Papers

Yash Kulkarni, Shubham Harkare, Arvind Suresh Yogesh Babu

For LLM semantic caches, focus on validating answer quality before optimizing eviction policy; LFU is a solid default, and geometry-aware policies don't help much because newly cached queries rarely have similar neighbors.

This paper evaluates different cache eviction policies for semantic caches that store LLM responses based on query embeddings. Testing seven policies across multiple datasets and cache sizes, the authors find that LFU (Least Frequently Used) performs best, but no policy significantly outperforms it.

Jayjun Lee, Jessica Yin, Asif Rana et al.

Pre-training on a generic manipulation task and using a stable fine-tuning recipe with behavior cloning and conservative updates enables efficient transfer of dexterous skills to new tasks and real robots without catastrophic forgetting.

ADEPT is a reinforcement learning framework that trains dexterous robot hands to perform complex manipulation tasks by first learning a general object-handling skill, then adapting it to specific downstream tasks. The system transfers from simulation to real robots with multi-fingered hands (23-29 degrees of freedom) using vision and touch sensors, solving long-horizon tasks at human-level speed.

Zhu Zhang, Jixun Wang, Xiaoang Xu et al.

When training smaller models on long-context tasks, combining token-level teacher guidance with task-level verifier rewards—by measuring and distributing their disagreement—works better than either signal alone.

This paper addresses a key problem in training smaller language models for long-context tasks: teacher models give token-level guidance that looks locally correct but misses global evidence or violates task constraints.

Sotirios P. Chatzis, Loukas Papadoulas

Lévy Attention replaces softmax with a probabilistic formulation that automatically outputs calibrated uncertainty estimates alongside predictions—no extra parameters or passes needed, making it practical for high-stakes applications like patient risk ranking.

This paper introduces Lévy Attention, a new attention mechanism for time series that predicts both values and uncertainty in a single pass. Instead of using softmax, it formulates attention as a stochastic integral over a Poisson random measure, which naturally captures prediction confidence through two signals: disagreement (value spread) and evidence (compatibility mass).

Tate Berenbaum, Muthaiah Venkatachalam

Pipeline parallelism with pre-compiled shards lets you run 70B models on clusters of consumer AI PCs at interactive speeds by splitting computation across machines and optimizing GPU utilization through careful kernel fusion and speculative decoding.

This paper shows how to run large language models (like Llama 70B) across multiple Intel AI PCs by splitting the model into layers and running each layer on a different machine. Using OpenVINO optimization, speculative decoding, and request interleaving, they achieve near-optimal speed while serving multiple users simultaneously on modest hardware.

Zhenyao Cui, Siyuan Kan, Dingkun Liu et al.

Brain-to-image decoding can work with far fewer repetitions by anchoring both neural and visual representations to a shared reference point, rather than just denoising the query alone.

This paper tackles brain-to-image retrieval with limited neural data. Current methods require averaging 80+ brain scans per image, but the authors show that low-repetition queries fail not just due to noise, but because brain and image representations misalign.

Jiaming Li

When you have a fast but imperfect analytical model and limited expensive simulation data, teaching a neural network to correct the analytical model's errors—or pre-training it on the analytical model first—can cut your data requirements dramatically.

This paper shows how to make machine learning models more data-efficient by combining cheap analytical equations with expensive high-fidelity simulations. Using Helmholtz resonators as a test case, the authors demonstrate two approaches: learning to correct analytical predictions, or distilling analytical knowledge into a neural network before fine-tuning with limited simulation data.

Reza Bayat, Ali Behrouz, Vahab Mirrokni et al.

Gradually expanding memory capacity during sequence processing—rather than using static memory—is a simple, cost-free improvement that works across different memory-based architectures and shows larger gains on longer contexts.

This paper introduces Proteus, a technique that improves how neural networks handle long text by gradually expanding memory capacity as more context arrives.

Ben Anson, Conor Houghton, Edward Milsom

Muon can now be used effectively with LoRA-style low-rank fine-tuning by approximating its solution mathematically, offering a practical way to combine two previously incompatible techniques.

This paper solves a mathematical incompatibility between Muon (an optimizer that works well for pretraining) and LoRA (a popular parameter-efficient fine-tuning method). The authors propose sMuon, which approximates Muon's benefits in the low-rank setting through linearization and least-squares, using only standard matrix operations.

Abhishek Shukla, Ankur Sinha, Faiz Hamid

By formulating architecture search as a linear program using gradient and curvature information, LP-NAS finds better neural network designs 2-3x faster than standard differentiable NAS methods while achieving higher accuracy.

This paper proposes LP-NAS, a neural architecture search method that uses linear programming to find better network designs faster. Instead of randomly exploring architectures, LP-NAS uses mathematical optimization principles to guide the search, resulting in architectures that generalize better and are found more quickly than existing methods like DARTS.

Ajith Anil Meera, Pablo Lanillos, Wouter Kouw

Expected Free Energy provides a single, tunable objective that lets robots efficiently explore unknown environments by balancing information gathering with resource constraints—useful for real-world autonomous missions.

This paper applies Expected Free Energy from active inference to robotic path planning for Mars exploration. The robot must balance two goals: building accurate maps and finding high-value regions (like water sources) while staying within travel and measurement budgets. The approach outperforms traditional information-seeking methods by unifying both objectives under one principled framework.

Ziyang Luo, Zhongyao Chu, Xinjie He et al.

You can improve a frozen language model's reasoning and confidence detection in one forward pass by reconstructing clean internal states after steering, rather than running separate passes or accepting interference between techniques.

This paper solves a key problem in frozen language models: they both misuse their internal knowledge and fail to recognize when they lack sufficient information to answer.

Weihan Meng, Hongzhu Guo, Yi Jing et al.

You can teach an LLM to explain what individual features in another LLM do by injecting those features and training it to verbalize their meaning—this works across different SAE dictionaries and even different models.

SAEVerbalizer trains language models to generate natural-language explanations for features extracted by sparse autoencoders (SAEs). Instead of observing model behavior to understand what features do, the system directly injects feature directions into the model and fine-tunes it to produce explanations, making the process faster and more scalable.

Tianyi Li, Yaxin Luo, Xinyi Shang et al.

DARTree achieves up to 9.73× speedup on language model inference by constructing and scoring candidate token trees in parallel, accepting nearly 13 tokens per verification step—nearly 3× more than prior methods.

DARTree speeds up language model inference by using a tree-based speculative decoding approach. Instead of generating one draft token at a time, it predicts multiple token candidates in parallel using diffusion, then verifies them together. The method uses an autoregressive correction head to score entire branches at once, accepting more tokens per verification round than existing methods.

Martin J. Wainwright

UGC provides a principled way to design adaptive masking schedules for discrete diffusion that are provably near-optimal, with potential √d speedups over fixed schedules by concentrating computational effort where data geometry demands it.

This paper introduces unmasking growth complexity (UGC), a geometric measure that controls how discrete diffusion models should reveal information during sampling. The key insight is that UGC increments directly bound discretization error, enabling optimized sampling schedules that adapt to data structure.

Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina et al.

You can build competitive frontier-level language models at 1B parameters using only openly licensed data, making it feasible for researchers and organizations to develop ethical AI without relying on scraped or restricted datasets.

Mimir v1 is a 1-billion-parameter language model trained entirely on permissible (legally and ethically sourced) data that achieves competitive performance with much larger models. It uses a Hierarchical Reasoning Model architecture and excels at English, math, code, and Danish tasks—showing that high-quality open-source models don't require massive proprietary datasets.

Yuchen Xin, Zhihua Zhang

For sampling from nonsmooth targets, tracking local curvature via the active trace gives tighter convergence guarantees than worst-case dimension-dependent bounds, achieving O(ε⁻²) instead of O(ε⁻³) for common structured problems.

This paper analyzes the Moreau-Yosida unadjusted Langevin algorithm (MYULA) for sampling from nonsmooth distributions. The key contribution is showing that the algorithm's convergence depends on an 'active trace' quantity (measuring local curvature) rather than global dimension, leading to improved complexity bounds—especially for structured penalties like lasso and total variation.

Zixuan Lan, Yanhong Li, Jiawei Zhou

You can speed up LLM inference by skipping less important values in matrix operations—attention layers benefit most from this, while MLPs are harder to compress—and the speedups scale better on longer inputs.

This paper introduces Reduced Matrix Multiplication (RMM), a training-free method that speeds up language model inference by selectively removing less important numerical values from matrix multiplications. The approach works without changing model weights and maintains accuracy while reducing computation, with benefits increasing at longer sequence lengths.

Aimilios Hadjiliasi, Louis Nisiotis

Small language models can efficiently run agent cognition (thinking and memory) on edge devices like Jetson boards, enabling virtual agents to operate independently in real-time without cloud latency.

This paper explores how small language models (SLMs) running on edge devices can power the cognitive processes of virtual agents in immersive worlds.

Muhammad Hannan Akram, Muhammad Abubakar Rashid, Wassi Haider Kabir et al.

Heterogeneous AI agents can efficiently synchronize their understanding in distributed networks by translating belief updates through edge-deployed models, without requiring shared training or identical architectures.

This paper proposes a framework for AI agents in 6G networks to synchronize their understanding (beliefs) despite using different AI models and operating under different constraints.

Cheng Qian, Wenting Zhao, Liangwei Yang et al.

You can transfer capabilities from strong to weak models at inference time through clever prompt engineering and code scaffolding, achieving similar gains to training-time distillation without any parameter updates.

This paper shows that stronger AI models can help weaker ones solve tasks better at test time by building 'harnesses'—structured inference-time scaffolds like code templates and routing logic—without retraining. Testing on reasoning benchmarks, the approach nearly doubles weaker model performance by offloading unstable reasoning into deterministic code and enforcing strict output formats.

Rafi Ibn Sultan, Chengyin Li, Yiannos Demetriou et al.

Combining neighborhood attention mechanisms with uncertainty-weighted loss functions enables accurate segmentation of tiny, low-contrast anatomical structures even with limited training data—a practical approach for clinical imaging tasks where manual annotation is expensive and ambiguous.

This paper presents NA-UNETR, a transformer-based deep learning model for segmenting the Left Anterior Descending artery in CT scans.

Junliang Liu, Ruoyu Li, Wenxin Tang et al.

LLM agents can be hijacked to waste computational resources through carefully crafted skills that appear legitimate but steer agents toward costly execution paths—a threat that bypasses traditional task-completion checks.

This paper reveals a new attack on LLM agents that manipulates how they choose and use third-party skills. By crafting deceptive skill descriptions and instructions, attackers can trick agents into taking unnecessarily expensive detours—using more tokens and time—while still completing tasks correctly.

Pedro Sousa, Will Tebbutt, Sadiq Jaffer et al.

Foundation models trained on satellite imagery can capture persistent surface properties (terrain, vegetation, water) that explain local weather variations better than hand-crafted features, enabling more accurate probabilistic weather predictions at arbitrary locations.

This paper shows that Earth observation embeddings from satellite data can improve weather downscaling—predicting local weather at specific locations from coarse global weather models.

Junyi Ye, Ivy Gateri Wanjiku

For financial forecasting models deployed at 4-bit precision, activation calibration method (how you estimate value ranges from past data) is critical—percentile calibration substantially outperforms standard approaches, but 8-bit activations remain the safest choice when accuracy matters most.

This paper studies how to properly calibrate neural networks for low-precision inference in financial forecasting. When deploying models with 4-bit quantization, the choice of how to estimate activation ranges from historical data significantly impacts prediction accuracy—more so than at 8 bits.

Nikolai Bolik, Lennart Stöpler, Artur Andrzejak

Sparse autoencoders don't compose semantically the way we might hope: their active feature sets track internal model structure rather than human conceptual categories, limiting their usefulness for understanding what language models actually represent.

This paper examines whether sparse autoencoders (SAEs) in language models can capture human-like semantic categories by analyzing which SAE features activate together.

Zetao Hong, Song Yuan, Yuanhao Ding et al.

When running multiple RL training pipelines on shared inference hardware, intelligent scheduling of which requests to accept and how to allocate GPU memory between them can dramatically improve efficiency without compromising training quality.

This paper addresses a scheduling problem in LLM inference services that run multiple types of reinforcement learning workloads simultaneously. Different RL approaches (RLVR, RLHF, agentic) have different memory and timing needs, causing them to compete inefficiently for GPU cache.

Kaan Buyukkalayci, Kyle Pak, Merve Karakas et al.

A recommendation-system framework can efficiently select the best subset of sensors for tracking by learning from acoustic features, outperforming simpler signal-strength methods while maintaining real-time performance.

This paper tackles sensor selection for tracking systems by using a recommendation-system approach. Instead of relying on simple signal strength measurements that fail under acoustic interference, the method uses frequency-based audio features and a two-tower neural network to intelligently choose which expensive sensors (like cameras) to activate.

Laurens Samson, Iva Gornishka, Gossa Lô et al.

When deploying LLMs in government, you can't optimize for everything—better factuality and honesty require accepting higher costs and energy use, and bias reduction requires separate attention regardless of model quality.

This paper presents a framework for evaluating large language models specifically for Dutch government use, identifying six key dimensions (factuality, honesty, bias, energy use, cost, and data transparency) and testing 30+ models. The research reveals important trade-offs: better accuracy comes with higher environmental and financial costs, while bias operates independently of both.

Björn Engdahl, Adrian Kosowski, Jan Chorowski et al.

Recurrent latent reasoning (thinking without verbalizing) combined with in-context learning enables smaller models to solve complex reasoning tasks more cost-efficiently than larger models that explain their work.

BDH-CQ is a reasoning model that learns from examples at inference time and solves problems through hidden, iterative computation rather than step-by-step explanations. On the ARC-AGI benchmark, a small 150M-parameter version achieves 29.5% accuracy while costing just $0.0007 per task—better cost-efficiency than previous approaches.

Boning Li, Yu Chen, Longbo Huang

You can stop evaluating agents the moment you have statistical proof one is better, rather than running predetermined game counts—this cuts evaluation costs dramatically while keeping your confidence level mathematically sound.

This paper makes agent evaluation cheaper and faster by combining variance reduction (AIVAT) with statistically valid early stopping (Confidence Sequences). Instead of running a fixed number of games, the method stops as soon as there's enough evidence to declare one agent stronger, reducing evaluation costs by up to 74x while maintaining statistical guarantees.

Daniel Paulin, Victor Elvira

VARMA models are now computationally practical for high-dimensional time series because optimization cost no longer scales with series length—enabling better forecasts than VAR models on real data.

This paper solves a long-standing computational problem in time series modeling: estimating VARMA models (which capture dependencies efficiently) at scale. The key innovation is making each optimization step independent of series length by using Fourier-based sufficient statistics and a reparametrization that ensures valid models by construction.

Yijiang Li, Bingyang Wang, Yijun Liang et al.

You can improve LLMs through self-distillation using only the model's own outputs and internal consistency, without needing ground-truth labels or external feedback—making it truly self-supervised post-training.

This paper introduces Unsupervised On-Policy Self-Distillation (U-OPSD), a method that improves language models without requiring external supervision like ground-truth answers or feedback from larger models.

Mutasim Fuad Sarker, Adiba Rahman Namira, Wafa Binte Alam et al.

Quantum-enhanced neural networks can match or exceed classical models for medical time-series prediction while using 100x fewer parameters—suggesting quantum computing may enable efficient clinical AI systems.

QuanTiMedAI combines quantum computing with AI agents to predict mortality in cardiac arrest patients using ICU time-series data. The system uses an LLM agent to select relevant clinical features and a quantum recurrent network to model how patient conditions change over time, achieving better accuracy with far fewer parameters than traditional approaches.

Johann Birnick, Rayan Saab

You can now quantize neural networks with better geometric information (two-sided Hessian) at roughly the same computational cost as simpler one-sided methods, enabling more accurate model compression.

BaKron speeds up neural network quantization by efficiently using two-sided Hessian information to round weights more intelligently. Instead of the expensive O(m²n²) computation, it achieves O(mn(m+n)) work while maintaining the same cubic scaling as existing methods, making it practical to apply richer curvature information during quantization.

Saad Ahmed, Md Khalid Syfullaha

You can build accurate, deployable sign language recognition systems from scratch without pretrained models—this lightweight attention-based approach matches heavy ImageNet models while using 8.5-68x fewer parameters and fitting in 0.48 MB on phones.

This paper introduces RSBdSL38, an expert-validated dataset of 10,874 images covering all 38 Bangla Sign Language hand signs, and a lightweight neural network (298K parameters) that achieves 96.37% accuracy while running on smartphones.

ZhiYan Hou, Xinyu Tang, Hongyan An et al.

Adaptive weighting of token-level supervision based on divergence sequences improves reasoning model training—the timing and context of prediction errors matter more than their magnitude alone.

This paper improves on-policy self-distillation for reasoning models by making supervision weights adaptive to how prediction errors evolve during generation. Instead of treating all mismatches between teacher and student equally, DASH uses gates to weight supervision based on local divergence patterns, improving math reasoning across model sizes without extra computation.

Sam Levang, Fran Bartolic, Ty Dickinson et al.

You can now configure forecast timesteps after training instead of before, letting one model handle both detailed short-range and stable long-range weather predictions without retraining.

GEM-3 is a weather forecasting model that solves a key problem: existing models must choose between short timesteps (detailed but error-prone) or long timesteps (stable but missing short-term details). This model lets you pick the timestep at inference time with a single trained model, and training on mixed timesteps makes forecasts more stable across longer periods.

Yixiong Xiao, Congxi Xiao, Jingbo Zhou

RAG can boost time series forecasting by retrieving similar past sequences and intelligently fusing them into the model, rather than just appending them like in language models.

This paper adapts retrieval-augmented generation (RAG) to time series forecasting by retrieving similar historical sequences and using special tokens to blend them with input data.

Anton Conrad, Rustam Isaev, Denis Belomestny et al.

Localized conformal prediction can now be used with formal guarantees that it provides accurate uncertainty estimates for individual test points, not just on average, with clear bounds showing how localization bandwidth and calibration size affect performance.

This paper improves conformal prediction—a method that adds uncertainty bands to any AI model—by making those bands more accurate for specific data points rather than just on average.

Aniri, Jinhe Bi, Peng Liao et al.

Modality imbalance (text drowning out vision) is a real bottleneck in multimodal reasoning. By detecting when visual input is being ignored and training only on those critical tokens, you can make self-distillation much more effective.

This paper identifies and addresses modality imbalance in multimodal language models—where text dominates over visual information during reasoning. OPD-V uses positive and negative teacher models with modified images to detect when the model isn't properly using visual input, then selectively applies self-distillation only on tokens where visual information matters most.

Adel Javanmard, David P. Woodruff, Vahab Mirrokni

SSTQ enables federated learning systems to compress communications to logarithmic bit-widths while guaranteeing local differential privacy—a significant improvement over prior methods that suffered from dimension-dependent variance.

This paper introduces SSTQ, a privacy-preserving vector quantization method for distributed learning that compresses high-dimensional data into fewer bits while protecting individual privacy.

Paritosh Parmar, Landy Lan, Hong Yang et al.

Motion-based reasoning (tracking how bodies move) is more efficient and generalizable than pose-based reasoning for incident detection, and knowledge distillation can compress this understanding into models small enough for real-world deployment.

This paper tackles classroom safety monitoring using privacy-preserving computer vision. The authors create a hybrid benchmark mixing synthetic and real classroom data, then propose a lightweight motion-reasoning model that captures how incidents differ in movement patterns (speed, direction, acceleration) rather than just body poses.

Yang Yang, Qinyu Zhao, Mouxiang Chen et al.

By sharing backbone parameters across parallel branches with task-specific computation allocation, you can improve multimodal model performance without increasing model size or inference latency.

ParVL introduces a framework for scaling multimodal AI models by running multiple parallel vision and language processing branches that share the same core parameters. Instead of making models bigger or slower, it reuses existing components more efficiently and lets different tasks use different amounts of vision vs.

Christopher Schröder, Lukas Gienapp, Ferdinand Schlatt et al.

ALiBi positional encoding silently fails on long sequences due to numerical precision issues, degrading retrieval performance; using log-scaled distances instead of linear scaling provides the most reliable fix.

This paper discovers that ALiBi positional encoding—a popular method for handling long sequences in transformers—has a critical flaw: its linear bias scaling causes floating-point underflow, making many attention weights zero and blinding attention heads. The authors analyze this problem, test fixes, and show it hurts token retrieval tasks while barely affecting standard benchmarks.

Mirac Suzgun, James Zou, Stuart M. Shieber et al.

This tool makes string algorithms transparent and inspectable: you can see not just the score but the exact reasoning (alignment paths, metric matches) behind it, all running fast in your browser without sending data anywhere.

string2string Studio is an interactive web-based tool for analyzing string-to-string algorithms across NLP, biology, and humanities. It runs locally in your browser (no data upload needed), offers six analysis modules, and shows you exactly how algorithms work by displaying their evidence—alignments, edit paths, and search results—making it easy to debug and compare methods.

Yordan Raykov, Rodrigo Veiga

By treating forward-policy training as a geometric optimization problem, you can leverage the target distribution's structure (factorization, locality) to compute better gradient updates, leading to faster convergence and better exploration in GFlowNets.

This paper reformulates how to train the forward policy in GFlowNets (a framework for sampling from complex distributions) using information geometry.

Hakan Ferhatosmanoglu, Kushal Kumar, Tal Wagner et al.

Predict recall progression curves per query before searching, then dynamically adjust search depth based on predictions—this eliminates expensive parameter tuning and achieves target recall with 80% fewer data accesses.

QASP predicts how recall improves as vector search explores more data, then uses that prediction to set search depth for any target recall level. Instead of tuning fixed parameters per dataset or running multiple searches, it learns once upfront and adapts per query, reducing computational cost while maintaining consistent quality across different queries.

Jiajia Tang, Sizhe Yuen, Francisco Gomez Medina et al.

Organizing tasks into separate optimization paths with independent adapters outperforms using a single shared adapter—even with the same total parameters—because it reduces task interference and preserves positive transfer between compatible tasks.

This paper addresses a key problem in fine-tuning large language models: when you try to adapt a single shared adapter to many different tasks, they interfere with each other, hurting performance. The authors propose automatically organizing tasks into groups and sequences, then training separate adapters for compatible tasks instead of forcing everything through one shared adapter.

Yao Xiao, Reuben Tan, Zhen Zhu et al.

A single trainable embedding can dramatically improve how vision-language models handle long visual context by acting as a retrieval target—selecting sparse, relevant tokens instead of processing everything, making it practical for real-world deployment.

ReToken is a lightweight technique that adds a single learnable token to vision-language models to intelligently select relevant visual information from long images and videos.

Md. Mehrab Hossain Opi, Robiul Islam Ryad, Md. Umar Faruk

Instead of using the same precision everywhere, MixFrag identifies which transformer layers are most fragile to quantization and protects them with higher precision while using lower precision elsewhere—delivering better accuracy under the same memory budget.

MixFrag is a method for compressing Vision Transformers by assigning different precision levels (bit-widths) to different layers based on how sensitive each layer is to quantization. It measures this sensitivity using KL divergence and solves an optimization problem to allocate bits efficiently, achieving strong results on image classification and object detection tasks.

Jiawei Xu, Minghui Liu, Juzheng Zhang et al.

By parameterizing the KL penalty weight (β) in self-distillation, you can trade off stability and teacher guidance more flexibly, and implement the optimal policy through simple logit mixing rather than expensive RL optimization.

This paper improves on-policy self-distillation for reasoning models by introducing β-OPSD, which treats the KL penalty weight as a tunable parameter rather than a fixed value. The method derives an optimal policy as a geometric blend between a reference policy and a teacher model, then implements this efficiently through logit mixing during distillation.

Iliya Mirzaei

When building with LLMs, don't assume fancy self-improvement techniques help—repeated sampling at the same token cost is a stronger baseline, and self-inspection methods often fail to improve over it.

This paper compares self-improvement methods (like self-critique, reflection, and debate) against simple repeated sampling at equal token budgets across models from 1.5B to 7B parameters. The key finding: repeated sampling consistently matches or beats elaborate self-inspection methods, and self-inspection methods actually get worse as models scale.

Woongkyu Lee, Jungwook Choi

Inference-time scaling helps frontier models but doesn't reliably improve local computer-use agents; success requires failure-aware design and selective compute allocation rather than just adding more computation.

This paper studies how to make computer-use agents (AI systems that control computers) work better on local devices with limited hardware. The researchers test different ways to use extra computation during execution—like remembering more history, planning longer, or trying multiple approaches in parallel—on smaller AI models.

Ruman Wang, Hangting Ye

LLMs can transfer medical knowledge into auditable, locally-executable feature programs instead of making direct predictions—improving data efficiency, privacy, and reproducibility in medical image classification.

ScaFE uses an LLM to generate executable programs that measure clinical scar features from photos, rather than asking the model to diagnose directly. These programs run locally, protecting patient data, and feed structured features into a simple Random Forest classifier.

Ali Rayat, Yunhao Fan, Gia-Wei Chern

GNNs can replace expensive electronic calculations for simulating spin dynamics in magnets by learning effective magnetic force fields, similar to how machine-learned potentials work for atomic systems.

Researchers developed a graph neural network framework that learns to predict magnetic forces in metallic magnets directly from electronic calculations. This approach eliminates expensive repeated electronic simulations during time evolution, enabling fast and accurate predictions of spin dynamics across different magnetic structures.

Alexander Boesgaard Lorup

Key-value cache precision and structure are the primary drivers of output divergence in language models, not the full decoder state—meaning you can replay intermediate reasoning stages reliably by preserving and transplanting caches across different precision levels.

This paper investigates why AI model outputs diverge when using cached intermediate states versus recomputing them from scratch. By systematically swapping key-value caches between different computation paths in a large language model, the authors show that cache precision and structure—not the full decoder state—primarily determines output differences.

Antonio Delgado-Rosa, David Muñoz-Valero, Enrique Adrian Villarrubia-Martin et al.

On-device inference on resource-constrained satellites combined with synthetic data generation can enable autonomous, real-time aircraft detection without overwhelming downlink capacity—shifting satellites from passive data collectors to active decision-makers.

This paper tackles satellite-based aircraft detection by running AI inference directly on small satellites (CubeSats) instead of sending raw images to Earth, and uses AI-generated synthetic images to train better detection models. The approach balances limited satellite bandwidth with scarce training data, enabling real-time autonomous surveillance from orbit.

Zechen Zhang, Rui Peng Li, Yousef Saad

GNNs can learn better preconditioners for sparse linear systems by adopting classical multigrid hierarchies as structural priors, but multilevel structure doesn't always outperform simpler single-level learned approaches—the tradeoff depends on the problem type.

This paper proposes Graph Neural Multilevel Preconditioners (GMP), which combines classical algebraic multigrid structure with graph neural networks to speed up solving large sparse linear systems. The method learns how to smooth, restrict, and interpolate data across multiple levels while maintaining compatibility with standard iterative solvers, tested on over 800 real sparse matrices.

Cheng Siong Chin, Jianhua Zhang, Mohan Venkateshkumar

Autonomous systems need to know they're aging: by continuously monitoring hardware health and adapting AI decisions in response, robots can gracefully degrade rather than suddenly fail, especially critical for systems you can't easily repair or replace.

This paper proposes Aging-Aware Autonomous Intelligence (AAAI), a framework that enables autonomous systems to monitor their own hardware degradation and adapt their behavior accordingly.

Yecheng Wu, Song Han, Han Cai

You can now use different models to generate training data and provide distillation supervision without performance loss, by filtering out style differences rather than requiring teacher consistency.

This paper addresses a practical problem in training large reasoning models: when the teacher model used for distillation differs from the one that generated the training data, performance often suffers.

Perry Dong, Ron Polonsky, Dorsa Sadigh et al.

When fine-tuning RL policies, pretraining Q-functions on offline data often doesn't help; instead, initializing Q-functions from diverse policy rollouts works better and is simpler.

This paper challenges the conventional wisdom that Q-functions should be pretrained for online RL fine-tuning. The authors show that naive Q-function pretraining often doesn't help because it targets the wrong policy, then propose IPE—a method that uses diverse pretrained policies to bootstrap Q-function learning, achieving 1.26x better performance than standard pretraining.

Shady E. Ahmed, Panos Stinis

You can solve regression problems by storing local patterns and using similarity matching instead of training large global models—this is faster, uses less memory, and works well for scientific data that repeats in certain regions.

This paper proposes a novel regression method inspired by how fruitflies sense their environment. Instead of building complex global models, the approach stores a library of representative local patterns and makes predictions by finding similar patterns to a query and combining their responses.

Ji Xin, Xiao Xiao, Ishan Bhatt et al.

By generating implicit user intents to expand search results, e-commerce platforms can surface complementary and substitute products while maintaining relevance, improving both user satisfaction and visibility for long-tail inventory.

This paper presents a system that improves product discovery in e-commerce by generating related search intents beyond exact query matches. Using large language models for popular queries and fine-tuned smaller models for niche queries, the system expands what products users see while keeping results relevant—increasing discovery coverage from 60% to 80% at lower computational cost.

Roshan Kenia, Stephanie L McNamara, William Lotter

You can efficiently adapt pre-trained medical imaging models to understand anatomy-level details by decomposing scans into regions, learning how they relate to each other, and aligning them with medical text—without retraining from scratch.

This paper presents Anatomy Contextualized Adaptation (ACA), a lightweight method that improves CT scan foundation models by aligning fine-grained anatomical regions with text descriptions from radiology reports while preserving whole-scan context.

Haolei Xu, Xiaowen Xu, Haiwen Hong et al.

When training smaller models to reason, letting a teacher briefly take over at critical failure points—rather than forcing the student to learn from its own mistakes—significantly improves both efficiency and accuracy.

This paper addresses a key failure mode in on-policy distillation where students get stuck in wrong reasoning paths. The authors propose Relay-OPD, which detects when a student diverges from the teacher's approach and temporarily hands off to the teacher to correct course, then resumes student training.

Sungjae Park, Shubham Tulsiani

Real-time reactive robot control is possible with diffusion policies by decoupling fast proprioceptive feedback from slower visual processing and adapting the denoising schedule to hardware latency—enabling closed-loop replanning at 25Hz instead of 6Hz.

This paper introduces πR², a method that makes diffusion-based robot manipulation policies reactive and fast enough for real-time control.

Tom Saliencro, Rohan Desai, Priya Nair et al.

You can route tokens to variable numbers of experts based on per-token confidence signals from the router itself, improving efficiency and performance without adding parameters or complexity.

This paper proposes CARE, a smarter way to route tokens through expert networks in fine-tuned language models. Instead of sending every token to the same fixed number of experts, CARE uses the model's own uncertainty signals to decide how many experts each token needs—easy tokens get fewer experts, hard tokens get more.

Solomon Micheal Serunjogi, Rachmad Vidya Wicaksana Putra, Ayat Taha et al.

Using spatial modes instead of wavelengths in photonic accelerators enables simpler, more efficient transformer inference with a single laser source and no spectral filtering, making photonic acceleration more practical for real-world deployment.

MDTransformer is a photonic hardware accelerator for transformer models that uses spatial light modes instead of wavelengths to perform matrix operations.

Yuan Yin, Elias Ramzi, Marc Lafon et al.

Training driving policies directly from egocentric camera images at scale is now practical—eliminating the need for privileged observations during training while maintaining competitive performance and improving real-world transfer.

Pictura is a GPU-accelerated driving simulator that renders camera views for each agent in real-time, enabling self-play training of driving policies directly from images rather than privileged vectorized data.

Neta Shaul, Chao Liu, Arash Vahdat et al.

PDD enables fast video/image generation (4-8 steps instead of 50+) with simpler training than current methods, while actually improving output diversity—making it practical for scaling generative models to production.

This paper introduces Parallel Decoding Distillation (PDD), a method to speed up image and video generation from diffusion models by predicting multiple denoising steps at once instead of one at a time.

Wenzhi Zhong, Edward Milsom, Michael Murray

Using spectral norm geometry in SAM's perturbation step combined with Muon optimization leads to better generalization on vision models—a practical improvement for training robust neural networks.

This paper improves Sharpness-Aware Minimization (SAM), a training technique that helps models generalize better, by using matrix-aware geometry. The authors propose using spectral norm-based perturbations for hidden-layer weights and combine this with the Muon optimizer, achieving better results on ImageNet with ViT and ResNet models.

Bingnan Li, Haozhe Wang, Haozhong Xiong et al.

When distilling guided diffusion models, directly matching teacher and student predictions can backfire if they have asymmetric access to information—you need to supervise different prediction branches separately to avoid antagonistic error dynamics.

This paper identifies a critical flaw in how diffusion models are distilled when using classifier-free guidance (CFG). The authors show that naive matching of teacher and student predictions can fail when the teacher has access to information the student doesn't, causing errors in different prediction branches to work against each other.

Zhen Huang, Yikun Wang, Shijie Xia et al.

Instead of applying uniform data processing rules, adapting the cleaning strategy per example—deciding what operation each piece of data needs—improves LLM pretraining efficiency and downstream performance.

DataOrchestra is a framework that customizes data processing for each example in pretraining, rather than applying one fixed strategy to all data. An orchestrator decides whether to drop, keep, or clean each data chunk, and if cleaning is needed, selects specific operations like editing or rewriting.

Jhonatan Tavori, Gur-Eyal Sela, Ion Stoica et al.

Distributed inference pipelines that merge fast and slow predictions are vulnerable to network-based denial-of-deadline attacks that don't require model access—attackers can degrade accuracy by simply congesting shared resources to miss latency deadlines.

This paper reveals a new vulnerability in distributed inference systems that combine fast local predictions with slower, more accurate remote predictions. Attackers can craft burst-shaped workloads that congest the network, delaying remote predictions past their deadlines so they get discarded.

Maruthi Vemula, Neeraj Praneeth Gajula

Memory eviction in LLMs can be viewed as estimating whether tokens will be reused by observing actual model attention over a short window, but this 'measuring' approach only beats simple attention-based methods when token reuse is sharp and predictable—which standard benchmarks don't exhibit.

This paper reframes the problem of deciding which items to keep in a language model's bounded memory as an estimation task. Rather than deciding immediately (like StreamingLLM) or guessing the future (like SnapKV), the authors propose fixed-lag smoothing: wait a few steps, observe what the model actually attends to, then decide what to keep.

Peng Zhao

Early-stopped negative-shifted gradient descent escapes fundamental limits of negative ridge regularization by creating adaptive, mixed-sign filters that improve risk by polynomial factors while recovering multiple scales simultaneously.

This paper studies how early stopping in gradient descent with a shifted learning rate can improve regularization in overparameterized linear regression. Unlike standard negative ridge penalties that have structural limits, the proposed approach creates mixed-sign filters that can better handle weak spectral directions, achieving polynomial improvements in risk under certain conditions.

Ritik Raj, Souvik Kundu, Sarbartha Banerjee et al.

For agentic workflows, route once per task using task-level feedback instead of routing each LLM call independently—this lets the system learn which models work best for different task types while optimizing accuracy-latency trade-offs.

TRACE-Router is a routing system that assigns agentic AI tasks to different LLMs once at the start, rather than deciding per-call. It learns which model works best for each task type using delayed feedback from task completion, balancing accuracy and speed. This approach outperforms existing routers by aligning decisions with actual task outcomes.

Yijun Liang, Yunjie Tian, Yijiang Li et al.

VCSD achieves significant improvements in vision-language model performance (up to 5% on benchmarks) by using contrastive image conditioning during self-distillation, requiring no external teachers, privileged data, or extra inference costs.

This paper introduces Visual Contrastive Self-Distillation (VCSD), a training method that improves vision-language models by having them learn from themselves without needing an external teacher.

Andrea Napoli

ARROW enables variance reduction for domain adaptation on streaming data by adaptively reweighting minibatches, making it practical for online and distributed learning scenarios where offline algorithms fail.

This paper introduces ARROW, an online algorithm for reducing variance in domain adaptation tasks using streaming data. Unlike existing offline methods, ARROW works with data arriving continuously by maintaining moving averages of alignment statistics and reweighting incoming batches to match reference statistics, enabling efficient learning in real-time settings.

Andrea Napoli

Pairing samples strategically during training can significantly reduce variance in domain adaptation, leading to better optimization and improved performance on shifted datasets.

This paper tackles high variance in domain adaptation losses by proposing PSDA, a sampling technique that pairs observations within and across domains to reduce gradient variance during training. The method reformulates the pairing problem as linear assignment and shows improved accuracy on domain shift tasks.

Aditi Gupta, Yossi Gandelsman

Vision-language models have a circuit-level sensitivity to input order that can be fixed with simple test-time adaptation, improving both consistency and accuracy without retraining.

Vision-language models perform differently depending on whether you show them the image or question first—a semantically irrelevant change that shouldn't matter. This paper identifies this "modality order" failure and proposes a test-time training method that makes models consistent across both orderings while improving overall performance.

Roger Sala Sisó, Tiago Silvério, Jakob Sand et al.

Deploying VLA robots in the real world is primarily a systems integration problem—thoughtful data curation, control alignment, and targeted fine-tuning on a single GPU can turn a failing policy into a working one.

This paper presents DEED, a practical framework for deploying vision-language-action (VLA) robots in real retail environments.

Niqi Lyu, Pengtao Shi, Wei Qiu et al.

Small models can learn to strategically delegate hard problems to large models at the token level, cutting total inference costs dramatically while preserving reasoning quality—no separate router or LLM retraining needed.

PyroDash is a framework that pairs small language models with large language models for cost-efficient inference. During text generation, the small model learns when to ask the large model for help by emitting a control token, enabling a single handoff for completion. This approach reduces inference costs by up to 96% while maintaining competitive accuracy on math reasoning tasks.

Qijia He, Jiayi Cheng, Chenqian Le et al.

For coding agents, execution feedback from failures creates opportunities for cheap-model recovery that cascade-only systems miss—a learned router with budget calibration can match expensive-model performance while using 35% of the cost.

This paper addresses a key challenge in coding agents: deciding whether to retry a failed task with a cheaper model or escalate to a more expensive one. The authors train a router that learns when cheap recovery attempts are worthwhile based on execution feedback, then add a calibration layer that lets the same router work under different budget constraints without retraining.

Yuchen Jiao, Na Li, Changxiao Cai et al.

You can now use diffusion models to solve inverse problems with mathematical guarantees—the algorithm switches between following the learned prior and using measurement data based on noise levels, making it both practical and provably correct.

This paper proposes PDDIM, a theoretically-grounded algorithm for solving inverse problems (like image restoration) using diffusion models. The method modifies standard DDIM sampling to incorporate measurement information through coordinate-wise updates based on signal-to-noise ratios, and proves it converges to the correct Bayesian posterior while being computationally efficient.

Chirag Vashist, Ke Li

Single-step generative models can match multi-step diffusion models' quality without gradual transformations—simplicity in training objectives and architecture doesn't sacrifice performance.

This paper challenges the assumption that gradual noise-to-data transformations are necessary for good generative models. The authors build a minimalist single-step generative model using Implicit Maximum Likelihood Estimation (IMLE) and a simple convolutional network, achieving competitive results (FID 2.56 on ImageNet 256) without iterative denoising or transformers.

Atahan Dokme, Larry Heck

Dynamic, state-aware adapters that track information across tokens and retrieve relevant context segments significantly outperform static low-rank adaptation for reasoning tasks, without retraining the base model.

This paper proposes two new adapter methods that improve how language models adapt to reasoning tasks. MaLoRA adds dynamic, input-dependent scaling to low-rank adapters using recurrent state at the token level, while MaRA selects relevant context segments before answering. Together, these methods boost reasoning accuracy by up to 18% over standard LoRA across multiple models and benchmarks.

Chen Ziheng

You can now build neural networks on curved spaces using reusable, theoretically-grounded components instead of ad-hoc manifold-specific solutions, with practical efficiency improvements for important geometric representations.

This thesis develops a unified framework for building deep neural networks that work with data on curved mathematical spaces (manifolds) rather than flat Euclidean space.