PreprintNot peer reviewedA preprint reports an LLM architecture that reads graph structure nativelyOriginal title: Teaching LLMs to See Graphs: Unifying Text and Structural ReasoningA preprint describes GTLM, which adds graph-aware attention to language models instead of compressing nodes into single tokens.Dario Vajda · Added 25 Oct 2026
PreprintNot peer reviewedStudy finds LLM agents’ numerical signals become structured but behaviour stays unpredictableOriginal title: When Numbers Start Talking: Numerical Signalling and Strategic Behaviour Among LLMsResearchers report that AI agents exchanging numbers develop consistent symbol patterns, yet this does not reliably predict cooperation outcomes across models.Alessio Buscemi et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom Qualcomm, NVIDIA, Microsoft, Google DeepMindMLCommons benchmark measures how easily jailbreaks bypass eight open-weight AI modelsOriginal title: MLCommons Jailbreak Benchmark v1.0A preprint from MLCommons reports a new pipeline for testing single-turn jailbreak attacks and measures how much safety degrades across eight open-weight models.Added 25 Oct 2026
PreprintNot peer reviewedNew benchmark pits humans against LLMs on graph neural network coding tasksOriginal title: GNN-CB: A Graph Neural Network Competition Benchmark for Human and LLM EvaluationA preprint introduces GNN-CB, a competition-style benchmark showing LLMs still trail top human coders on most GNN tasks.Murad Hossen et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom NVIDIAPreprint reports evaluator model lets frozen robot policies pick better actionsOriginal title: Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA ModelsA preprint describes SVA, a method that adds a learned action evaluator to frozen VLA robot policies, reportedly boosting real-robot success rates.Added 25 Oct 2026
PreprintNot peer reviewedResearchers report a context-learning framework doubles AI alignment with human moral judgmentsOriginal title: Morality is Contextual: Learning Interpretable Moral Contexts from Human Data with Probabilistic Clustering and Large Language ModelsA preprint presents COMETH, a modular system that learns context-specific moral judgments and roughly doubles alignment with human ratings versus direct LLM prompting.Geoffroy Morlat et al. · Added 25 Oct 2026
PreprintNot peer reviewedA preprint reports an anonymizer resisting agentic web-search re-identificationOriginal title: LLM Anonymization Against Agentic Re-IdentificationA preprint introduces AURA, a mask-reconstruct method the authors say cuts agentic re-identification while keeping more contextual detail than prior tools.Ziwen Li et al. · Added 25 Oct 2026
PreprintNot peer reviewedA preprint reports LLM agents can re-identify anonymised interviews via web searchOriginal title: AgentDoxx: Agentic Re-identification of Anonymized Text with Web SearchA preprint finds that search-enabled LLM agents can often re-identify anonymised interview transcripts, even after entity masking.Jianing Wen et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom Impossible, Inc.Researchers report hybrid attention cuts video generation latency with similar qualityOriginal title: Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video GenerationA preprint describes Video DeltaNet, a hybrid attention method that the authors say speeds up long video generation while keeping quality close to dense attention.Added 25 Oct 2026
PreprintNot peer reviewedPreprint reports small language models boosting intrusion detection via structured rulesOriginal title: SERA-IDS: Structured Experience Retrieval-Augmented Intrusion Detection with Small Language ModelsResearchers report that adding structured, error-derived rules sharply raises Macro-F1 for small language models detecting network intrusions.Md Shamimul Islam et al. · Added 25 Oct 2026
PreprintNot peer reviewedA preprint describes CacheTune, a system for reusing non-prefix KV caches in long-context LLM servingOriginal title: Adaptive KV Cache Reuse for Fast Long-Context LLM ServingA preprint reports a frequency-guided KV cache reuse system that speeds up long-context LLM serving while preserving output quality.Fei li et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom Crusoe AIResearchers report faster LLM decoding via asymmetric key-value attention designOriginal title: More Value per Key: Asymmetric Sparse Attention for Faster LLM DecodingA preprint describes an attention method that cuts decoding costs by using fewer key heads than value heads under sparsity.Added 25 Oct 2026
PreprintNot peer reviewedResearchers report internal signals can detect and curb unsafe CLI agent actionsOriginal title: Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI AgentsA preprint proposes reading an LLM agent's hidden states step-by-step to catch and steer away from harmful command-line actions.Weidi Luo et al. · Added 25 Oct 2026
PreprintNot peer reviewedResearchers report rhetorical tricks can sway human judges in AI safety debatesOriginal title: AI Safety via Debate is Compromised by Cognitive BiasesA preprint reports that cognitive-bias tactics like jargon and verbosity shifted human judgments in AI debate experiments, even when the manipulated side was wrong.Gefei Liu et al. · Added 25 Oct 2026
PreprintNot peer reviewedResearchers report a black-box attack method for testing LLM jailbreak defencesOriginal title: All you need to break LLMs are Black-Box, Adapting, Efficient, Transferable, Harmful, Applicable … AttacksA preprint presents a diffusion-model-based attacker the authors say works across undefended, defended and API-hosted LLMs without model access.Vincent Limbach et al. · Added 25 Oct 2026
PreprintNot peer reviewedResearchers report LLMs often get vulnerability verdicts right for flawed reasonsOriginal title: Correct Verdicts, Flawed Reasoning: Structured Auditing of LLM-based Vulnerability ReasoningA preprint finds correct vulnerability verdicts from LLMs often hide fabricated reasoning that free-form judges miss.Boyue Caroline Hu et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom Citigroup, Ernst & Young, NVIDIAA preprint proposes a framework to test enterprise AI workflows before scaling themOriginal title: EnterpriseVal: Quantifying the Efficacy, Reliability and Value of Generative AI in the EnterpriseA preprint presents EnterpriseVal, a system for assessing whether generative AI workflows are fit, reliable and worth scaling inside a firm.Added 25 Oct 2026
PreprintNot peer reviewedA preprint finds small LLMs cave fast to persuasion, confidence prompts backfireOriginal title: Vulnerability of LLMs' Stated Beliefs? LLMs Belief Resistance Check Through Strategic Persuasive Conversation InterventionsA preprint reports that LLM belief stability under persuasion varies widely by model, and that asking for confidence scores makes some models more, not less, vulnerable.Fan Huang et al. · Added 25 Oct 2026
PreprintNot peer reviewedPreprint predicts LLM code maintainability before generationOriginal title: PreMaQ: Predicting Maintainability-Related Quality of LLM-Generated Code Before GenerationA preprint reports a method that predicts code smell and maintainability scores from an LLM's internal state before code is generated.Taiki Wakamatsu et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom BaiduResearchers propose entropy-based filtering to stabilise asynchronous LLM reinforcement learningOriginal title: Deconstructing Off-Policy Ratios: Entropy-Normalized Trust Regions for Asynchronous Reinforcement LearningA preprint proposes scaling the trust region for off-policy updates by token entropy to stabilise asynchronous RL training.Added 25 Oct 2026
