PreprintNot peer reviewedFrom NVIDIAResearchers report offline distillation method matches on-policy teacher training at lower costOriginal title: Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy DistillationA preprint describes Lightning OPD, an offline distillation method the authors say matches standard on-policy distillation while using far less compute.Added 6 Oct 2026
PreprintNot peer reviewedPreprint describes TokenCake, a cache-aware serving system for multi-agent LLM workloadsOriginal title: TokenCake: A KV-Cache-centric Serving Framework for LLM-based Multi-Agent ApplicationsA preprint reports that TokenCake improves GPU memory handling and latency for multi-agent LLM serving by managing idle and contended caches.Zhuohang Bian et al. · Added 6 Oct 2026
PreprintNot peer reviewedFrom MeituanTGRL turns rollout temperature into a training signal for LLM reasoning, authors reportOriginal title: TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMsA preprint proposes TGRL, which converts temperature-based rollout diversity into token-level training credit for reinforcement learning with verifiable rewards.Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports a test-time training method for code LLMs robust to noisy promptsOriginal title: Noisy Test-Time Reinforcement Learning for Code LLMsResearchers propose NTRL-Code, a label-free test-time reinforcement learning method that helps code LLMs cope with noisy, error-prone prompts.Xikai Yang et al. · Added 6 Oct 2026
PreprintNot peer reviewedNew benchmark finds LLM agents still struggle with power-grid analysis tasksOriginal title: PowerBench: A Benchmark for Agentic Retrieval and Reasoning in Power SystemsA preprint introduces PowerBench, a synthetic power-system dataset and benchmark, and reports that top LLM agents still fall short on multi-step retrieval and reasoning tasks.Xijing Wang et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports a self-distillation method that targets only flawed reasoning steps in LLMsOriginal title: Beyond Imitation: Reflective On-Policy Self-Distillation for LLM ReasoningResearchers propose ROSD, which corrects only the erroneous part of an LLM's reasoning trace instead of imitating a full reference solution.Ziqi Zhao et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint describes MetaCrit, a four-agent method to calibrate LLM self-critique by taskOriginal title: Knowing When to Critique: Task-Adaptive Metacognitive Regulation for Reliable LLM ReasoningA preprint reports a multi-agent system that varies how much an LLM critiques its own reasoning, improving truthfulness and safety without hurting accuracy.Xinmeng Hou et al. · Added 6 Oct 2026
PreprintNot peer reviewedResearchers report no single AI model family wins across retinal imaging benchmarkOriginal title: FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMsA preprint introduces FOCUS, a ten-dataset benchmark showing that no retinal imaging model type reliably beats the others on all measures.David Restrepo et al. · Added 6 Oct 2026
PreprintNot peer reviewedFrom Together AIResearchers report QATFactory, a framework for training LLMs to survive 4-bit deploymentOriginal title: QATFactory: A Versatile, Deployment-Aligned Framework for Quantization-aware Training and Distillation of LLMsThe authors report their open framework improves quantized LLM accuracy over post-training quantization across formats, scales and inference engines.Added 5 Oct 2026
PreprintNot peer reviewedResearchers report a token-level credit signal that boosts LLM maths reasoning accuracyOriginal title: Semifactual Credit-Augmented Policy OptimizationA preprint introduces SCAPO, which uses semifactual prompt checks to refine token-level credit assignment in LLM reinforcement learning.Junshu Pan et al. · Added 5 Oct 2026
Journal articleReview status unknownResearchers report widespread safety failures in LLMs controlling medical robotsOriginal title: Benchmarking the Safety of Large Language Models for Robotic Health Attendant ControlA study evaluating 72 LLMs for robotic health attendant control finds many comply with harmful medical instructions.Mahiro Nakao et al. · Added 5 Oct 2026
PreprintNot peer reviewedPreprint surveys latency attacks that can slow or exhaust AI systemsOriginal title: Deep Learning Latency Attacks and Defenses: A Cross-Domain Survey of Availability ThreatsA preprint reviews research on attacks that inflate AI inference cost or delay, and the defences proposed against them.Zonghua Gu et al. · Added 5 Oct 2026
PreprintNot peer reviewedResearchers report MLLMs struggle to detect AI-generated multimodal fake newsOriginal title: Can Multimodal Large Language Models Generate and Detect Multimodal Social Media Fake News?A preprint reports that most multimodal language models fall well short of human accuracy at spotting AI-generated fake social media posts.Jiyao Yang et al. · Added 5 Oct 2026
PreprintNot peer reviewedPreprint reports process-level checks emerging from outcome-only RL trainingOriginal title: Inducing Process Supervision from Outcome-Only Reinforcement LearningA preprint describes TIPS, a method that trains reasoning verifiers using only final-answer labels, yet reportedly yields step-level error checks as a side effect.Shengda Fan et al. · Added 5 Oct 2026
PreprintNot peer reviewedA preprint describes CATCH, a testbed exposing how coding RL models learn reward hackingOriginal title: CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RLA preprint introduces CATCH, a controllable testbed that reproduces reward hacking in coding RL and tests ways to detect and curb it.Shouli Wang et al. · Added 5 Oct 2026
PreprintNot peer reviewedPreprint reports AI agents trained by rehearsing their own imagined environment responsesOriginal title: EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement LearningA preprint describes EnvACE, a method where an LLM agent imagines environment feedback internally rather than querying real or simulated environments during training.Zishan Xu et al. · Added 5 Oct 2026
PreprintNot peer reviewedA preprint reports SPLASH, a system that switches LLM attention layouts live during servingOriginal title: SPLASH: Switching Parallel Layouts of Attention with Seamless Handoff for LLM ServingThe authors describe SPLASH, a serving system that switches between attention layouts mid-run and reports notable throughput gains over fixed layouts.Chuan Liu et al. · Added 5 Oct 2026
PreprintNot peer reviewedFrom MetaPreprint proposes Meta-SecAlign to cut prompt-injection risk without losing agent utilityOriginal title: Meta-SecAlign: Training LLMs against Prompt Injection for Robust AgentsThe authors report a fine-tuning method that defends large language models against prompt injection while largely preserving their usefulness on agentic tasks.Added 5 Oct 2026
PreprintNot peer reviewedFrom Tencent HunyuanA preprint reports RL gains in LLMs trace to a low-dimensional activation geometryOriginal title: Learning to Steer, Steering to See: Unveiling the Geometry of RLVR in Large Language Models via Trainable VectorsThe authors find reinforcement learning's effect on large language models can be captured by a few trainable activation vectors, and build a stabiliser on this idea.Added 5 Oct 2026
PreprintNot peer reviewedResearchers report sparse circuit repairs can fix LLM reasoning errors without broad side effectsOriginal title: RESCUE: Repairing Language Model Errors to Sparse Circuits via Reinforcement LearningA preprint describes locating and retuning small LLM circuits to fix math and medical QA errors while preserving other skills.Chuanpu Liu et al. · Added 5 Oct 2026
