PreprintNot peer reviewedA preprint proposes a critic-free way to train RL agents on stale rolloutsOriginal title: On KL-Regularized Policy OptimizationA preprint introduces KLPO, which anchors KL regularisation at the data-collecting sampler to avoid importance weights and critics.Yifan Zhang · Added 10 Nov 2026
PreprintNot peer reviewedPreprint tests LLMs on wildfire tasks and finds hazards in public benchmarksOriginal title: AI4Fire: Evaluating Large Language Models on Wildfire TasksA preprint evaluates six core LLMs on wildfire tasks bare and grounded, finding tool access helps most but public datasets carry hidden flaws.Yue Zhao et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom Bespoke LabsResearchers report new benchmark testing AI agents on Slack, GitHub and docs togetherOriginal title: SlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code RepositoriesA preprint introduces SlackBench, finding even top AI agents struggle to combine chat, code and document evidence reliably.Added 10 Nov 2026
PreprintNot peer reviewedResearchers report a sparse mixed-precision attention method for faster long-context LLM inferenceOriginal title: TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM InferenceA preprint describes TriAttention, which mixes skipping, 4-bit and 8-bit attention to speed up long-context LLM generation while preserving quality.Hanzhi Zhang et al. · Added 10 Nov 2026
PreprintNot peer reviewedResearchers trace an LLM’s decision to call tools to a single internal vectorOriginal title: How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by SuppressionA preprint reports that a compact internal vector controls whether agentic LLMs decide to call a tool, shaped by suppression from analysis verbs.Xijie Gong et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom Lexsi LabsPreprint describes AgentTune, a unified library for training tool-using LLM agentsOriginal title: AgentTune: A Unified Training Lifecycle for Tool-Using LLM AgentsA preprint presents AgentTune, a source-available library that unifies building, training, evaluating and recovering tool-using LLM agents.Added 10 Nov 2026
PreprintNot peer reviewedFrom Microsoft ResearchResearchers report generative UI harness cuts agent dialogue and boosts task successOriginal title: When Interfaces Speak: Data-Aware Generative UI Harness for Active InteractionA preprint describes GenUI-Harness, which trains agents to build on-the-fly interfaces and reports large task-success and dialogue-reduction gains.Added 10 Nov 2026
PreprintNot peer reviewedResearchers report an evidence-review framework improves automated vulnerability repairOriginal title: EviRepair: Evidence-Driven Agents for Reliable Vulnerability RepairA preprint describes EviRepair, a multi-agent system that checks evidence consistency before generating vulnerability patches.Yilian Shi et al. · Added 10 Nov 2026
PreprintNot peer reviewedA preprint proposes giving AI personal agents memory with evolving relational statesOriginal title: EvoMemState: Evolving Memory State for Personal AgentsResearchers describe a framework that adds relational and reasoning states to stored facts so AI agents use memory more reliably.Zhishang Xiang et al. · Added 10 Nov 2026
PreprintNot peer reviewedA preprint proposes selective layer recovery to defend LLMs against malicious fine-tuningOriginal title: SLDR: Defending Against Malicious Fine-tuning via Selective Layers Recovery and Dynamic RoutingResearchers report a defense that routes a small safety adapter to malicious queries, cutting harmful outputs while keeping task accuracy.Hui Zhang et al. · Added 10 Nov 2026
PreprintNot peer reviewedFrom MicrosoftA preprint reports self-play method lets small models match larger ones at spotting hallucinationsOriginal title: Hallucination Self-Play: Bootstrapping Reinforced Detector via Evolved GeneratorResearchers describe a self-play framework where a generator and detector co-evolve to improve hallucination detection without extra human labels.Added 10 Nov 2026
PreprintNot peer reviewedA preprint reports smarter rollout budgeting improves search-agent training accuracyOriginal title: Information Gain-based Rollout Policy Optimization: An Adaptive Tree-Structured Rollout Approach for Multi-Turn Search AgentsResearchers propose allocating RL rollout computation by intermediate-state informativeness, reporting accuracy gains on search QA benchmarks.Yijun Zhang et al. · Added 9 Nov 2026
PreprintNot peer reviewedA preprint describes a defence against malware hidden inside shared neural network weightsOriginal title: NeuPerm: Disrupting Malware Hidden in Neural Network Parameters by Leveraging Permutation SymmetryResearchers report a method called NeuPerm that permutes neural network parameters to disrupt hidden malware without hurting model accuracy.Daniel Gilkarov et al. · Added 9 Nov 2026
PreprintNot peer reviewedFrom Alibaba GroupResearchers propose a graph-based method to trace failures in LLM memory systemsOriginal title: MemTrace: Tracing and Attributing Errors in Large Language Model Memory SystemsA preprint introduces MemTrace, a framework and benchmark for pinpointing which operation caused an LLM memory system to fail.Added 9 Nov 2026
PreprintNot peer reviewedFrom NEC Laboratories AmericaResearchers report skill-distilling reinforcement learning boosts LLM agent performanceOriginal title: SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement LearningA preprint describes SkillRL, which turns agent trajectories into reusable skills during reinforcement learning, reportedly beating memory-based baselines.Added 9 Nov 2026
PreprintNot peer reviewedFrom Lexsi LabsPreprint presents AuditKIT, a unified toolkit for evaluating language models and agentsOriginal title: AuditKIT: A Unified Evaluation Spine for Benchmarking, Judging, and Comparing Large Language Models and Agentic SystemsA preprint describes AuditKIT, a source-available library that unifies benchmarking, judging, and agent evaluation into one measurement framework.Added 9 Nov 2026
PreprintNot peer reviewedFrom Autolab AIResearchers report LLMs can partly predict which ML ideas will actually workOriginal title: Can LLMs predict the outcomes of machine learning experiments?A preprint finds frontier LLMs forecast ML experiment outcomes better than chance but still miss much of the variation.Added 9 Nov 2026
PreprintNot peer reviewedResearchers report multi-agent system with dual memory speeds up GPU kernel optimisationOriginal title: KernelClaw: A Multi-Agent Framework for GPU Kernel Optimization with Dual-Memory MechanismsA preprint describes KernelClaw, a multi-agent framework using expert knowledge and trajectory memory to optimise GPU kernels faster than prior methods.Qitong Sun et al. · Added 9 Nov 2026
PreprintNot peer reviewedResearchers find Bangla safety filters fail more when prompts are romanised or code-mixedOriginal title: BanglaShield: A Script-Robust Safety Classifier and Benchmark for Bangla Native, Romanized, and Code-Mixed PromptsA preprint reports LLM safety classifiers miss many unsafe Bangla prompts, especially when romanised, and introduces a compact classifier that performs more consistently.Mohsinul Kabir et al. · Added 9 Nov 2026
PreprintNot peer reviewedResearchers report a quality-diversity approach boosts research idea generation with LLM agentsOriginal title: IDEAgent: Agentic Quality-Diversity Search for Research Idea GenerationA preprint presents IDEAgent, a multi-agent system that the authors say generates more diverse, high-quality research ideas than prior methods.Varun Gumma et al. · Added 9 Nov 2026
