PreprintNot peer reviewedFrom Style3DAgentic system converts real robot videos into simulatable digital twinsOriginal title: Agentic Real2Sim: Physics-based World Modeling with Vision-Language AgentsA preprint describes an agent-based pipeline that turns recorded robot interactions into physics simulations for policy testing and training.Added 6 Nov 2026
PreprintNot peer reviewedSurvey maps how large language models could assist PDE research workflowsOriginal title: Large language models for partial differential equation workflowsA preprint reviews how LLMs might help discover, solve and optimise partial differential equation problems, and finds current benchmarks remain fragmented.Han Wan et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom XiaomiResearchers propose Prorsa to speed up agentic RL rollout via adaptive prefill-decode splitOriginal title: Prorsa: Unleashing the Potential of Prefill-Decode Disaggregation in Agentic Reinforcement LearningA preprint presents Prorsa, a framework that forecasts prefill and decode demand to speed up agentic reinforcement learning rollouts.Added 6 Nov 2026
PreprintNot peer reviewedPreprint proposes lightweight sanitiser to protect LLM judges from score-hijacking promptsOriginal title: StableJudge: A Lightweight Adversarial Defense Framework for LLM-as-a-JudgeA preprint introduces StableJudge, a small detachable module that strips injected instructions before an AI judge scores text.Anning Hu et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom AmazonResearchers report RL-trained agent judges and diagnoses competitive-programming code submissionsOriginal title: Beyond the Verdict: Agentic Code Judging with Verifiable FeedbackA preprint describes Arbiter, an agentic verifier trained with reinforcement learning to judge and diagnose code submissions using checkable counterexamples.Added 6 Nov 2026
PreprintNot peer reviewedFrom JD.com / TencentPreprint proposes attention-aware KV cache compression method called BELTOriginal title: BELT : BEhavior-Aligned Low-Rank KV Cache Compression with Learned Top-k Rank AllocationA preprint reports a KV cache compression scheme that targets attention-output quality instead of raw key reconstruction, claiming strong accuracy and speed gains.Added 6 Nov 2026
PreprintNot peer reviewedStudy finds LLMs often justify human-rights violations through reasonable-sounding logicOriginal title: UDHRBench: Benchmarking Human Rights in Large Language ModelsA preprint introduces UDHRBench, a dataset testing whether large language models align with the Universal Declaration of Human Rights.Alak Raad et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom Luma AIResearchers report mixed results on whether robot agents truly learn or just retryOriginal title: SEAR: Self-Evolving Agentic RoboticsA preprint finds robotic agents sometimes just resample attempts rather than genuinely learning, with transfer depending on code, not task-specific tricks.Added 6 Nov 2026
PreprintNot peer reviewedFrom logsight.aiResearchers report lattice-based quantization shrinks LLM key caches without quality lossOriginal title: E8 Lattice Key Quantization for Low-Bit KV-Cache CompressionA preprint describes an E8 lattice vector quantization scheme that compresses LLM key-value caches heavily while preserving near-baseline quality and decode speed.Added 6 Nov 2026
PreprintNot peer reviewedPreprint reports training-free way to share KV caches between different-sized language modelsOriginal title: Think Before You Speak: Enabling Training-Free Latent Communication Between Language ModelsA preprint describes CacheFuse, a training-free method that blends KV caches from two language models to lift accuracy and cut latency.Rishi Khare · Added 6 Nov 2026
PreprintNot peer reviewedResearchers report a stealthy backdoor that leaks private RAG documents via a third-party modelOriginal title: RAGLeak: Supply-Chain Privacy Backdoors in Retrieval-Augmented GenerationA preprint describes a supply-chain backdoor that makes a RAG generator secretly encode retrieved private documents for later recovery.Yuchen Chen et al. · Added 6 Nov 2026
PreprintNot peer reviewedA preprint maps how verifier mistakes shape reinforcement learning for reasoning modelsOriginal title: Checking the Checkers: A Synthesis of Verifier Noise and Robustness in RLVRA survey of 58 studies finds that whether a reward checker's errors can be exploited matters more than how often it errs.Savyasachi Gupta et al. · Added 6 Nov 2026
PreprintNot peer reviewedA preprint reports a graph-based retrieval method for more secure LLM code generationOriginal title: SEEK: Agentic Knowledge Graph Retrieval for Secure Code GenerationResearchers report that a CWE-structured knowledge graph helps LLMs generate more secure code without retraining.Varun Gadey et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom ScienceOne AIFormula-driven synthesis of physics problems boosts LLM training data qualityOriginal title: TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem SynthesisA preprint presents TIRE, a method building verifiable physics training data from formulas, reporting gains over existing datasets.Added 6 Nov 2026
PreprintNot peer reviewedPivotRL dynamically shifts GPUs across RL post-training stages to boost throughputOriginal title: Efficient Resource-Shiftable RL Post-training with Generative RewardsA preprint presents PivotRL, a system that reallocates GPUs on the fly between rollout, reward and training in RL post-training with generative reward models.Yanfeng Zhao et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom IBM / AmazonA preprint reports table-of-contents retrieval improves document retrieval accuracyOriginal title: STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentationA preprint introduces STAIR, an LLM-based retriever that uses a document's table of contents to improve retrieval accuracy over chunk-based methods.Added 6 Nov 2026
PreprintNot peer reviewedFrom AppleResearchers report frozen LLM critics sharpen credit assignment in agent RLOriginal title: LLM Value Functions for Credit Assignment in Agentic RLA preprint reports that using a frozen LLM as a value critic improves reinforcement learning for multi-step AI agents compared with GRPO and PPO.Added 6 Nov 2026
PreprintNot peer reviewedPreprint finds medical AI fine-tuning changes, not cuts, patient-facing risk languageOriginal title: Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation FrameworkA preprint reports that medically fine-tuned chatbots do not lower overall risk in patient answers, only shift which risky language appears.Savan Doshi · Added 6 Nov 2026
PreprintNot peer reviewedPreprint describes a single-injection attack turning AI agents into DDoS zombiesOriginal title: Mobius Injection: From One-Shot Prompt Injection to Sustained Resource Exhaustion in AI Agent InfrastructureA preprint reports that one hidden prompt injection can trap AI agents in self-sustaining loops, draining shared compute and degrading service for others.Zi Liang et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom AmazonResearchers report a multi-turn jailbreak method that beats prior attacks on protected LLMsOriginal title: ATTCK-R: Knowledge-Aware Long-Horizon Jailbreaking for Protected LLMsA preprint presents ATTCK-R, a planning-based red-teaming framework that the authors say outperforms prior multi-turn jailbreak methods on protected LLMs.Added 6 Nov 2026
