PreprintNot peer reviewedA preprint reports a reinforcement learning framework that blends retrieval with reasoningOriginal title: UR$^2$: Unify RAG and Reasoning through Reinforcement LearningUR2 trains language models to decide when to retrieve information, aiming to match larger commercial models on several tasks.Weitao Li et al. · Added 4 Oct 2026
PreprintNot peer reviewedPreprint proposes boosting diverse reward models to curb reward hacking in RLHFOriginal title: RMB: Reward Model Boosting Mitigates Reward HackingA preprint reports that combining diverse reward models through a boosting method improves reward accuracy and reduces reward hacking in RLHF.Jiabin Fan et al. · Added 4 Oct 2026
PreprintNot peer reviewedResearchers report models keep usefulness but lose safety under coded-text shiftsOriginal title: LLM Alignment–Utility Asymmetry under Semantic-Preserving TransformationsA preprint finds that once models learn a simple text-substitution code, they keep most task performance but become far more likely to comply with harmful requests.Mohan Li et al. · Added 4 Oct 2026
Journal articleReview status unknownA preprint reports no demographic bias detected in an LLM agent action benchmarkOriginal title: AgentFairBench: Do LLM Agents Discriminate When They Act?The authors present a benchmark for demographic disparity in LLM agent actions and report no bias signal for one tested model once noise is corrected.Triveni Morla et al. · Added 4 Oct 2026 · Released Jan 2026
PreprintNot peer reviewedResearchers report LLMs struggle to translate security requests into exact Kali Linux commandsOriginal title: KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable RewardsA preprint introduces KaliBench, a benchmark showing open-weight models often fail to turn plain-English security requests into correct CLI commands.Pengfei Li et al. · Added 4 Oct 2026
PreprintNot peer reviewedFrom FuriosaAIResearchers report faster RL rollouts via self-speculative decodingOriginal title: EfficientRollout: System-Aware Self-Speculative Decoding for RL RolloutsA preprint describes EfficientRollout, a system-aware self-speculative decoding method that speeds up LLM reinforcement-learning rollouts without hurting final quality.Added 4 Oct 2026
PreprintNot peer reviewedResearchers build a fine-grained Chinese LLM safety guardrail trained on new datasetsOriginal title: CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference AlignmentA preprint introduces CHILLGuard, a Chinese content-safety guardrail with a 31-category taxonomy and preference-tuned training.Wenbo Yu et al. · Added 4 Oct 2026
Journal article✓ Peer reviewedBenchmark finds LLM agents widely vulnerable to hidden tool-metadata attacksOriginal title: MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP ServersA new benchmark shows many LLM agents can be tricked by malicious instructions hidden in tool descriptions, not just tool outputs.Zhiqiang Wang et al. · Added 4 Oct 2026 · Released Mar 2026
PreprintNot peer reviewedResearchers propose an energy-landscape theory for why jailbreaks work on diffusion language modelsOriginal title: Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape AnalysisA preprint frames dLLM safety as an energy barrier and derives three training-free signals that detect known jailbreak attacks.Thong Bach et al. · Added 4 Oct 2026
Journal articleReview status unknownResearchers report an AI agent framework for controlling optical network hardware via natural languageOriginal title: Agentic AI for Scalable and Robust Optical Systems ControlA framework called AgentOptics lets large language models control real optical devices through a standard tool protocol, with authors reporting high task success rates.Zehao Wang et al. · Added 4 Oct 2026 · Released Jan 2026
Journal articleReview status unknownResearchers report widespread bias in LLM-based judges and propose a debiasing training methodOriginal title: Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing OptimizationA new benchmark finds LLM judges remain biased across 12 types, and a training method reduces this while keeping accuracy.Hongli Zhou et al. · Added 4 Oct 2026
PreprintNot peer reviewedResearchers report security flaws in world models used by AI agentsOriginal title: False Prophets: On the Security of World Models in Agentic SystemsA preprint finds text-based world models used by AI agents can be tricked into dangerous mispredictions up to 95% of the time.Erik Imgrund et al. · Added 4 Oct 2026
PreprintNot peer reviewedResearchers report encode-aware serving design boosts multimodal LLM throughputOriginal title: EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language ModelsA preprint proposes treating the Encode stage as the control point for serving multimodal LLMs, reporting large goodput gains over existing systems.Kunxiong Zhu et al. · Added 4 Oct 2026
Journal articleReview status unknownStudy tests ways to guide LLMs toward writing and fixing more secure codeOriginal title: Guiding AI to Fix Its Own Flaws: An Empirical Study on LLM-Driven Secure Code GenerationResearchers examine whether self-written vulnerability hints and detailed feedback help large language models avoid or repair insecure code.Hao Yan et al. · Added 4 Oct 2026
PreprintNot peer reviewedFrom GoogleSurvey maps how ‘agentic skills’ could make AI agents more reliable and secureOriginal title: A Systematic Survey of Agentic Skills: Architecture, Lifecycle, and SecurityA preprint surveys the emerging field of reusable 'agentic skills' for AI agents, covering design, memory, evaluation and security risks.Added 4 Oct 2026
Conference paperReview status unknownResearchers link knowledge graphs to LLM agents for automated hardware formal verificationOriginal title: Knowledge Graphs, the Missing Link in Agentic AI-based Formal VerificationA conference paper reports that a knowledge-graph-grounded multi-agent workflow improves LLM-generated formal assertions for hardware designs.Vaisakh Naduvodi Viswambharan et al. · Added 4 Oct 2026 · Released May 2026
PreprintNot peer reviewedA preprint describes ThinkingGuard, a guard model for detecting implicit multimodal safety risksOriginal title: ThinkingGuard: Decoding Implicit Hazards via Step-by-Step Risk Attribution in Multimodal Large Language ModelsA preprint presents a new dataset and training method that teach a guard model to reason about hidden safety risks across text and images.Ruochen Zhang et al. · Added 4 Oct 2026
PreprintNot peer reviewedA preprint reports a method to catch silent errors in LLM-generated optimisation codeOriginal title: ReLoop: Structured Modeling and Behavioral Verification for Reliable LLM-Based OptimizationResearchers describe ReLoop, a pipeline that combines structured prompting and solver-based checks to catch LLM optimisation code that runs fine but models the wrong problem.Junbo Jacob Lian et al. · Added 4 Oct 2026
PreprintNot peer reviewedSurvey maps how video generation models are post-trained to follow intent and stay coherentOriginal title: Video Generation Models: A Survey of Post-Training and AlignmentA preprint surveys methods for aligning pretrained video generation models with user intent, physics, and safety.Chaoyu Li et al. · Added 4 Oct 2026
Conference paperReview status unknownResearchers report LLMs often skip security protections in hardware code generationOriginal title: HardSecBench: Benchmarking the Security Awareness of LLMs for Hardware Code GenerationA new benchmark finds that LLMs frequently write functionally correct hardware code that still lacks important security protections.Qirui Chen et al. · Added 4 Oct 2026 · Released Jun 2026
