PreprintNot peer reviewedFrom BaiduResearchers propose highlight-then-summarize method to improve long-context LLM reasoningOriginal title: Highlight-Then-Summarize: Learning to Compress Evidence for Long-Context UnderstandingA preprint introduces H2S, a method that extracts evidence then summarises it before answering, reportedly boosting long-context benchmark scores.Added 7 Oct 2026
PreprintNot peer reviewedPreprint: cheap open-weight judge panels can often stand in for costly AI judgesOriginal title: RAIM: Robust Aggregation of Inexpensive Models for Hallucination DetectionA preprint reports that aggregating ten small open-weight language models can approach a frontier model's performance at a fraction of the cost for detecting hallucinations.Elia Onofri et al. · Added 7 Oct 2026
PreprintNot peer reviewedFrom IBMResearchers report belief queries help LLM agents act under uncertaintyOriginal title: Towards a Belief-Based World Model for LLM AgentsA preprint finds that letting language model agents query a world model's beliefs, not just simulate futures, improves decision-making under partial observability.Added 7 Oct 2026
PreprintNot peer reviewedA preprint proposes token-level advantage signals for training multimodal reasoning modelsOriginal title: Reinforcing Multimodal Reasoning via Token-Level Perception-Grounded Advantage EstimationResearchers report a method that scores individual tokens, not just whole answers, to guide reinforcement learning for multimodal reasoning.Zhihan Zhang et al. · Added 7 Oct 2026
PreprintNot peer reviewedFrom HuaweiTest-time method adapts LLM confidence calibration without labelled data, authors reportOriginal title: Test-time Calibration Learning for Large Language Model ReasoningA preprint proposes a label-free way to adjust an LLM's confidence estimates at test time using only its own generated answers.Added 6 Oct 2026
PreprintNot peer reviewedMethod estimates hidden pre-training gradients to curb LLM forgetting during fine-tuningOriginal title: Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language ModelsA preprint proposes generating pseudo pre-training data to estimate unknown gradients, helping LLMs retain general knowledge while learning new tasks.Bing Wang et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports an RL framework to help AI agents find more, not just one, correct answersOriginal title: SPADER: Step-wise Peer Advantage with Diversity-Aware Exploration Rewards for Multi-Answer Question AnsweringResearchers describe SPADER, a reinforcement learning method aimed at improving tool-using AI agents on questions with many correct answers.Qiming Shi et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports concept-graph distillation improves LLM-to-student training efficiencyOriginal title: Distilling LLM Reasoning into Graph of Concept PredictorsResearchers propose mirroring an LLM's reasoning steps as a graph of concept predictors to train cheaper student classifiers with less labelling.Ziyang Yu et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports LLMs struggle to clarify ambiguous coding requirementsOriginal title: ClarifyCodeBench: Evaluating LLMs on Clarifying Ambiguous Requirements for Code GenerationA new benchmark finds that even strong code-generating LLMs often fail to ask the right questions when requirements are unclear.Zheng Fang et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint links RLVR training collapse to loss of reasoning strategy diversityOriginal title: All Work And No Play Makes Jack a Dull Boy: Understanding and Preventing Catastrophic Strategy Collapse in RLVRResearchers propose a theory and fix for sudden accuracy drops during RLVR fine-tuning of reasoning LLMs.Qiyuan Huang et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports LLMs often misfire when recalling well-known algorithms, not just inventing new onesOriginal title: Are you Synthesizing or Recalling? Evaluating LLMs on Algorithmic Code RetrievalA preprint proposes testing whether LLMs can faithfully reproduce known algorithms, finding frequent silent errors across languages.Nickil Maveli et al. · Added 6 Oct 2026
PreprintNot peer reviewedFrom China Mobile Internet CompanyA preprint reports faster on-device LLM inference for multi-agent systems on edge chipsOriginal title: EdgeAgent: Orchestrating On-Device LLM inference for End-User Multi-Agent Systems on CPU-GPU Unified Memory ArchitecturesResearchers describe EdgeAgent, a system that speeds up on-device LLM inference for multi-agent workloads on CPU-GPU unified memory chips.Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports LLMs cling to internal concepts despite clearer promptsOriginal title: On the Limits of LLM Adaptability: Impact of Model-Internalized Priors on Annotation Task PerformanceA preprint finds language models rely on built-in notions of toxicity that prompting struggles to override.Etienne Casanova et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint benchmark compares KV-cache reuse methods across models and contextsOriginal title: KVShareArena: KV-Cache Reuse Across Contexts and Model CheckpointsA new benchmark called KVShareArena tests how well KV-cache repair methods hold up when context or model weights change.Xi Shi et al. · Added 6 Oct 2026
PreprintNot peer reviewedResearchers propose TRACE fine-tuning to handle knowledge conflicts in RAG systemsOriginal title: Parameters vs. Context: TRACE Fine-Tuning for Robust Retrieval-Augmented GenerationA preprint presents TRACE, a fine-tuning method using debate traces to help RAG models choose between retrieved context and internal knowledge.Zhengchen Huang et al. · Added 6 Oct 2026
PreprintNot peer reviewedA preprint reports LLMs miscoordinate in 38% of high-stakes multi-agent safety testsOriginal title: GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game TheoryA new benchmark finds frontier language models often fail to pick socially safe actions in game-theoretic safety scenarios.Pepijn Cobben et al. · Added 6 Oct 2026
PreprintNot peer reviewedPreprint finds LLMs can solve every math sub-step yet still fail to combine themOriginal title: Solving Every Step Is Not Enough: Milestone Oracles Reveal a Composition Gap in LLM Math ReasoningA preprint reports that giving models step-by-step milestones often fails to fix maths errors, pointing to a composition gap.Zhuohan Wang et al. · Added 6 Oct 2026
Conference paperReview status unknownStudy finds LLMs can often guess which AI model they are talking toOriginal title: Agent-to-Agent Theory of Mind: Testing Interlocutor Awareness among Large Language ModelsResearchers report that large language models can identify their conversational partners, and that this awareness creates both collaboration gains and new safety risks.Younwoo Choi et al. · Added 6 Oct 2026 · Released Jan 2025
PreprintNot peer reviewedPreprint reports targeted preference training cuts unsupported answers in a vision-language modelOriginal title: Reducing Hallucination in Multimodal Large Language Models through Hard Grounding Preference SupervisionA preprint describes refining a multimodal model with hard grounding preference pairs, reporting gains on benchmarks and judged pairwise comparisons.Qinwu Xu · Added 6 Oct 2026
PreprintNot peer reviewedPreprint proposes calibrated importance sampling to fix LLM RL training-inference mismatchOriginal title: Rethinking Training-Inference Mismatch in LLM Reinforcement Learning: Where It Arises and How to Correct ItA preprint reports a new correction, CIS, for mismatches between training and inference engines in reinforcement learning for large language models.Tianrun Yu et al. · Added 6 Oct 2026
