PreprintNot peer reviewedResearchers show open-weight LLM safeguards can be jailbroken without fine-tuningOriginal title: Open-Weight LLM Fine-Tuning Defenses are Susceptible to Simple AttacksA preprint finds that simple, no-training attacks bypass fine-tuning defenses meant to stop misuse of open-weight language models.Kevin Kuo et al. · Added 24 Oct 2026
PreprintNot peer reviewedFrom logsight.aiResearchers report dual-AI framework for robot control via natural languageOriginal title: Adaptive Code Generation for Controlling RobotsA preprint describes an LLM-and-VLM architecture letting robots turn open-ended language goals into verified, adaptable code.Added 24 Oct 2026
PreprintNot peer reviewedFrom Hewlett Packard EnterprisePreprint proposes probing LLM internals, not just benchmark scores, to spot hidden flawsOriginal title: Latent Performance Profiling of Large Language ModelsA preprint presents a framework that measures hidden activations in language models to reveal traits benchmarks miss.Added 23 Oct 2026
PreprintNot peer reviewedResearchers propose lifetime-aware KV cache placement to extend flash memory life in agent servingOriginal title: Lachesis: Lifetime-Aware KV Cache Placement for Agent Serving across HBM and High-Bandwidth FlashA preprint reports a placement layer that sorts AI agent memory by lifetime, extending flash memory's usable life over a simple baseline.Jaehoon Yang et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom HuaweiPreprint explores how intermediate retrieval rewards improve training of search agentsOriginal title: Beyond Outcome Rewards: Constructing and Assigning Retrieval Credit for Search AgentsA preprint reports that rewarding intermediate retrieval steps, not just final answers, improves search-agent training across seven QA benchmarks.Added 23 Oct 2026
PreprintNot peer reviewedA preprint links LLM reasoning to pattern recognition via a shared graph structureOriginal title: The Dichotomy Between Pattern Recognition and Step-by-Step ReasoningResearchers propose that step-by-step reasoning and pattern recognition in LLMs sit on one spectrum, governed by how much context matters.Amrut Nadgir et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom MetaResearchers report LLM agents designing new neural network architectures autonomouslyOriginal title: Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-DesignA preprint describes LLM agents that design and test new foundation-model architectures, some beating human baselines.Added 23 Oct 2026
PreprintNot peer reviewedStudy audits privacy risks introduced by LLM-enhanced graph neural networksOriginal title: Auditing Privacy Risks in LLM-Enhanced Graph Neural NetworksA preprint reports that enriching GNN node features with LLM-generated semantics raises exposure to link, label and membership inference attacks.Longzhu He et al. · Added 23 Oct 2026
PreprintNot peer reviewedA preprint finds LLM vulnerability-patching benchmarks easily inflated by evaluation designOriginal title: On the Reliability of LLM-Based Vulnerability Patching BenchmarksA preprint reports that small changes to how patching benchmarks are run can sharply shift headline pass rates without improving real fix quality.Dang K Le et al. · Added 23 Oct 2026
PreprintNot peer reviewedMulti-agent framework with rubric RL reports strong gains on end-to-end GPU code generationOriginal title: StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement LearningA preprint describes StitchCUDA, a three-agent GPU programming system trained with rubric-based reinforcement learning to improve end-to-end code generation.Shiyang Li et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom NEC Labs AmericaA preprint reports per-query multi-agent system design beats fixed workflowsOriginal title: OOPMAS: Object-Oriented Multi-Agent Systems for Query-Level Workflow GenerationA preprint describes OOPMAS, which builds a custom agent team and workflow for each query rather than reusing one fixed design.Added 23 Oct 2026
PreprintNot peer reviewedPreprint proposes DynaCore, a reshaped chip design for faster LLM servingOriginal title: A Shape-Adaptive Architecture with Disaggregated Quantization for Efficient LLM ServingA preprint describes a chip architecture that adapts to both the prefill and decoding phases of LLM inference to cut serving delays.Cong Guo et al. · Added 23 Oct 2026
PreprintNot peer reviewedResearchers report security risks hidden in efficiency-focused speculative decoding for LLMsOriginal title: Secure Speculative Decoding for Large Language ModelsA preprint finds that faster, lossy speculative decoding methods sharply raise jailbreak and prompt-injection risks while barely affecting utility scores.Yichi Zhang et al. · Added 23 Oct 2026
PreprintNot peer reviewedSurvey maps mismatches between RAG attacks and defencesOriginal title: A Survey of Secure Retrieval-Augmented GenerationA preprint survey of 135 studies finds RAG security attacks and defences are unevenly matched across targets, stages and signals.Yuming Xu et al. · Added 23 Oct 2026
PreprintNot peer reviewedStudy maps how language models recover after acting on false premisesOriginal title: PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMsA preprint introduces PHRBench, finding that models rarely correct hallucinated premises, but recovery can be predicted beforehand.Linghao Meng et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom Pathfinder Applied ResearchResearchers propose topology-based signals to detect and steer ill-posed LLM promptsOriginal title: TIPS: Topological Ill-Posedness Probing and Steering in Large Language ModelsA preprint reports that shape-based features of internal model states can flag ambiguous questions and guide LLMs to clarify or abstain.Added 23 Oct 2026
PreprintNot peer reviewedPreprint proposes skill lifecycle so LLM agent skill libraries don’t go stale during RLOriginal title: SkillForge: Co-Evolving Skills and Agents via Dynamic Skill LifecyclesA preprint describes SkillForge, which retires and mutates outdated agent skills during reinforcement learning rather than letting libraries grow unchecked.Yuyao Ge et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom XiaohongshuResearchers propose graph-based credit scoring to fix step selection in LLM agent distillationOriginal title: GraphOPD: Graph-Augmented On-Policy Distillation for LLM AgentsA preprint reports that routing teacher guidance to steps with high structural impact, not just high disagreement, improves agent training on several benchmarks.Added 23 Oct 2026
PreprintNot peer reviewedResearchers report decoding method cuts LLM hallucinations without retrainingOriginal title: APCD: Adaptive Path-Contrastive Decoding for Reliable Large Language Model GenerationA preprint describes a multi-path decoding technique that the authors say improves factual accuracy across several language models.Tianyu Zheng et al. · Added 23 Oct 2026
PreprintNot peer reviewedFrom WaldiezA preprint reports routing small models per task matches hosted LLMs in a home-automation systemOriginal title: Not Every Call Needs a Frontier Model: Per-Call-Site Evaluation of Small Language Models in a Deployed Agentic Home-Automation SystemA preprint finds that assigning different small language models to each type of call in a home-automation agent nearly matches hosted models at far lower cost.Added 23 Oct 2026
