PreprintNot peer reviewedA preprint reports personalized reward models miss preference signals in contextOriginal title: Using Context Is Not Enough: Test-Time Training for Personalized Reward ModelingA preprint proposes test-time weight updates to help reward models better use labelled preference pairs from user history.Bohao Wang et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom TencentA preprint reports hypergraph method for linking retrieved evidence in LLM reasoningOriginal title: HyperReCo: Retrieving and Connecting Evidence with Hypergraph Neural Networks for LLM Multi-hop ReasoningResearchers describe HyperReCo, a hypergraph neural network that retrieves and explicitly links multi-hop evidence for LLM question answering.Added 1 Nov 2026
PreprintNot peer reviewedPreprint proposes framework for detecting self-organisation in LLM agent societiesOriginal title: Population Physics, Population Problems: Safety and Emergence in LLM SocietiesA preprint introduces a statistical framework to detect unplanned collective behaviour across simulated LLM societies without reading agent language.Adrian de Wynter · Added 1 Nov 2026
PreprintNot peer reviewedFrom Microsoft ResearchA preprint reports RL-trained self-refinement lets small LLMs rival 32B-235B models on codingOriginal title: RefineRL: Advancing Competitive Programming with Self-Refinement Reinforcement LearningA preprint describes RefineRL, an RL method that teaches compact 4B language models to skeptically refine their own competitive-programming solutions.Added 1 Nov 2026
PreprintNot peer reviewedResearchers propose streaming benchmark for AI reading long ambulatory ECGsOriginal title: ECG-Scroll: A Long-Horizon, Streaming Benchmark and Agent Environment for Interpretation of Ambulatory ElectrocardiogramsA preprint introduces a streaming benchmark testing whether language-model agents can spot rare cardiac events in day-long ECG recordings as they unfold.Haitao Li et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom AmazonPreprint proposes letting LLMs judge their own training curriculum during fine-tuningOriginal title: Internalizing Curriculum Judgment for LLM Reinforcement Fine-TuningA preprint reports a method, METIS, where a language model predicts which prompts will teach it the most, speeding up reinforcement fine-tuning.Added 1 Nov 2026
PreprintNot peer reviewedA preprint reports multimodal models struggle to read physical fields in new benchmarkOriginal title: PhysFieldBench: Can Multimodal Models Understand Physical Fields?PhysFieldBench tests whether multimodal AI models can infer physics from field images, and finds current models perform poorly.Yuezhou Ma et al. · Added 1 Nov 2026
PreprintNot peer reviewedA preprint reports task type shapes speculative decoding acceptance more than tree depthOriginal title: Acceptance Dynamics Across Cognitive Domains in Speculative DecodingA preprint finds chat prompts accept more speculative tokens than code or maths, with task type mattering more than tree depth.Seifeldin Abdellatif · Added 1 Nov 2026
PreprintNot peer reviewedResearchers report a membership inference attack on federated prompt tuningOriginal title: Leveraging Soft Prompts for Privacy Attacks in Federated Prompt TuningA preprint describes PromptMIA, an attack where a malicious server infers client data membership via adversarial soft prompts in federated prompt tuning.Quan Minh Nguyen et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom NVIDIAResearchers report a cheaper way to fine-tune diffusion models for reward signalsOriginal title: Scaling Reinforcement Learning for Diffusion Models via Velocity MatchingA preprint proposes fine-tuning diffusion and video models by matching velocities weighted by reward, skipping likelihood estimation entirely.Added 30 Oct 2026
PreprintNot peer reviewedBenchmark tests whether LLM agents can reconstruct full APT attack chains from logsOriginal title: CyberClear: A Benchmark for LLM Agent Systems on APT Attack Chain ProvenanceA preprint introduces CyberClear, a benchmark and harness for testing LLM agents on reconstructing APT attack chains from raw security logs.Qi Chen et al. · Added 30 Oct 2026
PreprintNot peer reviewedStudy finds LLMs conform more to in-group peers, less to out-groupsOriginal title: LLMs Trust Their Own: Identity-Dependent Conformity in Multi-Agent SystemsA preprint reports that language models shift how much they conform to incorrect peers depending on whether those peers share their assigned social identity.Liron Soffer et al. · Added 30 Oct 2026
PreprintNot peer reviewedResearchers test how well LLMs resist confidently stated misinformation in health, finance, lawOriginal title: AdversaRiskQA: An Adversarial Factuality Benchmark for High-Risk DomainsA preprint introduces a benchmark testing whether LLMs detect and correct confidently phrased misinformation in high-risk domains.Adam Szelestey et al. · Added 30 Oct 2026
PreprintNot peer reviewedResearchers report LLMs’ own safety judgement can be weaponised to bypass refusalsOriginal title: Safety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior AttackA preprint describes a jailbreak that exploits a model's own safety-classification skill, finding better safety judgement means greater vulnerability.Long P. Hoang et al. · Added 30 Oct 2026
PreprintNot peer reviewedPreprint proposes an LLM agent that designs reward functions more efficientlyOriginal title: Beyond Scripted Search: Sample-Efficient Reward Discovery via Agentic Black-box OptimizationA preprint reports an agent-driven search system that finds reinforcement learning reward functions using far fewer costly training runs.Minghao Li et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom HuaweiA preprint reports a skill-based approach to train long-horizon code agentsOriginal title: CodeSkill: Latent Skill Abstraction for Long-Horizon Code AgentsResearchers propose CodeSkill, a framework that abstracts code-agent trajectories into reusable skills to improve reinforcement learning efficiency.Added 30 Oct 2026
PreprintNot peer reviewedA preprint proposes requesting short LLM hints to cut costly inference while keeping accuracyOriginal title: Pay for Hints, Not Answers: LLM Shepherding for Cost-Efficient InferenceResearchers report that giving small models short prefixes from a large model, rather than full answers, cuts costs while preserving accuracy.Ziming Dong et al. · Added 30 Oct 2026
PreprintNot peer reviewedPreprint proposes closed-form hindsight credit assignment for long-horizon LLM agentsOriginal title: GraphHCA: Closed-Form Hindsight Credit Assignment for Long-Horizon LLM AgentsA preprint presents GraphHCA, a model-free method that derives step-level credit for LLM agents without an auxiliary hindsight model.Haodong Zhu et al. · Added 30 Oct 2026
PreprintNot peer reviewedFrom Kivira HealthPreprint proposes K-Bench, a clinician-calibrated safety test for mental-health chatbotsOriginal title: K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversationsA preprint describes K-Bench, a benchmark testing 125 AI model setups on simulated high-risk mental health conversations against clinician judgements.Added 30 Oct 2026
PreprintNot peer reviewedFrom Ant GroupPreprint proposes relation tokens as building blocks for graph foundation modelsOriginal title: Relation-Aware Graph Foundation ModelA preprint introduces REEF, a graph foundation model that uses relation types rather than nodes or datasets as its basic unit for pre-training and transfer.Added 30 Oct 2026
