ThinkLLM
ModelsCapabilitiesUse CasesBenchmarksPapersGlossary
ModelsCapabilitiesUse CasesBenchmarksPapersGlossary
AboutPrivacyTermsRSS

ThinkLLM

Spot an error in our data? Let us know.

Papers

Recent AI research papers with accessible summaries. Updated daily from arXiv, summarized for developers who don't read papers regularly.

2455 papers46 this month12 topics
AllEvaluation 43Agents 42Applications 39Efficiency 29Reasoning 29Training 27Safety 23Multimodal 18Architecture 9Alignment 8Data 4scaling 2

Nov 16 – Nov 22(8)

RL Post-Training in Diffusion Transformers Is Activation Steering

Nov 19, 2026

Tuna Han Salih Meral, Kaan Oktay, Alec Helbling et al.

RL post-training of diffusion models works largely through input-independent activation steering, meaning you can recover most improvements with a tiny fixed vector added to frozen model outputs—this has major implications for understanding and efficiently improving generative models.

This paper shows that RL improvements to diffusion transformers come primarily from simple activation steering—adding fixed vectors to model outputs—rather than learning complex input-dependent changes.

trainingefficiency

TideCache: Efficient Hybrid LLM Serving with Transparent Cache Management and Layer-Wise Streaming

Nov 19, 2026

Jinbin Luo, Xiaozhe Yao, Shang Yang et al.

By treating different cache types as flexible pools and overlapping data transfers with computation, TideCache achieves 2x more concurrent requests and 20-59% lower latency—showing that intelligent memory management, not just bigger GPUs, unlocks efficiency gains.

TideCache is a serving system that optimizes GPU memory usage for hybrid LLM architectures by intelligently managing different cache types across GPU and CPU memory. It uses transparent page redistribution, smart prefetching during computation gaps, and adaptive request scheduling to fit more requests and reduce latency on CPU-GPU systems.

Nov 9 – Nov 15(13)

StreamEmbed: Real-Time Indexing from Pixels for Screenshot Retrieval

Nov 13, 2026

Dayoon Ko, Yichuan Wang, Soyeong Jeong et al.

By reusing computations across incrementally-changing screenshots, you can index screen history 3.7x faster in real-time without sacrificing visual understanding—making local visual assistants practical.

StreamEmbed enables real-time visual indexing of screenshots by exploiting pixel redundancy across frames. Instead of recomputing embeddings from scratch, it uses spatial token pruning to skip unchanged image blocks and temporal key-value reuse to cache computations across screenshots, achieving 3.7x faster encoding than standard visual encoders while maintaining accuracy.

efficiencymultimodalapplications

MeReader: An Offline, Privacy-Preserving, Progress-Aware AI Assistant for Narrative eBook Reading

Nov 13, 2026

Mohammed Efaz, Udo Bub, Itilekha Podder

You can build privacy-preserving, spoiler-free reading companions using small local language models (2-4B parameters) that match larger models' performance while running entirely offline on consumer hardware.

MeReader is an offline AI assistant for reading ebooks that keeps your reading private and prevents spoilers by only accessing content up to your current position in the book. It uses a hybrid search system combining multiple retrieval methods and works efficiently on consumer devices, showing that smaller local models can match larger cloud-based systems in comprehension quality.

Nov 2 – Nov 8(25)

CallCleaner: Lightweight Bioacoustic Denoising from an Audio Foundation Model

Nov 8, 2026

Michael Xie, Gašper Beguš, Morris Alper

By distilling a general-purpose audio foundation model, researchers created a practical denoiser for animal recordings that's both faster and more effective than specialized bioacoustic methods, showing that foundation models can be efficiently adapted for domain-specific audio tasks.

CallCleaner is a lightweight audio denoising tool for animal recordings that distills a large foundation model into a 19M-parameter neural network. It removes background noise from field recordings of animal vocalizations without requiring voice detection, runs 90x faster than real-time on a laptop CPU, and outperforms existing bioacoustic denoisers while maintaining call quality.

efficiencyapplications

Saidzoda-CA-EmbeddingGemma-300M: One Small Embedder for Tajik, Kazakh, Uzbek and Kyrgyz

Nov 8, 2026

Tohir Saidzoda, Nekruz Maxudov, Bedil Karimov et al.

A single small embedding model can serve multiple low-resource languages effectively through targeted fine-tuning with synthetic data, but scaling up to larger multilingual models remains necessary for competitive performance on these languages.

Researchers fine-tuned a small 300M embedding model (EmbeddingGemma) to work across four Central Asian languages—Tajik, Kazakh, Uzbek, and Kyrgyz—using synthetic question-answer pairs generated by a language model. The fine-tuned model significantly improved retrieval performance (14-26 percentage points) on a multilingual benchmark, though it still lags behind larger 560M multilingual models.

Oct 26 – Nov 1(30)

Flide: Accelerating Decision Inference with Exact State Forking

Oct 30, 2026

Bobo Li, Jiajia Song, Hao Fei et al.

For LLM agents making categorical decisions, Flide eliminates redundant computation by executing multiple decision branches simultaneously in one pass, delivering both speed and numerical precision that existing inference engines can't match.

Flide is an inference engine that speeds up decision models used in LLM agents by efficiently handling multiple decision branches in a single forward pass. Instead of recomputing shared parts of the model multiple times, it uses a technique called ForkAttention to track different decision paths through GPU memory, achieving up to 9.5x speedup while maintaining mathematical accuracy.

efficiencyagentsreasoning

Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention

Oct 30, 2026

Ales Kubicek, Andrea Caciolai, Liangzu Peng et al.

LSA enables practical ultra-long context (1M+ tokens) by replacing expensive full attention with hierarchical cluster scoring, achieving 22.7x speedup in prefill and 4.6x in decode while preserving model accuracy.

This paper introduces Latent Sparse Attention (LSA), a method that makes attention efficient at ultra-long context lengths (1M+ tokens) by hierarchically clustering keys in a learned latent space. Instead of comparing queries to every key, queries score clusters of keys, reducing computation from quadratic to subquadratic while maintaining accuracy.

Oct 19 – Oct 25(20)

Where Rounding Happens Matters: Rounding-Location Alignment Stabilizes LLM Reinforcement Learning

Oct 23, 2026

Yijing Ke, Tianle Cai, Ruosong Wang

Training instability in LLM RL often stems from mismatched rounding locations between inference and training—fixing this in just a few critical computation paths can prevent collapse while keeping everything else unchanged.

This paper identifies how rounding differences between inference and training destabilize reinforcement learning in large language models, even when using identical weights.

trainingefficiency

Knowledge Distillation across Item and Code Spaces for Parallel Generative Recommendation

Oct 23, 2026

Yixuan Xu, Hailin Liu, Linxiang Peng et al.

Knowledge distillation works better for semantic-ID recommenders when you match distributions either by reconstructing items from codes or by projecting item probabilities to individual code positions—both approaches beat standard hard-label training.

This paper tackles a key challenge in generative recommendation systems: how to transfer knowledge from traditional recommenders to newer semantic-ID models that predict items as sequences of discrete codes. The authors propose two knowledge-distillation methods that bridge the gap between item-level and code-level predictions, achieving significant improvements in recommendation quality.

Oct 12 – Oct 18(4)

Rethinking the design space of Vanilla Pixel Diffusion Transformers

Oct 16, 2026

Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava et al.

Pixel diffusion models benefit from adaptive prediction strategies that vary by capacity and data type, and post-hoc high-frequency refinement can significantly improve image quality without expensive retraining.

This paper improves pixel-space diffusion transformers by introducing MFlow, which adaptively learns to blend two prediction strategies (clean image vs. velocity) based on data characteristics and model size. The authors also propose ConvRefiner, a lightweight module that fixes a persistent high-frequency quality deficit in generated images without retraining the main model.

architectureefficiencytraining

Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere

Oct 16, 2026

Joshua Williamson, Niall Jeffrey, Lorne Whiteway

Encoding physical symmetries directly into neural network architecture—rather than letting the model learn them—dramatically improves data efficiency and accuracy for scientific problems with inherent geometric structure.

This paper presents a spherical convolutional neural network designed for data naturally living on spheres (like cosmic microwave background maps). By building in rotational symmetry mathematically rather than learning it, the network uses training data more efficiently and achieves better results on cosmology tasks with fewer simulations needed.

efficiencyarchitecturetraining

Bio-PM-Guided Synthetic IMU Generation

Nov 19, 2026

Deepali Singh, Deepesh Parpani, Aviral Gupta

Synthetic data generated in a pretrained encoder's embedding space can effectively augment real IMU data for activity recognition, especially for hard-to-distinguish activities, without requiring raw signal synthesis.

This paper generates synthetic IMU (motion sensor) data in the representation space of a pretrained biosignal encoder to improve human activity recognition across different people. Rather than creating raw sensor signals, the authors train decoders (autoregressive and diffusion-based) to generate token sequences in the encoder's space, then use these synthetic tokens to augment training data.

dataefficiencyapplications

ProtoDiffusion: Token-per-Layer Generative Modeling of Multi-Layer Images

Nov 19, 2026

Adrià Ciurana Lanau, Harleen Hanspal, Lusine Abrahamyan et al.

By sharing a spatial prototype across layers and using one compact vector per layer, ProtoDiffusion achieves efficient multi-layer image generation that's 19× faster than decomposing flat images, while maintaining high reconstruction quality.

ProtoDiffusion generates multi-layer graphic designs (like Photoshop files) by representing each layer compactly as a single vector, rather than full spatial grids. This reduces memory from N×H×W tokens to H×W+N tokens. The model can generate designs from images or text, automatically determining layer count and layout without requiring them as input.

multimodalarchitectureefficiency

Cost-Aware Agentic Feature Optimization System for Fleet-Scale Ranking Models

Nov 19, 2026

Chang Yang, Yao Xuan, Wendan Yan et al.

LLM agents can effectively optimize complex multi-model systems by proposing hypotheses and using deterministic tools to verify them, achieving real production gains while respecting hard resource constraints.

A system that uses LLM agents to automatically optimize which features ranking models use, balancing performance improvements against computational costs. The system proposes feature combinations, ranks them by importance, and iteratively refines selections under budget constraints—improving production metrics by 0.20% while reducing pipeline costs.

agentsapplicationsefficiency

Cap-and-Track: Optimizing LLM Inference Request Dispatch Across Shifting Capacity

Nov 18, 2026

Benjamin Novotny, Doug Szeto, Chang Wang et al.

By treating server refusals as feedback signals rather than failures, you can build adaptive request dispatchers that automatically track available capacity without predicting output lengths or manually tuning rates.

Cap-and-Track is a system for efficiently routing inference requests to LLM serving clusters that dynamically adjusts to changing capacity. Instead of using fixed dispatch rates, it uses refusal signals from overloaded servers to probe the current capacity ceiling and adjust request routing in real-time, reducing GPU requirements by 43% while maintaining quality.

efficiencyagents

Local AI Agents in Power Electronics Simulation: Deployment, Measurements and Costs

Nov 18, 2026

Yuzhuo Li, Pasan Gunawardena, Yunwei Li

Local AI agent deployment for power electronics simulation trades off data privacy and control against measurable costs in power, energy, and time—organizations need quantitative benchmarks to decide whether local or cloud deployment makes sense for their specific tasks.

This paper evaluates locally-deployed AI agents for power electronics simulation tasks, measuring their power consumption, energy use, token generation, and execution time. Unlike cloud-based approaches, local deployment keeps sensitive designs and data on-site but creates computational overhead.

applicationsefficiencyevaluation

Can Programs Replace LLM Judges? Executable Program Judges for Scalable and Efficient Relevance Labeling

Nov 18, 2026

Eugene Agichtein, Charles L. A. Clarke, Kaustubh Dhole et al.

You can replace expensive LLM judges with one-time program generation that's faster, cheaper, and fully inspectable—achieving comparable accuracy to frontier models while costing just $0.17-0.49 per topic instead of per-document LLM calls.

This paper proposes ExecJudges, a method that uses LLMs to write readable programs instead of making individual relevance judgments. These programs use lexical matching and NLP tools to automatically label documents, eliminating expensive per-document LLM calls while remaining interpretable and reproducible.

evaluationefficiencyapplications
applicationsefficiencysafety

Learning When Diffusion Language Models Should Regenerate

Nov 12, 2026

Minhae Oh, Jeonghye Kim, Jinwoo Koo et al.

A learned verifier can efficiently improve diffusion language model outputs by deciding when to regenerate blocks, achieving better accuracy-to-cost tradeoffs than traditional search-based decoding.

This paper introduces DiVR, a verifier that learns when diffusion language models should regenerate text blocks during decoding. Rather than relying on token confidence, DiVR predicts whether a partially generated sequence will succeed, enabling selective remasking and regeneration.

reasoningefficiencytraining

CoReMem: Cost-Internalizing Memory for Successive Coding Tasks with Prospective Valuation

Nov 12, 2026

Peiwen Li, Harshit Verma, Sophia Zhou et al.

By measuring memory value prospectively on the next task and assigning credit only to steps that actually use memory, agents can build efficient, cost-aware memory systems that improve coding task performance without bloating context.

CoReMem is a framework that helps coding agents decide what information to remember between tasks. It measures how valuable different memory pieces will be for future unknown tasks, balancing the cost of storing and retrieving memory against its usefulness. The system improves a 27B model's ability to solve coding tasks from 21% to 25.7% while using fewer steps and tokens.

agentsefficiencyreasoning

Sketching Language Models for Context Attribution

Nov 12, 2026

Christopher Liaw, Aranyak Mehta, Grigoris Velegkas et al.

You can estimate token-level attribution in LLMs much faster than brute force by applying sketching to the backpropagation circuit, reducing evaluations from linear in prompt length to a small constant factor.

This paper develops an efficient algorithm to compute how removing individual tokens from a prompt affects a language model's output. Instead of evaluating the model n times (once for each token removal), the algorithm uses sketching techniques to estimate all attribution values with only O(√log(1/ε)/ε) evaluations, matching a proven lower bound.

evaluationefficiency

Classical Energies Learn to Reason 1,000x Faster than Looped Models

Nov 12, 2026

Benjamin Hoover, Hugh Blayney, James C. R. Whittington

Energy-based models offer a faster, more efficient alternative to looped reasoning architectures for constraint satisfaction problems, suggesting that iterative reasoning may be overkill for many tasks.

This paper shows that energy-based models (specifically a learned version of Hopfield networks) solve reasoning puzzles like Sudoku and Mazes 1,000x faster than looped reasoning models, while using far fewer computations. The approach learns to encode puzzle constraints directly into an energy function, enabling efficient problem-solving without iterative loops.

reasoningefficiencyarchitecture

Metal Sandbox: Isolated Execution and Command Scheduling for GPU Coding Agents on Apple Silicon

Nov 12, 2026

Haozhe Wang, Lina Chen, Fanqi Wan et al.

When building AI agents that generate GPU code, isolating execution environments and intelligently scheduling commands across multiple devices can nearly double throughput and significantly reduce idle time.

Metal Sandbox is infrastructure for efficiently training and evaluating AI agents that generate GPU code for Apple Silicon. It isolates agent workspaces, schedules GPU commands across multiple Macs, and lets waiting commands migrate to available machines—achieving 1.68–2.05× throughput gains and reducing tool waiting time by 18.8%.

efficiencyagentsapplications

Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery

Nov 12, 2026

Samidha Verma, Ankit Anand, Sayan Ranu

You don't need the most expensive LLM for every task—smart routing between models using bandit algorithms can produce better solutions at lower cost by adapting which model to use based on real-time performance.

This paper presents Ballad, a bandit algorithm that dynamically selects which LLM to use at each step of automated heuristic discovery for combinatorial optimization. Rather than assuming larger models are always better, Ballad learns to orchestrate multiple LLMs into an ensemble that outperforms any single model while cutting costs by up to 66%.

agentsefficiencyreasoning

Maximizing Capacity Utilization of LoRA Adapters for Mitigating the Learning-Forgetting Trade-off

Nov 12, 2026

Tomohiro Sawada, Kartik Goyal

LoRA adapters waste capacity by not fully utilizing their configured rank; adding an effective-rank regularizer during training fixes this and improves the learning-forgetting trade-off without extra computational cost.

This paper identifies why LoRA adapters underperform full fine-tuning: they don't use their full capacity. The authors show that LoRA updates have much lower effective rank than their configured rank, limiting how broadly parameter updates spread.

efficiencytrainingalignment

Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination

Nov 12, 2026

Wei Feng, Meiling Chen, Boxiao Chen et al.

Coordinating GPU resource allocation, cache-aware scheduling, and environment execution can significantly speed up batch processing of agent workloads without sacrificing accuracy.

This paper optimizes batch execution of LLM agent tasks by coordinating inference and environment interaction. It introduces resource allocation strategies for hybrid LLMs, trajectory-aware routing for cache reuse, and Chain-Advancing Action Scheduling to handle environment bottlenecks—achieving 10-65% makespan reductions on real benchmarks.

efficiencyagentstraining

Folio: Recoverable Multi-Resolution Context Folding for Long-Horizon Language Agents

Nov 12, 2026

Yubo Ruan, Tao Peng, Jiao Zhang et al.

By recording context at multiple levels and folding programmatically rather than summarizing with the model, agents can handle longer histories more efficiently—solving 66/75 tasks vs 54 with model summarization, in half the time and 2.4x lower cost.

Folio is a context manager for long-horizon language agents that replaces model-based summarization with a multi-level recording system.

agentsefficiencyreasoning

ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows

Nov 10, 2026

Kaiyuan Hou, Yang Ping Kuo, Indranil Bhattacharya et al.

You can replace expensive LLM calls in repetitive business processes with rule-based code by learning decision logic from agent execution traces—achieving near-agent performance at a fraction of the latency and cost.

ReTrace converts LLM agent workflows into fast, cost-efficient code-based processes by extracting decision rules from agent reasoning logs. It analyzes why agents took actions, distills these explanations into compact rule documents, and generates executable workflows that run 130x faster at 24% lower cost while maintaining 93.8% success on production tasks.

agentsefficiencyapplications

C$^2$-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference

Nov 10, 2026

Guorui Lu, Qinyu Chen

Cropping and compression are complementary strategies for reducing VLM inference cost—use cropping for detail-focused questions and compression for context-heavy ones, with a learned safety mechanism to protect accuracy.

This paper proposes C²-Routing, a method that adaptively chooses between cropping (for local details) and compression (for context) to reduce computational costs in vision-language models for VQA tasks. It includes a safety gate to prevent accuracy loss and achieves 36.53% FLOP reduction with minimal performance drop.

efficiencymultimodalapplications
multimodaltrainingefficiency

Semantically Aligned KV Cache Compression for Efficient Agentic Autoregressive Video Generation

Nov 8, 2026

Weizhong Huang, Huaijia Kou, Tianyu Xie et al.

By aligning KV cache compression with semantic planning constraints, you can reduce memory usage by 3.35x and speed up video generation by 1.22x without retraining, making agentic video generation more practical for resource-constrained settings.

This paper presents SemanticKV, a method to compress the key-value cache in video generation models by using semantic constraints from a planner to decide which cached information to keep.

efficiencymultimodalagents

How Much Training-Inference Mismatch Can On-Policy Distillation Tolerate?

Nov 8, 2026

Cong Yu, Ziting Wang, Songlin Jiang et al.

On-policy distillation tolerates significant training-inference mismatch from quantization and asynchronous rollouts without quality loss, enabling 1.6× faster training with FP8 precision while maintaining performance.

This paper studies how much training-inference mismatch on-policy distillation (OPD) can tolerate when using quantized inference and asynchronous rollouts to reduce training costs.

trainingefficiency

Flex-DDCM: Zero-Shot Task-Oriented Image Compression via Diffusion Trajectory Guidance

Nov 8, 2026

Yixin Gao, Runsen Feng, Xin Li et al.

You can compress images for specific AI tasks (like object detection) without retraining by steering a pretrained diffusion model's denoising process with task guidance signals—enabling one codec to serve multiple purposes.

This paper presents Flex-DDCM, a zero-shot image compression method that uses diffusion models to compress images while optimizing for specific downstream tasks. By guiding the diffusion denoising process with task-specific signals, it achieves flexible trade-offs between image quality and task performance without requiring task-specific training.

efficiencymultimodalapplications

CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse

Nov 6, 2026

Peter Belcak, Greg Heinrich, Saurav Muralidharan et al.

You can extract and cache the reasoning logic from LLM problem-solving into reusable CPU programs, eliminating redundant computation and dramatically reducing inference costs while maintaining performance.

This paper introduces agentic reasoning caching (ARC), a method that extracts reasoning logic from LLM traces and converts it into efficient, CPU-executable programs. Instead of re-solving problems from scratch, ARCs apply learned solution patterns to new instances, achieving comparable performance to frontier LLMs while being 16-291x cheaper and fully explainable.

efficiencyreasoningagents

On the Reliability of Majority-Vote Pseudo-Labels in Test-Time Reinforcement Learning

Nov 6, 2026

Haoyu Wang, Ru Wang, Yuanyuan Wang et al.

When adapting language models at test time using self-generated labels, you should weight updates based on how reliable the voting signal is for each prompt—this prevents the model from learning from noisy pseudo-labels.

This paper addresses a key problem in test-time reinforcement learning: how reliable are the self-generated labels used to adapt language models? The authors analyze majority-vote pseudo-labels, identify sources of error, and propose PC-Dr.DPO, a method that makes the model more cautious when updating from unreliable prompts. They validate their approach across math and reasoning tasks.

trainingalignmentefficiency

CodeQuant2: Enhanced Outlier Mitigation and Adaptive Clustering for Low-Precision LRMs

Nov 6, 2026

Xingyu Liu, Xiangyang Yin, Tianhua Xia et al.

Reasoning models need special quantization handling because they show larger activation fluctuations early in generation—CodeQuant2 captures this with adaptive clustering and better calibration, achieving 2.85× speedup with minimal accuracy loss.

CodeQuant2 is a quantization technique that compresses large reasoning models to low precision while maintaining accuracy. It addresses unique challenges in reasoning models by detecting and handling unusual activation patterns in early decoding steps, using clustering to represent weights efficiently, and adapting compression levels per layer.

efficiencytraining

BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

Nov 6, 2026

Dayou Du, Woon Yee Ng, Guangming Sheng et al.

You can compress LLM key-value caches to 4-bit precision without significant accuracy loss by carefully handling quantization errors at the algorithm level and co-designing with hardware capabilities—delivering major serving speedups in practice.

BitDecoding-FP4 enables efficient long-context LLM serving by storing key-value caches in ultra-low-precision 4-bit floating-point format while maintaining accuracy.

efficiencytrainingarchitecture

User Embeddings are Superpositions of Interpretable Behavioral Modes

Nov 5, 2026

Jonathan Kozaczuk, Zheng Dong, Noyan Evirgen et al.

User embeddings contain interpretable behavioral patterns that can be extracted post-hoc using sparse autoencoders and automatically labeled—making opaque recommendation systems more transparent without changing existing infrastructure.

User embeddings in recommendation systems are dense and hard to interpret. This paper shows that embeddings can be decomposed into sparse, interpretable behavioral modes using sparse autoencoders. The recovered modes can be automatically labeled in natural language, enabling zero-shot audience retrieval without retraining the original system.

applicationsefficiency

SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

Nov 5, 2026

Qile Jiang, George Em Karniadakis

Agent swarms communicating in latent space without text generation can solve complex optimization problems 7-20x faster than traditional LLM-guided search, even with smaller models.

SwarmEvolve uses hundreds of lightweight AI agents working together in swarms to improve programs for math and science problems. Instead of generating long text explanations, agents reason directly in latent space and share information through KV caches, with a novelty filter ensuring diverse ideas. The system is 7-20x faster than existing methods while using smaller models.

agentsreasoningefficiency

Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing

Nov 5, 2026

Michal P. Podolinsky, Yunhao Yang, Daniel Little et al.

By probing a VLA's latent representations during task execution, you can detect failures 30% earlier than existing methods and maintain reliable failure predictions even when the robot encounters new tasks it wasn't trained on.

VLA-Tug is a lightweight monitoring system that detects when vision-language-action robots are about to fail by analyzing their internal representations. It predicts both whether failure will occur and what type of failure it will be, enabling corrective actions while running efficiently on edge devices.

safetyevaluationefficiency

Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement Learning

Nov 5, 2026

Chenruo Liu, Zihan Wang, Yijun Dong et al.

LoRA's optimization dynamics differ fundamentally between SFT and RL due to baseline subtraction in RL; adapting LoRA's spectrum management to each stage—spreading updates in SFT, rotating subspaces in RL—improves fine-tuning efficiency.

This paper analyzes how LoRA (a parameter-efficient fine-tuning method) behaves differently during supervised fine-tuning versus reinforcement learning. By tracking the mathematical properties of LoRA updates, the authors discover that RL concentrates weight changes along fewer directions while SFT spreads them out, and they trace this difference to how baseline rewards are subtracted in RL.

trainingefficiency

StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

Nov 5, 2026

Yang Ouyang, Jung-Eun Kim

Instead of resampling complete reasoning trajectories, you can pinpoint error locations at the step level and only regenerate from there—matching accuracy with significantly lower computational cost.

StepBack improves the efficiency of chain-of-thought reasoning by identifying and resampling only the steps where errors likely occur, rather than regenerating entire reasoning chains. It achieves the same accuracy as standard self-consistency while using up to 41% fewer tokens.

reasoningefficiencyevaluation

FedReason: Large Language Models as Weight-Space Meta-Aggregators for Heterogeneous Federated Learning

Nov 5, 2026

Ali Mohammadi Ruzbahani, Abbas Yazdinejad, Hadis Karimipour

LLMs can act as adaptive meta-aggregators in federated learning without server-side training, dynamically selecting aggregation strategies based on weight analysis—though performance depends heavily on proxy feedback and the approach doesn't yet match specialized methods like SCAFFOLD.

FedReason uses a large language model to dynamically choose aggregation rules for federated learning by analyzing client model weights and testing candidates on a small proxy dataset. The LLM runs code in a sandbox to inspect weights, selects or writes Python aggregation programs, and the server applies them until the next LLM call, falling back to FedAvg if invalid.

trainingefficiencysafety

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

Nov 5, 2026

Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer

Linear attention models can be made significantly more efficient by allocating parameters based on sensitivity analysis: use low-rank gates for state erasure (which matters less) and invest savings into wider networks and better readout mechanisms.

Muir improves linear attention models by analyzing how state updates work differently across channels. The key insight is that erasing old information requires less capacity than writing new information, so the model uses low-rank gates for erasure and reinvests those parameters into a wider network.

efficiencyarchitecturetraining

StrataFlux: Memory-Hierarchy-Native Low-Bit LLM Inference

Nov 5, 2026

Jake Hyun, Ziyu Zhao, Jae-sun Seo et al.

You can quantize LLM weights and KV caches to 2-3 bits without accuracy loss by selectively streaming precision from slower storage during inference—the key is overlapping transfers with computation so decoding speed doesn't suffer.

StrataFlux recovers accuracy lost to extreme quantization (2-4 bits) in LLM inference by strategically fetching higher-precision data from RAM and SSDs during generation. It overlaps these transfers with computation so they don't slow down decoding, enabling models like Llama-3.1-8B to run on GPUs with half the memory while matching the accuracy of less-aggressive quantization schemes.

efficiencytraining

SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains

Nov 5, 2026

Siva Rajesh Kasa, Yasong Dai, Sumit Negi et al.

You can generate text 1.7-4.4x faster from diffusion models by proposing tokens speculatively and verifying them in parallel, while mathematically guaranteeing the output distribution stays identical to sequential generation.

SpecCTMC speeds up text generation from masked diffusion language models by speculatively proposing multiple token updates in parallel, then verifying them exactly without changing the output distribution. Unlike typical speedup methods, it requires no separate draft model and mathematically guarantees it preserves the original sampling behavior.

efficiency

SpecCompact: Accelerating LLM Agent Context Compaction with Slot-Parallel Semantic Speculative Verification

Nov 4, 2026

Yulu An, Yifei Feng, Xiangyang Yin

Parallel drafting with verification can make context compression 2-5x faster for AI agents, enabling them to handle longer interactions without sacrificing speed or information quality.

SpecCompact speeds up how AI agents compress long conversation histories by using a small model to draft multiple summary sections in parallel, which a larger model then verifies and fixes. This reduces the time agents spend summarizing while keeping important information intact, achieving 2-5x faster compression on real agent tasks.

efficiencyagentsreasoning

Balancing hierarchical representation and quantization: A performance trade-off in text embeddings

Nov 4, 2026

Kai Ponel, Marco Polignano, Erasmo Purificato et al.

Int8 quantization is a practical efficiency win for text embeddings with minimal quality loss, but binary quantization and aggressive dimensionality reduction hurt retrieval performance significantly—the trade-off varies by task type.

This study benchmarks how hierarchical text embeddings and quantization—two efficiency techniques—interact across 38 retrieval tasks. Testing 168 configurations per task, the authors find int8 quantization performs nearly as well as full precision (losing only 1.5% effectiveness), while binary quantization significantly hurts performance.

efficiencyevaluation

CompressionBench: How can your coding agents compress LLMs?

Nov 4, 2026

Xinrui Zhong, Zhifeng Li, Zhiming Wang et al.

Coding agents can autonomously compress LLMs with reasonable success, but their performance depends heavily on how they design experiments, interpret feedback, and validate results—not just on compression algorithm choice.

This paper studies whether AI coding agents can autonomously compress large language models to fit storage constraints while maintaining performance. Agents receive a source model, storage limit, and time budget, then design compression algorithms and write inference code.

efficiencyagentsevaluation

DreamX: Video Diffusion Model is a Universal Latent Action Learner

Nov 3, 2026

Jisoo Kim, Chaehyun Kim, Junsu Kim et al.

Video diffusion models can be repurposed as action encoders to create embodiment-agnostic action representations that transfer across robots and humans without retraining, while being 300× more compact than raw video.

DreamX repurposes a pretrained video diffusion model to learn universal action representations that work across different robots, humans, and environments.

multimodalagentsefficiency

Adversarial Speculative Reasoning Improves Long Horizon Agents

Nov 3, 2026

Hongyin Luo, Wei Fang, Lijie Yang et al.

Speculative decoding isn't just a speed trick—it changes how models reason over long horizons. By training draft models to fail at critical moments, you can force the main model to handle important reasoning steps with better precision, improving agent accuracy on complex tasks.

This paper shows that speculative decoding—a technique usually just for speed—actually affects reasoning quality in long-horizon AI agents. The authors discovered that numerical noise accumulates during the verification process, degrading reasoning accuracy over many steps.

reasoningefficiencyagents

CATF: A Manager-Worker Framework for Context-Aware Time-Series Forecasting

Nov 3, 2026

Yueqiao Chen, Dongyu Liu

Hard routing through a Manager model lets specialized Workers handle different data regimes automatically, improving forecasting accuracy while reducing inference costs compared to soft-routing ensemble approaches.

CATF is a forecasting framework that uses a Manager-Worker architecture to predict time-series values when multiple contextual factors (like weather or calendar) influence the data. A Transformer-based Manager intelligently routes each input to a specialized Worker model based on context, allowing workers to automatically specialize without manual data splitting.

architectureefficiencytraining

PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data

Nov 3, 2026

Kewei Zhang, Juncheng Ma, Guodong Lv et al.

LLM agents can autonomously generate high-quality robot training data by building simulations and rendering photorealistic videos, achieving better generalization than human-collected data at a fraction of the cost.

A system where an LLM agent acts as a virtual teleoperator to generate robot training data. The agent receives a real robot scene, writes simulation code and controllers, then uses a video model to render photorealistic training episodes. The resulting synthetic data trains policies that match or exceed those trained on real teleoperated data, while costing 60-70% less than human teleoperation.

agentsapplicationsefficiency
efficiencyarchitecturescaling

A Better Critic Matters for Distribution Matching Distillation

Oct 30, 2026

Zongze Wu, Xianbang Wang, Zhengyang Geng et al.

For distillation to work well, the critic that evaluates student quality must be scaled proportionally with the teacher and should condition on generation steps—otherwise training can collapse or students fail to improve.

This paper improves Distribution Matching Distillation (DMD), a method for training fast student models from slower teacher models. The authors show that the critic network—which guides student training—needs to be scaled alongside the teacher and should be aware of generation steps to provide reliable training signals.

trainingefficiency

DecisionRank: On-Policy Distillation for Decision-Model Rerankers

Oct 30, 2026

Yanran Wu, Dian Yu, Chongyang Bai et al.

Decision models can match or exceed generative LLM ranking quality through on-policy distillation, while being orders of magnitude faster—enabling practical deployment for retrieval tasks.

DecisionRank trains a decision model to rank documents by predicting scores directly, avoiding slow autoregressive decoding. It uses a novel on-policy distillation method where a generative teacher LLM provides feedback on the student's predicted rankings, enabling a 1.7B model to outperform its 27B teacher while running 24× faster.

efficiencytrainingevaluation

ACCT: Asymmetric Cross-LLM Context Transfer

Oct 30, 2026

Yuzhe Liu, Haolin Tian, Tonghan Wang

You can transfer pre-computed context between different-sized models in the same family, cutting context processing time by 6-21x—useful when deploying multiple model sizes and reusing the same input across them.

ACCT enables reusing computed context (KV cache) across different LLM sizes within the same family, avoiding redundant computation when switching models. It transforms a source model's KV cache into sparse, compressed states that a smaller target model can directly use, achieving 6.5-21x speedup on context processing while maintaining performance.

efficiencyarchitecture

Subquadratic Attention with a Hierarchical Tree-Structured Indexer

Oct 30, 2026

Leon Lufkin, Cutter Dawes, Alexander Speicher et al.

ZHA achieves subquadratic attention scaling by using a hierarchical tree indexer, potentially enabling LLMs to handle contexts of hundreds of millions to billions of tokens—a major step toward solving the long-context bottleneck without sacrificing expressiveness.

This paper introduces Zyphra Hierarchical Attention (ZHA), a method that reduces the computational cost of attention in large language models from quadratic to subquadratic scaling.

efficiencyarchitecturescaling

RangeSWA: Training-Free Length Extrapolation for Continuous Diffusion Language Models

Oct 30, 2026

Guangfu Guo, Xiaoqian Lu, Linsey Pang et al.

You can extend diffusion models to generate much longer text without retraining by mixing local attention with long-range context retrieval and adaptive denoising strategies.

RangeSWA enables diffusion language models to generate longer text without retraining by combining sliding-window attention with remote summaries and dynamic retrieval. It fixes common problems like repetition and quality loss when generating beyond training length through positional remapping and adaptive scheduling.

efficiency

RAMA: Retrieval-Augmented Multi-Agent Routing with Episodic Memory for LEO Satellite Networks

Oct 30, 2026

Xin Ming Li, Peng Hu

Adding episodic memory retrieval to multi-agent reinforcement learning dramatically improves routing decisions in dynamic satellite networks, enabling agents to reuse successful past strategies without retraining.

RAMA is a routing system for satellite networks that combines reinforcement learning with memory retrieval. Instead of just learning from current observations, each satellite remembers successful routing decisions and retrieves similar past experiences when making new routing choices. This approach reduces delivery delays by 13-93% compared to existing methods, especially as networks scale up.

agentsreasoningefficiency

Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

Oct 29, 2026

Jiongfan Yang, Weichen Fan, Haiwen Diao et al.

By splitting image frequency components across local and global attention paths, you can generate finer image details without the quadratic cost of full global attention—achieving better quality at lower computational cost.

SpecSlot is a pixel-space image generation model that separates low-frequency and high-frequency image details into different processing paths. Low-frequency information uses global attention for context, while high-frequency details are processed locally, reducing computational cost while maintaining image quality.

architectureefficiencymultimodal

An ODE-based Sampler for Discrete Diffusion Models

Oct 29, 2026

Samson Gourevitch, Yazid Janati, Eric Moulines et al.

Discrete diffusion models can now use deterministic ODE-based sampling (like continuous diffusion models do), enabling faster inference and better distillation without sacrificing generation quality.

This paper develops a deterministic sampling method for discrete diffusion models by introducing exponential residual flows (ERFs), which use ODEs to guide token generation.

efficiencytrainingarchitecture

TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races

Oct 29, 2026

Tom Sander, Pierre Fernandez, Itai Gat et al.

TextSeal2 achieves distortion-free LLM watermarking that survives speculative decoding by using copula-based key merging and memoryless exponential races, detecting watermarks 8-22% more reliably than prior methods at the same false-positive rate.

TextSeal2 is a watermarking technique for large language models that embeds invisible signatures in generated text without changing output quality or speed.

safetyefficiency

When to Assist and What to Suggest: Joint Gating and Memory-Grounded Hint Generation in One Compact Decoder

Oct 29, 2026

A. J. A. Jubayer Talukder, Mohammad Ariful Haque

A single small model can effectively decide when to assist and generate helpful hints by leveraging memory grounding, outperforming larger multi-model systems on speed while maintaining quality—memory context matters far more than retrieval or model size.

This paper tackles proactive assistance in conversations by combining gating (deciding when to help) and hint generation (what to suggest) into a single compact model. The authors show that a 1B parameter decoder can match larger two-model pipelines while being 1.31-1.76x faster, with memory-grounded hints being the key driver of quality.

efficiencyapplicationsreasoning

Multi-View MoE: Letting Every Router Shape Expert Computation

Oct 29, 2026

Zelin Zhao, Tianchi Ma, Junyuan Shang et al.

Using multiple routers with different expert preferences improves expert selection in MoE models, and reusing routing signals to customize expert inputs boosts performance without increasing compute cost.

This paper improves sparse mixture-of-experts models by using multiple routers to select experts and modulate their computation. Instead of one router deciding which experts to use, multiple independently-trained routers vote on expert selection, and their preferences are reused to customize each expert's input representation.

architectureefficiencytraining

Every Token Needs Its Own Teacher: Multi-teacher Adaptive On-policy Self-distillation

Oct 29, 2026

Yifu Luo, Chuang Ma, Zeyu Chen et al.

Token-level teacher selection during training outperforms fixed or response-level teacher assignment—different parts of a model's output benefit from different supervision signals.

This paper proposes MTA-OPSD, a method that assigns different teacher models to individual tokens during language model training. Unlike existing approaches that use a single fixed teacher or vary teachers only between responses, this method lets each token choose its best teacher through voting and KL divergence selection.

trainingefficiency

Training Foundation Models with a Single Algorithm

Oct 29, 2026

Yuda Song, Alvin Zhang, Fahim Tajwar et al.

Foundation models don't need separate training algorithms for different stages—you can train everything with one likelihood-based method by treating RL objectives as fitting to model-generated distributions, which improves efficiency and capability coverage.

This paper shows that foundation model training stages (pretraining, fine-tuning, reinforcement learning) can be unified under a single likelihood-based algorithm. Instead of switching between different training methods, the researchers reformulate RL objectives as maximum-likelihood fitting, allowing all stages to use the same core algorithm.

trainingefficiencyreasoning

Retrieval as Traversal: Value-Indexed Memory for Stateful Long-Context Agents

Oct 29, 2026

Daniel Strauß, Zhenshan Bing, Alois Knoll

Instead of searching all memories every forward pass, stateful retrieval navigates locally through memory neighborhoods, potentially making long-context reasoning more efficient by exploiting overlap between consecutive retrieval problems.

This paper proposes a new approach to memory retrieval for long-context AI agents that treats retrieval as persistent navigation through an address space, rather than searching from scratch each time. By separating memory content from its location and charging for access cost, the method encourages related memories to cluster together, reducing redundant retrieval work across similar tasks.

reasoningefficiencyagents

On the Fundamental Limits of Prefill-Based Decoding Length Prediction

Oct 29, 2026

Haowen Shi, Xuzhe Yan, Fahao Chen et al.

Prefill-based decode-length prediction has hard limits set by autoregressive stochasticity, not predictor design—changing the model, data, or sampling breaks existing benchmarks because they measure task difficulty, not predictor quality.

This paper reveals why token-length prediction for LLMs fails when conditions change, showing the problem isn't bad predictors but fundamental randomness in autoregressive decoding. The authors derive mathematical bounds on prediction accuracy and show that decode length follows a normal distribution determined by the model's stopping probabilities, not the predictor architecture.

efficiencyevaluationreasoning

Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese

Oct 29, 2026

Pedro Carvalho Brom

Small language models trained on consumer hardware can match or exceed much larger models on specialized domains, but memorization and domain mismatch remain key challenges—this work provides practical benchmarks and a reusable toolkit for building domain-specific SLMs.

A tutorial on building a small language model (11-12M parameters) from scratch on a single consumer GPU for Brazilian Portuguese legal and general text.

trainingefficiencyevaluation

NEOPIC: A Neural Operator Framework for Particle-based Kinetic Plasma Simulations

Oct 29, 2026

Sriramkrishnan Muralikrishnan, Chelsea Maria John, Jeffrey Kelling et al.

Neural operators can efficiently replace expensive field solvers in particle simulations by learning directly from particle data, with proven error bounds and linear computational complexity independent of mesh resolution.

NEOPIC is a neural operator framework that learns to replace the field solver in particle-based plasma simulations. Unlike grid-based approaches, it operates directly on particle positions without requiring a mesh, avoiding noise aliasing issues. The framework is plug-and-play, works with different simulation codes, and comes with theoretical guarantees on accuracy and computational complexity.

architectureefficiencyreasoning

Preserving Utility in Continual LLM Unlearning via Orthogonal Projections

Oct 29, 2026

Juan Belieni, Ana Carolina Erthal, Eliezer de Souza da Silva et al.

COP enables multiple sequential unlearning requests by mathematically preventing new updates from undoing previous deletions, storing only low-rank information per layer and working with existing unlearning methods.

This paper addresses the problem of repeatedly removing sensitive data from language models without retraining. When unlearning requests happen sequentially, standard methods degrade in performance.

safetytrainingefficiency

KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry

Oct 27, 2026

Wenqing Peng, Michael Boy, Zhi-Song Liu

By making atmospheric chemistry models differentiable, KPPax enables automatic computation of parameter sensitivities and hybrid physics-ML models without hand-coded adjoints—reducing emission estimation errors by 2.4–5× in demonstrated applications.

KPPax converts atmospheric chemistry models from KPP format into differentiable JAX code, enabling automatic differentiation and machine learning integration.

trainingefficiencyapplications

MuonGS: Pretraining LLMs with Relaxed Weight-Norm Control

Oct 27, 2026

Athanasios Glentis, Dawei Li, Mingyi Hong

Relaxing weight-norm constraints to allow controlled growth (proportional to √t) improves optimizer efficiency and model convergence speed compared to strict norm-fixing approaches, with practical gains of ~15% on large-scale pretraining.

This paper analyzes how weight-norm control affects optimizer performance during language model pretraining and proposes MuonGS, a new optimizer variant. The key insight is that controlling how fast weight norms grow—rather than fixing them—allows better exploration during training.

trainingefficiency

Multi-State Linear Attention with Generalized Second-Order Routing

Oct 27, 2026

Hee Seung Hwang, Sanghyuk Chun, Xindi Wu et al.

By incorporating variance information into how segments are aggregated in linear attention models, MERLIN enables better long-context understanding without the quadratic complexity of standard attention—useful for scaling LLMs to longer sequences.

This paper improves Multi-State Linear Attention (MSLA) for long-context language models by tracking both the mean and variance of key alignments across segments, rather than just the mean.

efficiencyarchitecturereasoning

Metacognitive Decoding: Efficient Inference for Diffusion Language Agents

Oct 27, 2026

Mingyu Zhao, Shuhang Lin, Ligong Han et al.

Diffusion models can generate multiple tokens in parallel, but you need a smart controller to decide when to stop and how much to generate; this work shows you can cut inference costs by ~20% by training a lightweight decision module on agent trajectories.

This paper introduces a metacognitive controller that optimizes how diffusion language models generate text during agent interactions. By learning when to stop generating tokens and how long responses should be, the controller reduces computational cost by 20.6% while maintaining task performance—making parallel token generation more practical for real-world agents.

efficiencyagentsreasoning

EnvSched: Environment-Aware Scheduling for Rollout-Efficient Agent Reinforcement Learning

Oct 27, 2026

Sihan Yang, Weichen Bai, Ruichuan An et al.

By treating environments as first-class scheduling objects and predicting reward variance across task-environment pairs, EnvSched achieves 3x faster training than random selection—reaching in 40 steps what random sampling needs 120 steps to accomplish.

EnvSched is a scheduling system that makes reinforcement learning for language agents more efficient by intelligently selecting which tasks to train on within limited rollout budgets.

trainingefficiencyagents

MoI-Bench: A Benchmark for Image Generation and Editing on Mobile Devices

Oct 27, 2026

Kailai Feng, Haoxiang Rao, Yuxiang Wei et al.

Mobile image models need different evaluation criteria than server models—this benchmark shows that expanding task coverage matters more than further optimizing deployment efficiency, and specific capability gaps (like text replacement) should be prioritized over broad improvements.

MoI-Bench is a benchmark for evaluating image generation and editing models on mobile devices, with 1150 samples across 64 subtasks defined by combinations of 16 operations (like removal, replacement) and 8 visual targets (portraits, text, etc.).

evaluationefficiencyapplications

Retriever-Free Retrieval: Navigating Documents with Decision Models

Oct 26, 2026

Chao Wan, Paul Jünger, Beixi Du et al.

Language models can retrieve documents more effectively by making comparative judgments across candidates in shared context, rather than relying on fixed embeddings or lexical matching—eliminating the retriever bottleneck that even strong generators cannot overcome.

DecisionRAG replaces traditional retrieval with a decision model that navigates document hierarchies like browsing a library, evaluating multiple candidate passages simultaneously without generating tokens. It outperforms dense retrieval systems by 13-17 points on standard benchmarks while remaining fast (0.7-1.8 seconds per query on millions of documents).

reasoningefficiency

Preserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR

Oct 26, 2026

Yuchen Zhang, Haralambos Mouratidis, Ravi Shekhar

When adapting multilingual speech systems to new languages, simple fine-tuning causes catastrophic forgetting of old languages—but selective merging strategies like SSU-Wanda+TIES can preserve both old and new language performance simultaneously.

This paper addresses expanding multilingual speech recognition systems to new languages while preserving performance on existing languages, without replaying original training data. The researchers test different adaptation and merging strategies for a connector module in speech-based LLMs, finding that careful update merging balances learning new languages with retaining old ones.

trainingmultimodalefficiency

Reva: Real-Time Vulnerability Detection through Contextual Probing of Large Language Models

Oct 26, 2026

Amine Lbath, Massih-Reza Amini

LLM internal representations already contain rich security insights—you don't need expensive multi-step reasoning to detect vulnerabilities. A trained probe reading hidden states can match complex agentic detectors while being 250× faster.

REVA is a vulnerability detector that achieves real-time security screening by extracting threat signals from an LLM's internal representations, rather than running expensive multi-step reasoning.

safetyefficiencyevaluation

Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models

Oct 26, 2026

Aidar Myrzakhan, Zhiqiang Shen

Diffusion LLMs can generate much faster by strategically choosing which tokens to commit to first—prioritizing tokens that unlock predictions for other positions, not just the easiest ones to predict.

This paper introduces dSpark, a method for speeding up diffusion language models by learning which tokens to reveal first during generation. Instead of just predicting easy tokens, the approach identifies tokens whose revelation helps other positions become easier to predict, reducing the total decoding rounds needed.

efficiencytrainingreasoning
trainingefficiency

If at First You Don't Succeed, Backtrack: Lightweight Recovery for Frozen VLAs

Oct 23, 2026

Thomas Landeg, Thierry Tambe

You can add robust failure recovery to frozen robotic VLAs with minimal overhead by training a small failure detector on the model's existing representations and using simple motion replay—no retraining or new data needed.

Backtrack is a lightweight failure recovery system for frozen Vision-Language-Action (VLA) robotic models that detects when a task will fail and automatically retraces the robot's recent motion to retry from a safer state.

efficiencyagentsapplications

Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model

Oct 23, 2026

Yalda Taheri, Mohammad Hassan Heydari, Armon Rasooli

Language-specific retrieval models can match much larger models through targeted fine-tuning with distillation, and open-sourcing training data and procedures enables reproducible progress in underserved languages like Persian.

Yalda Embedding is a 596M-parameter Persian retrieval model fine-tuned from Qwen3-Embedding using over 1 million query-document pairs. It outperforms comparable 0.6B models on Persian retrieval tasks and approaches 4B model performance, with gains from knowledge distillation and contrastive learning. The authors release all code, weights, and training data to support Persian NLP research.

trainingmultimodalefficiency

Adaptive Frequency Threshold Byte Pair Encoding for Mixed Natural Language and Structured Symbolic Corpora

Oct 23, 2026

Pietro Di Nicco, Lucio La Cava, Andrea Tagarelli

Tokenization methods should adapt their merging criteria based on token characteristics—shorter tokens and domain-specific patterns need different frequency thresholds to get fair representation in the vocabulary.

This paper introduces AdaBPE, an improved tokenization method that adapts how subword units are created based on token length and type (general vs. domain-specific language).

trainingefficiency

Direct Pass@$k$ Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning

Oct 22, 2026

Mahdi Salmani, Adel Javanmard, Devansh Gupta et al.

Directly optimizing pass@k for pruning outperforms separate quality and diversity strategies by naturally balancing both objectives, achieving better results especially when keeping moderate numbers of candidate solutions.

This paper tackles test-time reasoning efficiency by proposing a pruning method that directly optimizes pass@k—the probability of finding at least one correct solution. Rather than separately balancing solution quality and diversity, the method unifies both objectives through a single metric that naturally rewards good partial solutions while penalizing redundancy.

reasoningefficiencyevaluation

Low-Bit N-Gram Embeddings for Language Models

Oct 22, 2026

Stephen Chung, Wenyu Du

You can compress n-gram embedding tables to 4-bit precision and save 75% of their memory with minimal performance loss—or use the freed memory to make the model better overall.

This paper shows that n-gram embedding tables in language models can be stored at low precision (4 bits) instead of full precision, reducing memory usage by ~75% while keeping 99% of the performance benefit. The authors test quantization methods on models from scratch and on large existing models like DeepSeek, finding that saved memory can be reinvested to improve overall performance.

efficiencytrainingarchitecture

What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

Oct 22, 2026

Jinya Sakurai, Xun Xu, Jaehong Yoon

For faster parallel token generation in diffusion LMs, a small learned policy (856K params) using confidence + hidden states can reliably decide which predictions to commit, improving throughput without sacrificing accuracy—and the policy transfers across different base models.

This paper studies how to decide which token predictions to commit to in parallel-decoding diffusion language models. The authors systematically analyze what makes a good unmasking policy—a learned model that decides when predictions are confident enough to finalize—by testing different supervision signals, inputs, and model sizes.

efficiencytrainingevaluation

Rethinking Sequence Compression with Test-time Regression

Oct 22, 2026

Sihan Yang, Jie Huang, Hanxuan Li et al.

Test-time regression offers a more effective way to compress token sequences than standard pooling or projection methods, enabling better efficiency in both language and multimodal models without sacrificing performance.

This paper proposes TTR-Compressor, a new method for compressing sequences of tokens in language and multimodal models. Instead of using pooling or learned projections, it frames compression as a regression problem solved at test time with coefficients computed efficiently in closed form.

efficiencyarchitecturemultimodal

LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data

Oct 21, 2026

Zhennan Wan, Jianfei Chen

You can train LLMs to handle longer contexts better using only self-generated data—no external corpora or teacher models needed—by having the model create its own reasoning tasks and learning from them.

LongZero enables LLMs to improve their long-context reasoning abilities without external data by having models generate their own training tasks. The method creates multi-hop reasoning and retrieval tasks by first synthesizing compact question-answer pairs with evidence, then expanding that evidence into long multi-document contexts, before training with reinforcement learning.

trainingreasoningefficiency

PhysiCalWorld: In-Context Physics Calibration for World Models

Oct 21, 2026

Yunhao Luo, Wentao He, Yinpei Dai et al.

World models can learn to calibrate themselves to scene-specific physics by encoding exploratory interaction data into reusable descriptor tokens, improving prediction accuracy without requiring physics labels or fine-tuning.

PhysiCalWorld improves world models by learning to identify and adapt to the unique physics properties of a scene (like object mass or friction) through brief exploratory interactions. Instead of fine-tuning or using physics equations, it encodes these properties into a compact token that guides predictions, making visual planning more accurate in novel environments.

reasoningefficiency

Asymmetric On-Policy Distillation Learns In-Context Exploration

Oct 21, 2026

Sriyash Poddar, Yanda Bao, Jacob Krantz et al.

You can teach agents efficient exploration by having a privileged teacher label actions conditioned on the student's real exploration history—no RL needed, just supervised learning with asymmetric information access.

ASTEROID is a student-teacher learning framework that teaches agents to explore efficiently under partial observability without reinforcement learning. A privileged teacher with full state information labels actions based on the student's actual exploration history, allowing the student to learn when and how to gather information.

trainingagentsefficiency

VX-One: Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift

Oct 21, 2026

Juan S. Santillana

Typed decision heads are faster and safer than generative models for security triage, but their calibration degrades significantly under temporal drift, and they still fail to reliably detect truly novel attacks without explicit unknown options.

VX-One is a small neural network that makes fast, structured security decisions for network intrusion detection by outputting typed primitives (binary gates, closed-set choices, ordinal scores) instead of raw predictions.

safetyevaluationefficiency

Industrial ReMEmbR: Scaling Spatiotemporal Memory to Robot Fleets

Oct 20, 2026

Jad Yammine, Malek Wahidi, Abrar Anwar et al.

Splitting large vision-language models at their architecture boundary—vision encoder on robots, language decoder on servers—enables efficient fleet-scale memory systems that generate descriptions on-demand rather than at write time, cutting server costs while preserving retrieval quality.

IndustrialRemembr extends a spatiotemporal memory system to robot fleets by splitting vision-language models between edge devices and servers. Robots encode video into embeddings and tokens without uploading raw frames, while an AI agent retrieves memories on demand and generates descriptions only when needed, reducing computation and storage costs while maintaining GDPR compliance.

efficiencyagentsmultimodal

Efficient Edge-aware Attention Network for Graph Generation

Oct 20, 2026

Vincent Jung, Alba Carballo-Castro, Yiming Qin et al.

Graph generative models can achieve significant memory and speed improvements by redesigning attention to work with hardware-optimized implementations, making graph foundation models more practical to scale.

This paper proposes an efficient attention mechanism for graph generation models that reduces GPU memory usage by 65% while maintaining generation quality. The key innovation is a head-wise score modulation approach that incorporates graph structure into attention, implemented using PyTorch's FlexAttention to leverage modern hardware optimizations.

efficiencyarchitecturescaling

Text-to-Image Generative Modeling Without Language Pretraining

Oct 19, 2026

Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas et al.

You can build competitive text-to-image systems without relying on pretrained language models, which could make these systems cheaper and more accessible to build from scratch.

Halcyon shows that text-to-image models don't need large language pretraining to work well. By training from scratch on 100M image-text pairs, Halcyon achieves better results than existing systems while using far less compute—18× to 275× less than competitors. This challenges the assumption that internet-scale language models are essential for image generation.

trainingmultimodalefficiency

WeAct: Dynamic Weight Switching as Agent Actions

Oct 19, 2026

Rui Shi, Xunjian Yin, Jiacheng Sang et al.

Instead of permanently merging all skills into one model, let agents choose which skill module to activate at each step—this prevents forgetting old capabilities while enabling flexible, modular skill composition.

WeAct lets AI agents dynamically switch between specialized skill modules (LoRA adapters) during task execution, treating adapter selection as an action the agent can take. This solves catastrophic forgetting—where learning new skills degrades old ones—by keeping a frozen backbone and training separate adapters for different capabilities.

trainingagentsefficiency

LatComm: Reusable Context Compression for Latent Communication between Large Language Models

Oct 19, 2026

Huawei Lin, Jinting Liu, Tan Yu et al.

Separating context encoding from reading allows LLM systems to compress context once and reuse it across multiple questions with different models, dramatically reducing computation and latency in multi-turn scenarios.

LatComm enables efficient communication between multiple LLMs by compressing context once into a compact, reusable representation that different models can selectively read.

efficiencymultimodalarchitecture

Kronecker Delta Rule for Long Context Linear Attention

Oct 19, 2026

Anamaria-Roberta Hartl, Levente Zólyomi, Michael List et al.

The Kronecker Delta Rule offers a theoretically grounded way to update associative memories in attention models that's both stable and efficient for long-context tasks, outperforming existing linear attention methods.

KronLSTM introduces a new memory update mechanism for linear attention models that modifies memories along a single direction, reducing interference with stored information. This approach provides mathematically stable memory dynamics without extra normalization, making it effective for processing long documents and sequences.

architectureefficiencyreasoning

Competing Learning in Hybrid Architectures

Oct 19, 2026

Xingyu Shen, Yingfa Chen, Ziqing Qiao et al.

In hybrid architectures mixing recurrence and attention, recurrence dominates learning of retrieval functions. Staged training that temporarily freezes recurrence allows attention to develop, significantly improving model performance.

This paper identifies a 'Competing Learning' problem in hybrid language models that combine recurrent and attention layers: recurrence learns retrieval functions faster, blocking attention from learning. The authors propose staged training—freezing recurrence updates mid-training—to let attention catch up, improving performance on language modeling and retrieval tasks.

trainingarchitectureefficiency
architectureefficiencyreasoning

Acceptance Is Not Speed: Length Inflation in Training-Free Lossy Speculative Decoding

Oct 16, 2026

Chiatzen Wang, Bill Xu, Haochen Gu et al.

Accepted tokens per round is a misleading metric for speculative decoding speedup—you must also measure completion length and verification rounds, as relaxed verification often trades token acceptance for longer outputs with minimal real-world speed gains.

This paper reveals a critical flaw in how we measure speedup in speculative decoding with relaxed verification: while accepting more tokens per round looks faster, completions actually become 1.54× longer, offsetting efficiency gains.

efficiencyevaluation

Steering beyond Language: Improving Multilingual Training-Free Text Embeddings via Paired-Residual Subspace Removal

Oct 16, 2026

Ailiang Lin, Zhuoyun Li, Yusong Wang et al.

Language-specific noise in LLM embeddings can be removed at inference time using paired multilingual sentences, significantly improving cross-lingual and monolingual embedding quality without any training.

This paper tackles multilingual text embeddings from frozen language models by identifying that semantically equivalent sentences misalign across languages due to language-specific geometry.

multimodalefficiencyevaluation