PreprintNot peer reviewedFrom MetaPreprint proposes multi-sample tuning to fix diversity and bias in generative modelsOriginal title: Preference Optimization with Multi-Sample ComparisonsA preprint introduces mDPO and mIPO, which compare groups of model outputs rather than single pairs to improve diversity and reduce bias.Added 7 Oct 2026 · Released Mar 2025
PreprintNot peer reviewedResearchers report a faster self-speculative decoding method called SEEDOriginal title: SEED: Self-Speculative Decoding via Implicit Encoder-DecoderA preprint describes SEED, which reuses cached deep representations to speed up LLM decoding without hurting output quality.Hankun Lin et al. · Added 7 Oct 2026
PreprintNot peer reviewedResearchers propose MASBench to test multi-agent LLM teamwork with hidden informationOriginal title: MASBench: Benchmarking LLM-based Multi-Agent Collaboration under Partial ObservabilityA preprint introduces MASBench, a benchmark that evaluates LLM multi-agent collaboration mechanisms under partial observability rather than global visibility.Qizhi Chu et al. · Added 7 Oct 2026
PreprintNot peer reviewedPreprint describes method to extract a safety-steering signal from aligned LLMsOriginal title: Reward Stealing Attack on Large Language ModelsA preprint reports a technique that reverse-engineers an aligned model's hidden safety reward to craft transferable jailbreak attacks.Jiaming Qian et al. · Added 7 Oct 2026
PreprintNot peer reviewedResearchers profile how robot vision-language-action models run on edge hardwareOriginal title: Beyond LLM Serving: Characterizing Vision-Language-Action Workloads for Embodied AI System DesignA preprint examines how vision-language-action robot models behave when run one-at-a-time on edge GPUs and onboard chips.Seonghun Jung et al. · Added 7 Oct 2026
Conference paperReview status unknownUnsupervised training method cuts LLM answer flip-flopping across reworded promptsOriginal title: Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMsResearchers report an unsupervised method that makes language models answer more consistently when prompts are reworded, without needing labelled data.Parsa Hejabi et al. · Added 7 Oct 2026 · Released Jan 2026
PreprintNot peer reviewedA preprint reports a hyperparameter-free way to combine distillation and verifiable rewardsOriginal title: On-policy Distillation with Verifiable RewardResearchers describe OPDVR, a gating method that merges on-policy distillation with verifiable-reward reinforcement learning for reasoning models.Wenze Lin et al. · Added 7 Oct 2026
Conference paperReview status unknownFrom GoogleResearchers report AI coding agents struggle to safely migrate Java projects to newer JDKsOriginal title: FreshBrew: A Benchmark for Evaluating AI Agents on Java Code MigrationA new benchmark finds AI agents often migrate Java codebases successfully but sometimes do so by gaming the tests.Added 7 Oct 2026 · Released Apr 2026
PreprintNot peer reviewedResearchers propose a memory system that delays compressing evidence in long documentsOriginal title: CoEM: Empowering Long-Context Reasoning with Commit-on-Evidence MemoryA preprint describes CoEM, a memory agent that keeps uncertain evidence verbatim until later context clarifies its relevance, improving long-context question answering.Jingguang Li et al. · Added 7 Oct 2026
PreprintNot peer reviewedFrom ZoomPreprint finds safety-layer defenses can be localised yet still bypassedOriginal title: Refusal Localizes, the Damage Relocates: Safety Layers Under Few-Sample Fine-TuningA preprint reports that freezing or repairing the layers tied to refusal can still be defeated by attackers who know the defense.Added 7 Oct 2026
PreprintNot peer reviewedResearchers build a chess-puzzle benchmark to test prompt optimisation on frozen LLMsOriginal title: Benchmarking Prompt Optimization of Large Language Models With ChessA preprint uses Lichess puzzles to compare six automatic prompt-optimisation methods across eight frozen language models.Timothée Lesort et al. · Added 7 Oct 2026
PreprintNot peer reviewedimecResearchers report a tool that audits model specs for internal contradictionsOriginal title: Detecting Inconsistencies in Model Specifications with LLM-as-Verifier ReasoningA preprint describes VeriSpec, which uses an LLM verifier to find conflicting rules in OpenAI's Model Spec.Added 7 Oct 2026
PreprintNot peer reviewedResearchers report reinforcement learning fixes an AI theorem-falsification blind spotOriginal title: Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement RepairsA preprint finds that counterexample-only imitation training harms false-theorem detection, while reinforcement learning with a verifier reward repairs and improves it.Omar Farouk Zouak et al. · Added 7 Oct 2026
PreprintNot peer reviewedFrom PleIAsResearchers report a fully synthetic Wikipedia-derived corpus trains competitive small LLMsOriginal title: It's All Training: A Fully Synthetic Single-Stage Recipe for LLMsA preprint reports that models trained only on a synthetic corpus derived from Wikipedia can match similarly sized open models using far fewer tokens.Added 7 Oct 2026
Conference paperReview status unknownFrom IBMLLM-based tool flags Linux kernel patches fixing serious memory bugsOriginal title: What Do They Fix? LLM-Aided Categorization of Security Patches for Critical Memory BugsA conference paper presents DualLM, a tool combining an LLM and a fine-tuned small model to spot Linux patches fixing out-of-bounds and use-after-free bugs.Added 7 Oct 2026 · Released Jan 2026
PreprintNot peer reviewedA preprint proposes a framework to test whether LLMs reflect group differences in surveysOriginal title: Population Fidelity: Evaluating Population Representativeness in LLMsResearchers introduce Population Fidelity, a framework separating accuracy, variation and structure to evaluate how well LLMs represent subgroup differences in survey responses.Neemias B. da Silva et al. · Added 7 Oct 2026
PreprintNot peer reviewedResearchers report a method blending RL and distillation for LLM post-trainingOriginal title: Distilled Reinforcement Learning for LLM Post-trainingA preprint proposes Distilled RL, which folds teacher guidance into reinforcement learning to help student models learn new skills more selectively.Chen Wang et al. · Added 7 Oct 2026
PreprintNot peer reviewedA preprint reports a gating method to separate number IDs from maths in LLMsOriginal title: ANI: Adaptive Numerical Injection for Unifying Semantic and Arithmetic Representations in Numerical ReasoningResearchers propose ANI, which selectively injects numerical embeddings into LLMs only when a number is used arithmetically, not as an ID.Jinsung Jeon et al. · Added 7 Oct 2026
PreprintNot peer reviewedResearchers report LLMs’ reasoning text often doesn’t match their internal computationsOriginal title: Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability AlignmentA preprint introduces a metric to measure whether an LLM's chain-of-thought matches what the model actually computes internally, and shows post-training can improve this.Yihuai Hong et al. · Added 7 Oct 2026
PreprintNot peer reviewedPreprint proposes dependency graphs to give RL agents step-by-step creditOriginal title: Dependency-Aware Reward Shaping for Agentic Reinforcement LearningA preprint reports a reward-shaping method that tracks task prerequisites to assign step-level credit in agentic reinforcement learning.Ziyi Chen et al. · Added 7 Oct 2026
