accuracy.

When an AI model can achieve high scores by paraphrasing or trivially matching evaluation criteria without genuine understanding.

Detailed feedback that scores responses across multiple specific evaluation criteria rather than a single overall score.

A language model trained to evaluate the quality of another model's outputs and identify areas for improvement.

Using a separate model to evaluate and reject outputs that contain errors, improving final answer quality.

Using a neural network to predict game payoffs for different action combinations, amortizing learning across multiple game states.

Identifying which specific error in a failed trajectory was responsible for the final failure outcome.

A position in code generation where errors are most likely to occur and cause downstream semantic failures.

An agent that reviews and validates the recommendations and execution plan of other agents to ensure correctness and coherence.

Providing corrective feedback or explanations of errors as examples to steer model behavior toward better outputs.

A standardized format for describing machine learning datasets with structured information about their contents and properties.

Mechanism allowing one sequence to attend to and focus on another sequence.

Transferring knowledge between models with fundamentally different designs, attention mechanisms, or tokenizers.

A neural module that merges information from two sources by learning which parts of each are most relevant.

Mapping the same source text across multiple independent interpretations to enable direct comparison of how different schools read identical material.

Testing whether a model trained on one dataset generalizes to perform the same task on a different dataset.

A creativity technique where ideas from one unrelated domain are applied to solve problems in another domain.

When optimizing multiple objectives jointly causes conflicts where improving one metric degrades another.

Training a single model on data from multiple robot types or humans to learn generalizable physical principles.

Learning to control one body type (like a humanoid robot) using data from a different body type (like humans).

A model architecture that takes a query and document together as input and directly outputs a relevance score, unlike dual-encoders that score them separately.

A loss function that measures how well a predicted probability distribution matches a target distribution.

Running an AI model in different network environments or systems than the one it was trained on.

A technique that uses cross-validation to estimate and remove systematic style differences between teacher models before applying distillation.

A technique that splits data into parts to estimate nuisance parameters on one part and the target parameter on another, reducing bias.

The ability to understand relationships and transfer knowledge between different languages, such as answering a question in one language based on text in another.

Matching equivalent sentences or meanings across different languages.

The ability of a model to understand and relate concepts across different languages, allowing it to find similarities between text in different languages.

The ability of a model to understand and work with multiple languages, sometimes even translating concepts between them.

The ability of a model to represent similar meanings in different languages as nearby points in its vector space, so translations and equivalent concepts are treated as semantically close.

The ability of a model or probe trained on one language to work effectively on other languages.

The ability to find and compare similar content across different languages by representing them in a shared mathematical space.

The ability to find relevant documents or text in one language when searching with a query in a different language.

The ability to recognize that sentences or phrases in different languages have the same or similar meaning and represent them close together in numerical space.

The ability to measure how similar two sentences are even when they are written in different languages.

The ability of a model trained on multiple languages to apply knowledge learned from one language to understand or generate text in another language.

The ability of a model to comprehend relationships and meanings across different languages, enabling tasks like translation and multilingual reasoning.

Connecting representations from different types of data (like speech and text) so they work together effectively.

An attack that manipulates multiple input types (like images and text) together to deceive a model.

A mechanism that aligns and weights information between different modalities like images and text.

Ensuring semantic consistency between different modalities (e.g., text and images) in generated content.

Ensuring that representations across different modalities (images, 3D, text) align and reinforce each other.

Alignment in how models from different modalities (e.g., vision and language) represent the same stimulus.

Transferring knowledge from a teacher model in one modality (text) to a student model in another modality (audio).

The process of combining information from multiple modalities (e.g., vision and text) into a unified representation.

When a model produces contradictory predictions for the same concept represented in different modalities.

The process of connecting and aligning information from different modalities (like text and 3D geometry) within a single model.

The ability to find relationships between different types of content, such as matching natural language descriptions to code snippets.

Learning representations by predicting one modality from another.

The ability to connect and reason about information from different input types (like audio and video) together to draw conclusions.

Relationships and connections between different types of data (e.g., how text and images relate to each other).

The ability to search and find relevant items across different data types, such as finding images using text queries or vice versa.

Finding similar scenes across different input types (e.g., finding a visual scene matching an audio description).

Computing relevance or similarity scores between information from different modalities (e.g., text and images).

The ability of a model to share semantic understanding between different input modalities like vision and text.

The ability to measure how closely related content from different types of input (like images and text) are to each other.

Training a model using aligned labels from different data sources or modalities (e.g., video and biomechanical data).

Exchanging information between different input types (text and vision) to guide compression decisions.

Applying knowledge or rules learned from one model to improve performance in a different, often smaller model.

The ability of AI tools to access information from other modules and make decisions based on shared context.

How the demand for one product changes when the price of a different product changes.

The ability of a model to perform consistently when input text or audio switches between different writing systems or languages.

Multi-step problems where each step requires a different reasoning skill and depends on previous outputs.

Linking and connecting information from multiple different data sources to understand how they relate to each other.

The process of comparing and resolving conflicting information from multiple sources to determine accurate answers.

A model's ability to work on new individuals without retraining, despite differences in neural anatomy.

A technique that splits data into multiple subsets to train and test models, reducing overfitting risk.

A mechanism that transfers motion information from one camera viewpoint to another while maintaining consistency.

The degree to which internal representations align when processing the same task in different formats or modalities.

When a model fails to recognize that the same object in different camera views is the same entity, leading to counting errors.

Aligning images captured from different viewpoints (e.g., street-level and overhead) to find correspondences.

Data or labels collected from many independent contributors rather than a single expert source.

A metric that evaluates probabilistic forecasts by measuring how well predicted probability distributions match actual outcomes.

Calderbank-Shor-Steane quantum codes combining two classical error-correcting codes for quantum error correction.

A 3-dimensional algebraic variety defined by a degree-3 polynomial equation.

A mathematical structure representing data as a collection of cubes at different scales, used in topological analysis.

NVIDIA's parallel computing platform that runs code on GPUs to process many tasks simultaneously.

Optimized GPU code that performs specific computational operations efficiently.

Systematic preference for values, norms, or behaviors from one culture over others in model outputs.

Using computational methods to quantify and analyze cultural phenomena like dialogue patterns, social interactions, or linguistic variation.

The ability to understand and infer cultural context, significance, and metadata from visual or textual information.

Statistical measures that describe probability distributions, used to track activation behavior.

Training data that has been carefully selected and filtered to include only high-quality examples relevant to specific tasks or domains.

Carefully selected and filtered training examples chosen for quality rather than quantity, often resulting in models that produce more structured and reliable outputs.

RL approach where agents explore by seeking states where their world model makes poor predictions.

Training strategy that gradually increases task difficulty to help models learn robustly.

Training strategy that presents examples in increasing order of difficulty.

A training constraint that penalizes curved or winding paths in the learned representation space.

Allowing users to configure AI behavior, roles, and interaction patterns to match their specific needs and preferences.

A constraint requiring a model to reconstruct its original output after transforming through intermediate steps.

A metric measuring how many different paths code can take; lower values mean simpler, easier-to-maintain code.

A consensus mechanism using a directed acyclic graph structure to order and validate transactions or events without requiring a single chain.

An interactive learning method where a human corrects the model's mistakes during training to fix distribution mismatch.

Smaller, controlled refinement iterations that reduce the magnitude of updates to stabilize computation.

Combining observed measurements with model predictions to calibrate and improve model parameters.

Measuring how much each training example contributes to a model's final performance using gradient-based methods.

Technique to increase training data by creating variations or new samples from existing data.

When test data accidentally leaks into training, artificially inflating a model's measured performance.

The process of carefully selecting, cleaning, and organizing training data to improve model quality; better curated data often leads to better model performance.

Predicting how a model would behave if specific training examples were excluded without retraining.

Predicting how a model's behavior would change if specific training data were excluded without retraining.

The variety of examples in a dataset across different attributes, styles, or contexts to improve model generalization.

Achieving better model performance using less training data through improved data quality or learning methods.

The process of selecting, organizing, and preparing training data to improve model performance.

The process of automatically identifying and pulling relevant information from unstructured or scattered data sources.

The accuracy and correctness of data representation in a visualization or output.

A self-reinforcing cycle where a system identifies gaps, generates data to fill them, and uses that data to improve itself.

Policies and procedures that control how sensitive data like medical images are stored, accessed, and used.

The tendency for computation to move toward where large datasets are stored, rather than moving data to where computation happens.

Variation in data distribution across different sources or groups.

Combining data from multiple sources into a unified, usable format for analysis or querying.

Security controls that prevent sensitive data from leaving a restricted environment without authorization.

Gaps or missing values in a dataset caused by sensor failures, blinks, or other interruptions.

The expense of changing, augmenting, or purchasing training data to improve model fairness and reduce bias.

The sequence in which training examples are presented to a model during training.

Adversarial manipulation of training data to degrade model behavior while maintaining normal performance on standard metrics.

A sequence of automated steps that transform raw data into a format suitable for model training.

The relevance, accuracy, and usefulness of training data, which can be more important for model performance than simply having more data.

The practice of carefully selecting and filtering training data for relevance and accuracy rather than simply using larger amounts of raw data.

The process of checking whether data in a dataset meets required standards and accurately represents what it claims to represent.

The proportion of training data allocated to different modes or tasks (e.g., 70% thinking vs. 30% non-thinking).

Mistakes where LLMs incorrectly cite, omit, or misread values from tables despite understanding the table structure.

A centralized catalog storing metadata about available data sources and their query interfaces.

A guarantee that your data is stored and processed only in a specific geographic region, helping meet regulatory requirements.

When researchers use datasets from previous studies in their own research rather than collecting new data.

A controlled variation in the order or selection of training data used to train a model, allowing researchers to isolate how data ordering affects learning.

Choosing a subset of training data based on quality or relevance metrics rather than using all available data.

Automatically generating training data from existing datasets to teach models new tasks.

Automated checks that verify data meets quality and correctness requirements before use.

An approach that focuses on improving data quality and efficiency rather than just model architecture.

Achieving good performance with minimal training data through smart collection or learning strategies.

A quantum operation that transforms quantum states based on classical input data.

Modifying training data itself rather than changing model architecture or training procedures.

Distributing training data across multiple GPUs that compute gradients independently then synchronize.

Compressing a large dataset into a smaller synthetic version preserving key information.

A classification system that organizes datasets by their characteristics like sparsity, scale, and sequential structure.

A neural network design pattern that serves as the structural foundation for this model, determining how it processes and generates text.

Simplified power flow approximation that ignores reactive power and voltage variations for faster computation.

A faster sampling method for diffusion models that skips intermediate steps while maintaining quality.

A mathematical result showing that coherent probability assignments must satisfy certain consistency conditions that can be tested via betting.

Creating entirely new protein sequences from scratch rather than modifying or copying existing ones.

Removing duplicate detections of the same object across multiple video frames to avoid counting it twice.

A transformer-based language model architecture that uses disentangled attention mechanisms to improve how the model weighs different parts of the input text when making predictions.

Each agent independently executes its assigned action without real-time coordination, based on a centrally-computed plan.

Systems where rules and enforcement are managed by participants themselves rather than a central authority.

A training approach where a model is developed across multiple independent computers or organizations rather than in a single centralized facility, allowing distributed collaboration.

Outputs that mislead a recipient, whether or not the model has an internal mechanism designed to deceive.

A loss function that directly penalizes the cost incurred by making decisions based on a model's predictions.

The ability to replay and understand why a system made a specific decision after the fact.

A tree-based model that recursively partitions data using threshold rules to make predictions.

Educational approach where learners make choices that shape outcomes, reinforcing clinical decision-making skills.

Combining predictions or confidence scores from individual modalities to make a final decision.

A mechanism that selects actions based on current state, goals, and expected outcomes to maximize success.

A tool or system that provides information and analysis to help humans make better decisions without replacing human judgment.

An approach that evaluates systems based on the quality of decisions they enable under different costs and benefits.

A formal description that binds visual elements to data fields, separating what to show from how to render it.

A component that converts compressed internal representations back into human-readable outputs like audio or images.

A vector in an autoencoder that reconstructs or represents a specific learned feature from compressed latent space.

The inherent randomness in a model's output generation, even when given identical inputs multiple times.

A type of LLM that generates text one token at a time, like GPT models.

Language model design that generates text sequentially without a separate encoder, like GPT models.

Converting model outputs into human-readable text or structured predictions.

The speed at which a model generates output tokens one at a time, a critical bottleneck in long-context scenarios.

Methods for generating text from a language model, such as greedy selection, beam search, or temperature sampling.

A parameter controlling randomness in AI text generation; higher values increase diversity, lower values increase predictability.

The sequence and order in which tokens are generated or unmasked during the model's iterative generation process.

Converting compiled binary code back into human-readable pseudo-code approximations.

Breaking a complex problem into smaller, simpler sub-problems that are easier to solve and understand.

A reward system that breaks complex requests into atomic, checkable questions to provide interpretable feedback for model training.

An architecture that separates visual understanding from information retrieval, forcing exhaustive visual analysis before external search.

Training separate reward objectives for different tasks (e.g., binary judgment vs. error localization) instead of optimizing them jointly.

The separation of economic output growth from human welfare, where GDP expands while human ownership share and consumption decline.

The process of removing duplicate or near-duplicate examples from training data to improve model efficiency and prevent overfitting to repeated content.

An AI system that performs multi-step research by reasoning through problems and making multiple search queries.

A context-dependent rule that determines which attacks between arguments succeed based on the current context or regime.

Automatically identifying problems or errors in software artifacts, such as incomplete or ambiguous descriptions.

An extension of Gaussian Splatting that allows the 3D primitives to deform over time to handle dynamic scenes.

The number of independent ways a mechanical part can move or rotate in an assembly.

Consequences of an agent's actions that appear many steps later, making it harder to learn cause-and-effect relationships.

A reward signal that only becomes available after multiple steps or actions have been completed.

Feedback or verification of agent actions that arrives after a delay, requiring the agent to maintain accountability over time.

A failure mode where removing steps from a plan increases its evaluation score despite making it incomplete.

A form of democracy where citizens and representatives engage in reasoned discussion to reach decisions.

Real-time natural language guidance that shapes agent reasoning without requiring model parameter updates.

The difference between a reasoning-tuned teacher model and its base pre-trained version, used as a training target.

Using machine learning to predict how much of a product customers will buy given prices and other factors.

Removing or hiding demographic information (like gender) from model inputs to reduce bias in decision-making.

Learning which demographic attributes (race, age, etc.) are most influential in predicting how annotators will judge subjective content.

Training examples collected from real robots performing tasks, used to teach the model how to execute similar actions.

A training approach where the model learns to reconstruct clean audio from corrupted or noisy versions, improving its ability to extract meaningful features.

A neural network trained to reconstruct clean text from corrupted or noisy versions, learning to remove noise while preserving meaning.

A training approach where a model learns to reconstruct clean audio from noisy versions, making it better at understanding speech in real-world conditions.

A technique where a model learns to gradually remove random noise from data to reconstruct meaningful content, used as an alternative to traditional token prediction.

A training objective that learns to predict noise in corrupted data, used in diffusion models for stable gradient-based optimization.

Generating detailed, comprehensive descriptions of images that capture rich visual information and relationships rather than brief summaries.

A compact vector representation where most dimensions contain meaningful information, as opposed to sparse embeddings that are mostly zeros.

Vector representations where most or all of the numbers contain meaningful information, as opposed to sparse embeddings where most numbers are zero.

Fine-grained, continuous-valued signals that provide detailed information about solution quality for training or selection.

A neural network where all parameters are active for every input, in contrast to sparse architectures like mixture-of-experts that selectively activate different parts.

A technique that converts documents and queries into dense vectors so that relevant passages can be found by comparing their numerical representations rather than matching keywords.

A compact numerical format where meaning is captured in a fixed-size list of numbers, making it efficient for storage and similarity comparisons.

A search method that converts text into a single, compact numerical vector and finds similar documents by comparing these vectors.

A retrieval system using learned embeddings to find semantically similar documents via vector similarity.

Training data that provides detailed annotations for every part of an input, rather than just overall labels.

A compact numerical representation where most values are non-zero, used to efficiently store and compare the meaning of text.

A compact numerical representation of text that captures its meaning, allowing the model to compare how similar different pieces of text are to each other.

Numerical representations of text where each word or sentence is converted into a list of numbers that capture its meaning, allowing the model to compare semantic similarity.

A compact numerical format where text is encoded as a list of numbers that capture its meaning, allowing efficient similarity comparisons.

A mathematical space where text is represented as vectors of numbers, positioned so that similar meanings are located close together.

Compact numerical representations where most values are non-zero, used to encode the meaning of text in a form that computers can compare mathematically.

Automatically generating descriptions for multiple events throughout a long, untrimmed video with precise timing.

Dense embeddings use all dimensions with non-zero values (like traditional neural embeddings), while sparse embeddings mostly contain zeros and are more interpretable and storage-efficient.

Measuring how crowded solutions are in the search space to maintain diversity in the population.

A mathematical representation of a quantum state that describes the statistical properties of a quantum system.

Local peaks in a probability distribution where data is most concentrated; useful for summarizing multimodal data.

A method that aligns models by learning from the geometric clustering of accepted responses in the model's representation space.

A visual map showing which models, datasets, and tools a system relies on and how they connect.

Finding a tree structure that represents grammatical relationships between words or elements in a sequence.

How changes to one software component automatically affect other components that depend on it.

The ability to understand how multiple facts relate to and affect each other when making decisions.

The pattern of how information at different distances in text relates to each other, characterized by width and scale.

The specific technical setup of an LLM in production, including safety layers, routing rules, and interface choices.

Risk that a single model's values or biases get applied uniformly at scale, eliminating the diversity of perspectives that would naturally exist with multiple decision-makers.

An image where each pixel's brightness represents how far away that object is from the camera.

Computing per-pixel distance from camera to scene surfaces to reconstruct 3D geometry from 2D images.

Assigning different amounts of model parameters to different layers based on their functional importance rather than uniformly.

Improving model performance by increasing computational depth without adding new parameters, achieved through layer reuse.

A technique that creates a larger model by combining and stitching together layers from smaller pre-trained models rather than training a new model from scratch.

A convolution operation that applies separate filters to each input channel, reducing parameters compared to standard convolution.

The process of restoring a compressed model's weights to higher numerical precision, improving quality but requiring more memory.

A new model created by modifying or fine-tuning an existing base model rather than training from scratch.

Derive is deterministic computation over available state; infer is LLM-mediated judgment under declared context and capability constraints.

A numerical representation that captures the visual characteristics around a detected keypoint, allowing the model to match similar points across different images.

Generating model weights using text or structured descriptions of the target architecture and task as input.

Systematic evaluation of different design choices to find optimal trade-offs between performance, power, and area.

The loss of professional expertise and judgment that occurs when workers rely on automated systems instead of developing their own capabilities.

A mathematical model that generates diverse sets of items by penalizing similarity, useful for ensuring variety in generated outputs.

Automated verification rules that produce the same result every time, used when there is clear evidence of task completion.

A guaranteed non-AI backup process that ensures the system can continue operating safely if AI components fail.

A system that executes predefined process steps in a fixed, repeatable sequence without adaptive reasoning.

Converting a game with hidden information into multiple fully-observable versions by randomly filling in unknown information.

The script used to write Marathi, Hindi, and several other Indian languages.

An early, pre-release version of a model used for testing and refinement before public release.

Explicit signals that alert an agent when its current reasoning or conclusions diverge from task requirements or ground truth.

Fine-grained, skillful robotic hand control requiring precise coordination of many joints.

The reasoning and explanation behind why a specific code location is likely buggy, not just the location itself.

AI process of identifying root causes or problems from observed symptoms.

Converting visual diagrams into executable or markup code, such as LaTeX or TikZ, automatically.

The patterns and interactions that emerge in conversation, including how participants exchange information and coordinate actions.

The process of an AI model creating natural conversational responses based on input text.

A compact storage mechanism that maintains compressed conversation state across multiple dialogue turns.

A metric measuring overlap between predicted and ground-truth segmentation masks, ranging from 0 to 1.

The process of finding a set of basis vectors (dictionary) that can reconstruct data through sparse combinations.

The ability to understand and apply code changes (diffs) to existing files rather than generating code from scratch.

A contiguous section of a code patch showing added, removed, or modified lines in a specific file.

The ability to compute gradients through a system, enabling end-to-end learning.

A property of operations that allows gradients to flow through them during backpropagation for model training.

Smooth mathematical function approximating non-differentiable operations for training.

A NAS method that relaxes discrete architecture choices into continuous variables to enable gradient-based optimization.

Logical reasoning operations (like applying rules) implemented as differentiable functions so gradients can flow through them.

A dynamical system whose equations can be differentiated to compute gradients, enabling gradient-based optimization.

Mathematical functions that measure how far a model's output is from desired behavior, designed to be optimizable via gradient descent.

A learnable memory retrieval mechanism that can be trained end-to-end to recall relevant past episodes for current decision-making.

A physics solver built into a neural network so that gradients can flow through physical laws during training.

A rendering system that allows gradients to flow backward from rendered images to 3D parameters, enabling optimization of visual properties.

A reward function whose gradients can be computed, allowing optimization of model outputs toward desired properties.

A reward function designed to be differentiable so gradients can flow through it during training.

A sparse attention method that supports gradient computation, enabling end-to-end training with learned sparsity patterns.

A list of possible medical conditions ranked by likelihood, used by clinicians to guide further testing.

Mathematical equations describing how systems change over time, naturally solved by analog hardware.

A mathematical framework that adds controlled noise to data to protect individual privacy while enabling statistical analysis.

A technique to systematically increase problem complexity to better differentiate model capabilities.

Adjusting problem complexity to match model capacity, ensuring training data is neither too easy nor too hard.

Predicting how hard a task is to automatically adjust the amount of computational effort needed.

An internal indicator that estimates how hard a problem is, used to guide model behavior.

Selecting training examples based on their difficulty level and diversity to improve model learning efficiency.

A passive optical component that uses diffraction to manipulate light without moving parts or power.

A training technique for diffusion models that uses per-position noise schedules to enable flexible inference-time control.

A language model that generates text iteratively by refining noisy predictions, allowing generation in arbitrary word order rather than strictly left-to-right.

Language models that generate text by iteratively refining noisy predictions into coherent words.

Generative model that creates images or videos by gradually removing noise from random data.

AI models that generate images by learning to reverse a noise-adding process, starting from pure noise.

A generative approach that iteratively refines predictions by gradually removing noise from random initial states.

A policy representation that uses diffusion models to generate action sequences from observations.

A learned distribution that guides diffusion models toward realistic outputs in a specific domain.

A generation method that iteratively refines outputs by gradually removing noise, rather than predicting tokens one at a time from left to right.

Iterations in a diffusion model that gradually refine noise into a final image or video output.

A transformer architecture adapted to work with diffusion-based generation processes.

A neural network design that generates outputs by iteratively refining noisy predictions into clear results, rather than building text one token at a time like traditional language models.

A method where a model generates text by iteratively refining noise into coherent output all at once, rather than predicting one word at a time.

A language model that generates text by iteratively predicting and refining masked (hidden) tokens across the entire output, rather than predicting one token at a time from left to right.

Using diffusion models to generate realistic robot motion sequences that can be used as training data.

A raster dataset representing terrain elevation, often outdated but useful as a geometric prior for reconstruction.

A virtual simulation model of a physical system used to predict behavior and test changes before real-world deployment.

A linguistic situation where two distinct language varieties coexist in a community, each used in different social contexts.

A convolutional operation that skips input elements to capture patterns at multiple scales without increasing parameters.

A technique to simplify high-dimensional parameter spaces by identifying and focusing on the most critical variables.

Directing different types of evidence or features to specialized processing paths based on their characteristics.

Evaluating a system across multiple independent criteria or dimensions rather than a single overall score.

Techniques that compress high-dimensional data into fewer dimensions while preserving important patterns.

A training technique that teaches a model to prefer certain outputs over others by learning from examples of better and worse responses.

Training method that aligns models with human preferences by directly optimizing the difference between preferred and dispreferred outputs.

A graph structure representing causal relationships where arrows point from causes to effects with no cycles.

A workflow representation where tasks are nodes and dependencies are directed edges with no circular paths.

A method for comparing texts that preserves the direction of change, showing how one text transforms another rather than just measuring similarity.

A probability distribution over probability distributions, used here to model uncertainty over class predictions.

A measure of smoothness on a graph that quantifies how much node values vary across connected edges.

A probability distribution used to smooth categorical data by adding prior counts to observed frequencies.

Clear, explicit communication to users that content contains paid promotion or advertising.

A quantum NLP framework that models language using grammar-aware categorical structures for text analysis.

The logical flow and consistency of ideas across sentences in a text or conversation.

Examining how language serves specific communicative purposes in conversation, like validating feelings or paraphrasing.

Words or phrases that signal relationships between ideas, like 'therefore' or 'for example'.

Small words or phrases like 'well' or 'kind of' that convey emotion, intention, and interpersonal meaning in conversation.

The challenge of moving from discovering causal rules to engineering them into working systems.

A generative model that iteratively removes noise from discrete tokens (like words) to generate text, as an alternative to autoregressive decoding.

Generative models that iteratively denoise discrete tokens (like words) from noise to produce text.

Compressed representations of audio data stored as specific, distinct values rather than continuous numbers, making them efficient for storage and processing.

A mathematical framework for defining calculus operations (gradient, curl, divergence) on discrete geometric structures like cell complexes.

A real-valued alternative to FFT that decomposes signals into cosine and sine components without complex numbers.

A compressed representation where continuous data is converted into distinct, countable tokens or categories.

A communication channel where each transmitted symbol is corrupted independently with no memory of past transmissions.

A mathematical space representing probability distributions over discrete categories, used for modeling sequences and graphs.

Individual units of quantized information that represent audio in a compressed, symbolic form rather than continuous values.

Converting continuous numerical values into discrete bins or categories for processing by algorithms.

The error introduced by approximating a continuous process with discrete steps, a key concern in sampling algorithms.

The ability of a model to generalize across different mesh resolutions or numerical discretizations of the same continuous problem.

A pattern in token gradients that effectively distinguishes high-reward responses from low-reward ones.

A learned representation designed to distinguish between different classes or outcomes, rather than reconstruct input.

Separating different factors of variation (like expression and identity) in a model's learned representations.

When a policy or algorithm produces unequal outcomes for protected groups, even if not intentionally discriminatory.

Simple heuristic strategies for assigning jobs to machines, like prioritizing shortest jobs first.

Automatically generating training labels using heuristics or patterns instead of manual annotation.

A smaller, faster version of BERT that retains most of its language understanding ability while using fewer parameters and less computational power.

A technique that compresses a large, complex model into a smaller one by training it to mimic the larger model's behavior, resulting in faster inference with minimal loss of quality.

A model that has been compressed by training a smaller model to mimic a larger, more capable model, reducing size and computational requirements while retaining performance.

A smaller, faster version of a larger model created by training it to mimic the larger model's behavior, reducing computational requirements while maintaining reasonable performance.

Irrelevant information in the input that doesn't help solve the task but can distract models from focusing on key evidence.

A harmful task split across multiple user accounts so each individual transcript appears benign.

Using multiple computers or servers across a network to share the computational work of training or running a model, rather than relying on a single machine.

Adjusting a model so its learned patterns match the actual distribution of data in a target domain.

A learned mechanism that controls which image features to modify based on their distance from the target feature distribution.

When the data distribution used for training differs from the distribution encountered during deployment, causing performance degradation.

Modifying a model's output probability distribution at inference time to satisfy constraints without changing the model's weights.

When a policy becomes overly specialized in reproducing successful behaviors without learning to handle diverse situations or recover from failures.

When a model encounters data that looks different from what it was trained on, causing performance to drop.

A guarantee that works regardless of the underlying data distribution, requiring no assumptions about how data is generated.

When a model's behavior diverges from the original training data distribution during fine-tuning or RL.

A mathematical space where words are represented as vectors based on their usage patterns in text, like GloVe or Word2Vec.

Ensuring benefits and harms are equitably distributed across agents rather than concentrated in hubs or privileged positions.

How well generated music matches the statistical distribution of real music across acoustic and structural features.

A systematic difference in how two groups (humans vs. LLMs) distribute their outputs across categories.

Forcing a model's output distribution to match a target distribution, here used to normalize reward structures across different tasks.

A measure comparing probability distributions rather than individual samples, useful for evaluating generative model quality.

Learning to predict probability distributions over outputs rather than single deterministic predictions.

When the statistical properties of data change over time, making old patterns unreliable for future predictions.

Finding solutions that work well across all possible data distributions within a defined uncertainty set.

A regularization technique that limits how far a model's distribution can drift from a reference distribution during training.

A penalty that prevents a model from changing its behavior too drastically by measuring the statistical distance between old and new policies.

The temporal pattern of prediction mismatches between teacher and student models across generation steps.

A mathematical property ensuring that a velocity field conserves mass (no fluid is created or destroyed at any point).

When a model trained with RL produces repetitive, similar outputs instead of varied responses, reducing usefulness.

A metric measuring the quality of unique answers generated relative to the best possible answer set of the same size.

A ranking method that prioritizes both relevance and variety, ensuring results cover different perspectives or approaches.

Breaking a large problem into smaller independent subproblems, solving each separately, then combining results.

Safety checks that DNA synthesis providers use to block orders for sequences that could be used to create dangerous pathogens.

The ability to identify and explain which retrieved documents contributed to a generated answer.

The natural division between separate documents used as a constraint to group tokens for shared expert selection.

The process of breaking long documents into smaller pieces before embedding them, which this model is optimized to work with effectively.

Automated checking of whether a document (like a contract) meets organizational policies and requirements.

The process of automatically pulling out specific structured information from documents, like names and dates from a resume.

Anchoring AI responses to specific source documents to ensure answers are based on provided content.

The ability to automatically extract, understand, and convert information from document images (like scans or forms) into structured, machine-readable formats.

The process of identifying and understanding the structure of a document, such as text regions, tables, and columns.

The process of automatically reading and extracting structured information like text, tables, and layout from documents.

Finding the relevant documents or passages from a large collection that are needed to answer a question.

The ability to maintain the original layout, formatting, and organization of a document when extracting text, rather than just outputting raw characters.

The ability to read and extract meaningful information from structured documents like receipts, invoices, and forms by recognizing both text and layout.

A task where a model reads a document image and answers natural language questions about its content by understanding both the visual layout and text.

Tasks that require processing, searching, and reasoning over large collections of documents to find answers.

Understanding and answering questions that require information from multiple parts of a full document.

The process of an AI agent consulting, reading, or creating documentation during software development tasks.

Training a model on data from multiple specialized fields (like general text, scientific papers, and medical literature) so it works well across all of them.

A specialized expert in an MoE model trained to handle reasoning or task-specific knowledge rather than raw perception.

Deep knowledge and experience in a specific field or function that helps users leverage AI tools effectively.

The difference between synthetic training data and real-world data that can hurt model performance.

Training models to work well on new, unseen domains beyond their training data.

A technique that automatically creates many fake domain names to evade detection and maintain control of malicious infrastructure.

How well a model's responses are anchored in accurate, specialized knowledge specific to a field rather than generic or hallucinated information.

Specialized expertise and facts about a particular field or subject area that an AI model has learned during training.

When a model encounters data from a different source or environment than it was trained on, causing performance to drop.

A model fine-tuned or trained specifically to handle a particular data distribution or task variant.

When a model is trained to excel at a specific task or set of languages rather than being a general-purpose tool.

Programming languages designed for specialized tasks in particular industries or fields.

A predefined categorization system that organizes data sources (e.g., code, web, books, academic papers) into distinct domains.

Training technique using adversarial objectives to make model representations invariant across different data domains.

A model that works effectively across many different subject areas and use cases without needing to be retrained for each one.

Abstract problem formulations that can be recognized and solved across multiple unrelated academic fields.

A model's ability to understand and respond accurately to topics within a specific field or area of expertise it was trained on.

A language model specialized and fine-tuned for a specific subject area (like chemistry) rather than trained as a general-purpose system.

An AI planning algorithm that solves problems in any domain without domain-specific customization.

Learned representations that remain useful across different data sources or conditions despite their differences.

A distillation approach that routes different domains to specialized teachers, then combines their knowledge into one model.

A model trained specifically on data and tasks from a particular field (in this case, chemistry) to achieve higher accuracy in that domain than general-purpose models.

Tailored or optimized for a particular field or type of content, such as news, reviews, or scientific writing.

A test designed to measure AI performance on specialized tasks within a particular field or industry.

Assessment tailored to a particular field (like law) using metrics and error types relevant to that domain.

Training a model on specialized data from a particular field (like medicine) so it becomes expert at tasks in that domain rather than being a generalist.

The ability to generate text tailored to a particular field or context, such as legal documents, Wikipedia articles, or product reviews.

Specialized expertise required for a particular field, like vendor-specific scanner operations in medical imaging.

Specialized vocabulary and terminology unique to a particular field or industry, like medical jargon in healthcare or mathematical notation in physics.

A language model trained exclusively on text from a particular field or subject area, making it much better at understanding and generating content in that domain than general-purpose models.

A language model trained specifically on data from one field (like biomedical research) rather than general internet text, making it excel at specialized tasks.

Training a model to excel at tasks within a particular field (like legal documents) rather than being a general-purpose model.

Training a model on specialized data from a particular field (like biomedical literature) rather than general internet text, making it much better at understanding that field's concepts.

Specialized workflows and methodologies unique to a particular field that require expert knowledge to execute correctly.

Training a model exclusively on data from a narrow domain (like Python code) rather than general text, making it highly specialized but less versatile.

Training or adapting a model to specialize in a particular field (like biomedicine) rather than performing equally well across all topics.

A model's understanding of specialized terms and concepts unique to a particular field, like medical terminology in biomedical text.

A mechanism property where bidders maximize profit by truthfully reporting their values, regardless of others' bids.

Velocity measurements from radar based on frequency shifts, indicating object motion relative to the sensor.

The ability to distinguish between different motion speeds based on frequency shifts in reflected radar signals.

A fine-tuning method that adapts model weights by separately learning magnitude and direction changes, extending LoRA.

A method of comparing two vectors by multiplying their components and summing the results, where vector magnitudes (length) affect the final score.

A causal inference method that remains valid if either the outcome model or treatment model is correct, providing robustness to misspecification.

A square matrix where all rows and columns sum to 1, used to represent valid probability distributions for mixing multiple streams.

Reducing an image's resolution by removing pixels, making it smaller and faster to process.

Converting coarse-resolution predictions (e.g., 25 km grid) to fine-resolution local predictions at specific locations.

A specialized AI model that receives requests routed to it by another system and performs the actual task or generates the final response.

A specific NLP application or problem that uses the output of a pre-trained model, such as classification, search, or similarity matching.

Specific applications or problems that use the output of a pretrained model, such as predicting protein structure or identifying protein function.

The smaller neural network component in speculative decoding that quickly generates candidate tokens before verification by the main model.

A smaller, faster model used in speculative decoding to quickly propose token sequences before a larger model verifies them.

A tree structure of multiple candidate token sequences proposed by a draft model, allowing parallel verification of multiple continuations.

A quantitative finance metric measuring the largest peak-to-trough decline in wallet balance and subsequent recovery.

Continuous tracking of whether a model's performance or fairness properties change over time as new data arrives.

A control policy that reduces the expected growth of queue sizes by prioritizing service to stabilize the system.

The process of automatically identifying and classifying different driving behaviors (e.g., aggressive vs. normal) from sensor data.

Forecasting how a patient or cell will react to a specific drug treatment based on molecular or clinical data.

A system with two separate neural networks—one that processes questions and one that processes documents—both converting their inputs into comparable vector embeddings.

The parallel development and deployment processes for machine learning models and traditional software components.

The danger that AI technology can be misused for harmful purposes despite benign original intent.

A model with separate encoders for two input modalities that map them into a shared embedding space.

Organizing information at two levels of detail: high-level task guidance and low-level step-by-step actions.

An approach that conditions a model on two complementary graph structures simultaneously to capture multiple constraints.

A system with separate memory components (Example and Reflection) that stores patterns and feedback to improve future decisions without model retraining.

Using two complementary memory systems to store and retrieve different types of information for decision-making.

Training models to process and align two different types of input data (like RGB and infrared images) simultaneously.

An approach combining two complementary methods—one for logical reasoning and one for learning patterns—to solve a problem better than either alone.

A single model trained to perform multiple distinct tasks, such as both text generation and embedding, rather than being specialized for just one.

Synthesizing an image that combines content structure from one reference image with visual style from another reference image.

A neural network design using two separate computation pathways for different functions.

An architecture using two parallel processing streams with different time scales—one dense and one sparse.

Knowledge or skills that can be applied in both harmful and benign contexts, requiring selective removal.

A minimal, non-functional model used for testing infrastructure and workflows without the computational cost of a real model.

Persistent memory that tracks an agent's progress, decisions, and learned strategies across multiple task attempts and sessions.

The ability to generate responses with a specific target length or speaking time.

A set of bets that guarantees profit against someone's probability assignments, revealing logical inconsistencies in their beliefs.

A social interaction or conversation between exactly two people.

Testing code by executing it to confirm whether identified weaknesses are actually exploitable.

An assessment approach that measures learning potential by observing how students respond to graduated hints and scaffolding.

Training approach that evaluates which skills remain helpful during learning and selectively retains only those that improve the current policy.

A setting where conditions, rules, or task requirements change over time rather than remaining static.

A formal system for reasoning about how beliefs and knowledge change when new information is revealed.

Building a network representation that changes over time to reflect evolving relationships, like road connectivity adjusted for traffic incidents.

Combining task-specific model parameters at inference time based on input features, rather than using a fixed merged model.

Automatically choosing the best execution approach (LLM reasoning, tool use, or code) for each step based on task requirements.

The ability to automatically adjust how many of a model's parameters are actively used based on available computational resources, allowing the same model to run efficiently on different hardware.