PreprintNot peer reviewedPreprint proposes hybrid reward model that switches between fast and slow judgingOriginal title: Fast-Slow Thinking RM: Efficient Integration of Scalar and Generative Reward ModelsA preprint reports a reward model that mixes quick scalar scoring with chain-of-thought reasoning, cutting token use while keeping accuracy.Jiayun Wu et al. · Added 30 Oct 2026
PreprintNot peer reviewedPreprint proposes letting reasoning models consult helper agents during trainingOriginal title: AIPO: Learning to Reason from Active InteractionA preprint reports a training method where language models query verification, knowledge and reasoning helpers mid-rollout to learn beyond their own exploration limits.Junnan Liu et al. · Added 30 Oct 2026
PreprintNot peer reviewedStudy finds LLMs judge morality differently by language but reason similarly underneathOriginal title: One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral ReasoningA preprint reports that language shifts LLMs' moral verdicts, yet their underlying reasoning stays broadly convergent across six languages.Sualeha Farid et al. · Added 30 Oct 2026
PreprintNot peer reviewedResearchers propose a reliability-engineering framework for autonomous and self-evolving AI systemsOriginal title: Reliability Engineering for AI Systems: Challenges, Methods, and DirectionsA preprint adapts classic reliability engineering—FMEA, monitoring, FRACAS—to assess AI agents and self-evolving systems beyond benchmark accuracy.Rong Pan et al. · Added 30 Oct 2026
PreprintNot peer reviewedResearchers build a taxonomy to profile how AI agents behave on everyday tasksOriginal title: AgentHabit: Characterizing Distinct Behaviors of Agents on Everyday TasksA preprint introduces HABIT and AgentHABIT, tools that profile 18 AI agents on 23 behavioural axes beyond task success.Woojung Song et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom Adobe ResearchPreprint proposes splitting long-context reasoning across worker and driver LLMsOriginal title: DISCO: Distributed Long Context Scaling with Grounding-Reasoning DisaggregationA preprint reports that disaggregating grounding from reasoning across distributed LLMs curbs long-context accuracy collapse.Added 30 Oct 2026
PreprintNot peer reviewedFrom DeepAuto.aiResearchers propose entropy-based credit assignment to improve LLM reasoning explorationOriginal title: Surprising Success, Repeated Failure: Entropy-Guided Credit Assignment for Exploration in LLM ReasoningA preprint reports a token-level credit method that treats confident failures and uncertain successes differently to boost reasoning exploration in LLMs.Added 30 Oct 2026
PreprintNot peer reviewedFrom Kensho TechnologiesStudy finds LLM judges falter on finance questions they cannot answer themselvesOriginal title: No Free Labels: Limitations of LLM-as-a-Judge Without Human GroundingA preprint finds LLM-as-a-Judge systems agree with human experts mainly when they can already answer the question or get a correct reference.Added 30 Oct 2026
PreprintNot peer reviewedA preprint reports fine-tuning disturbs LLM alignment more than RL-based adaptationOriginal title: How LLM Task-Adaptation Reshapes Alignment: A Multi-dimensional Study of Behavioral and Representational DriftResearchers report that supervised fine-tuning causes far more alignment drift than RL-based task adaptation, with KL regularisation helping but not matching RL.James Elcock et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom XiaohongshuResearchers report training models to reason over documents using perturbed public textsOriginal title: Learning to Learn from Context: Synthetic Training from Perturbed Public DocumentsA preprint describes a pipeline that perturbs public documents to generate synthetic training data, boosting a student model's context-reasoning scores.Added 30 Oct 2026
PreprintNot peer reviewedA preprint reports forcing one ‘safety trigger token’ can curb LLM jailbreaksOriginal title: Decoding One Safety Trigger Token for Balancing Safety and Usability in Large Language ModelsResearchers report that decoding a single learned safety token as a prefix cuts harmful outputs while barely affecting usability.Haoran Gu et al. · Added 30 Oct 2026
PreprintNot peer reviewedA preprint reports exact aggregation fix for federated LoRA fine-tuningOriginal title: CFLoRA: Federated Fine-tuning of LLMs with Complementary Factors for Error-free AggregationA preprint presents CFLoRA, a federated LoRA scheme that splits update channels to avoid aggregation errors while updating both factors at once.Yanan Ma et al. · Added 30 Oct 2026
PreprintNot peer reviewedA preprint reports a knowledge-graph fast-slow framework for LLM agents under partial observabilityOriginal title: NeSyFS: A Neuro-symbolic Fast-Slow Thinking Framework for LLM Agent under Partial ObservabilityResearchers describe NeSyFS, which pairs a knowledge graph with fast and slow thinking modules to help LLM agents act under partial observability.Duo Xu et al. · Added 30 Oct 2026
PreprintNot peer reviewedA preprint benchmark tests how easily network-ops LLM agents are tricked by fake instructionsOriginal title: NetInjectBench: Benchmarking Indirect Prompt Injection in Tool-Using Large Language Model Agents for Network OperationsA preprint introduces NetInjectBench, finding naive LLM agents often execute unsafe network changes unless metadata-based checks are added.Ruksat Khan Shayoni et al. · Added 30 Oct 2026
PreprintNot peer reviewedResearchers report partial guidance improves LLM agent coordinationOriginal title: Symbolic Guidance for LLM Agents in Distributed Multiagent CoordinationA preprint finds that giving LLM agents moderate, pseudocode-light guidance outperforms both unguided and fully scripted coordination.Ben Rachmut et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom Changan AutomobileResearchers propose runtime guard that intervenes mid-task to block unsafe AI agent actionsOriginal title: PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time SafetyA preprint presents PROACT-Agent, a runtime safety guard for LLM agents, alongside a new bilingual benchmark of labelled interaction states.Added 29 Oct 2026
PreprintNot peer reviewedFrom Cisco ResearchResearchers report order-sensitive retrieval-and-policy training boosts multi-hop QA accuracyOriginal title: BRIDGE: Bilevel Retrieval-Credit-Aware Agentic Reinforcement LearningA preprint introduces BRIDGE, a bilevel method that trains retriever and language model jointly, reporting gains on multi-hop and medical QA tasks.Added 29 Oct 2026
PreprintNot peer reviewedA preprint describes GPU-native KV injection to speed up personalised LLM memory servingOriginal title: InferScale: GPU-Native KV Injection for Personalized LLM ServingA preprint reports that injecting precomputed memory as KV state instead of re-prefilling text cuts latency while keeping most of the accuracy.Peter Li et al. · Added 29 Oct 2026
PreprintNot peer reviewedResearchers report rendering untrusted text as images curbs prompt injectionOriginal title: Render Before Reading: Visual Rendering as a Prompt Injection DefenseA preprint finds that turning untrusted text into images makes vision-language models far less likely to obey hidden instructions.Jie Zhang et al. · Added 29 Oct 2026
PreprintNot peer reviewedResearchers report AI-authored stocktake of machine learning’s role in LHC physicsOriginal title: Machine learning for the LHC physics program: a 2025-2026 stocktakeA mostly AI-generated review argues machine learning has become infrastructure for LHC physics, not just a research tool.Jesse Thaler · Added 29 Oct 2026
