PreprintNot peer reviewedPreprint describes sleeper-style attacks that hide in AI agent memory and skillsOriginal title: Plant, Persist, Trigger: Sleeper Attack on Large Language Model AgentsA preprint reports that adversarial content can lie dormant in an LLM agent's state and later trigger harm via an unrelated, benign request.Yongxiang Li et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom Relty TechnologyPreprint proposes token-selection method to linearise LLM attention more efficientlyOriginal title: STILL: Selecting Tokens for Intra-Layer Hybrid Attention to Linearize LLMsA preprint describes STILL, a method that picks important tokens for full attention and summarises the rest linearly, aiming to cut compute costs.Added 18 Oct 2026
PreprintNot peer reviewedFrom SnowflakeA preprint describes EviLLM, an attack that makes AI coding tools quietly inject vulnerabilitiesOriginal title: Beware EviLLM: Enabling Vulnerability Injection via Large Language ModelsA preprint reports a third-party attack, EviLLM, that manipulates ChatGPT-based code generation to inject vulnerabilities most study participants failed to notice.Added 18 Oct 2026
PreprintNot peer reviewedFrom MicrosoftPreprint finds sliding-window attention often beats retrofitted linear attention in LLMsOriginal title: Sliding-window beats linear attentionA preprint reports that simple sliding-window attention with sinks matches or outperforms retrofitted linear attention, especially on long-context tasks.Added 18 Oct 2026
PreprintNot peer reviewedFrom MartianSteering vectors can cut self-preference bias in LLM judges, preprint reportsOriginal title: Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLM EvaluatorsA preprint reports that activation steering vectors reduce unjustified self-preference in LLM evaluators far more than prompting or DPO.Added 18 Oct 2026
PreprintNot peer reviewedFrom Microsoft ResearchResearchers propose training safety routers on data isolated from harmful contentOriginal title: Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety RoutingA preprint describes SafeMoE, a system that isolates harmful knowledge into expert modules and routes around it to give safer, more informative answers.Added 18 Oct 2026
PreprintNot peer reviewedA preprint reports a training method for more calibrated LLM-based numeric predictionsOriginal title: Reinforcement Learning over Predictive Distributions for LLM RegressionA preprint introduces a reinforcement learning reward that trains LLMs to produce better-calibrated predictive distributions alongside more accurate point estimates.Jungsoo Park et al. · Added 18 Oct 2026
PreprintNot peer reviewedFrom NTTResearchers report prompt injection can fool LLM-based phishing detectorsOriginal title: Clouding the Mirror: Stealthy Prompt Injection Attacks Targeting LLM-based Phishing DetectionA preprint finds that hidden instructions in phishing sites can mislead multimodal LLM phishing detectors, and tests a defence to reduce this.Added 18 Oct 2026
PreprintNot peer reviewedResearchers report gains from delaying KV cache eviction until first decode stepOriginal title: DeferKV: Rethinking Eviction Timing for One-Shot KV Cache CompressionA preprint proposes waiting until the first decoding step before trimming the KV cache, reporting consistent gains over prompt-only compression methods.Zhe Wang et al. · Added 18 Oct 2026
PreprintNot peer reviewedA preprint proposes a theory for when external guidance helps LLM reasoning trainingOriginal title: When Does External Guidance Help LLM Reasoning? A Bias-Variance Theory of Guidance-Augmented GRPOA preprint offers a bias-variance theory and optimal weighting rule for guidance-augmented reinforcement learning on reasoning tasks.Sofia Torres et al. · Added 18 Oct 2026 · Released Jul 2026
PreprintNot peer reviewedA preprint describes a cost-aware agent system for proving code correctness in Lean 4Original title: Cost-Efficient Theorem Proving via Agent Orchestration in Program VerificationResearchers report a theorem-proving agent that balances solve rate against token cost across Lean 4 verification benchmarks.Shuangjie Yao et al. · Added 18 Oct 2026
PreprintNot peer reviewedA preprint benchmark tests whether LLMs judge microcontroller electrical safety correctlyOriginal title: ElecSafety: Diagnosing Large Language Model Safety Judgments for Microcontroller BoardsA preprint introduces ElecSafety, a benchmark finding that LLMs often give correct safety labels for the wrong reasons.Linjian Yang et al. · Added 18 Oct 2026
PreprintNot peer reviewedStudy questions how hallucination benchmarks label LLM answers as wrongOriginal title: The Labeling Problem in Hallucination Detection Benchmarks: An Empirical EvaluationResearchers find that automated labelling methods for hallucination detection often disagree with human judgement and with each other.Jorma Valjakka et al. · Added 18 Oct 2026
PreprintNot peer reviewedPreprint describes tile-aligned mixed-precision quantization for faster LLM inferenceOriginal title: AlignQuant: Tile-Aligned Mixed-Precision Quantization for Efficient LLM GenerationA preprint reports a quantization method that aligns precision choices with GPU tiles to speed up LLM generation while preserving quality.Hanzhi Zhang et al. · Added 8 Oct 2026
PreprintNot peer reviewedA preprint reports LLMs show stronger illusory pattern perception than humansOriginal title: Illusory Pattern Perception Drives Spurious Inference in Large Language ModelsA preprint finds large language models tend to infer false patterns from random data more readily than human participants do.Peihua Mai et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom HuaweiA preprint reports LLMs often answer even when given context is insufficientOriginal title: Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive PerspectiveA preprint introduces a paired-question benchmark finding most LLMs fabricate answers rather than abstaining when context is insufficient.Added 8 Oct 2026
PreprintNot peer reviewedResearchers propose communication-efficient federated reasoning for medical LLMsOriginal title: FedCoT: Communication-Efficient Federated Reasoning Enhancement for Large Language ModelsA preprint describes FedCoT, a federated learning framework that improves LLM reasoning on medical QA while cutting communication overhead and keeping data local.Chuan Li et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom summary.aiA preprint proposes a semantic-aware distillation method for shrinking large language modelsOriginal title: WASD: Wasserstein-based Knowledge Distillation for Large Language ModelsResearchers report that aligning teacher and student models by token meaning, not just probability, improves knowledge distillation across several LLM families.Added 8 Oct 2026
PreprintNot peer reviewedFrom Microsoft ResearchPreprint proposes benchmark and two world-model approaches for language-guided navigationOriginal title: Language-Conditioned World Modeling for Visual NavigationA preprint introduces a large instruction-following navigation benchmark and compares two world-model paradigms for guiding robots by language alone.Added 8 Oct 2026
PreprintNot peer reviewedFrom XiaohongshuA preprint proposes unbiased reward models trained from implicit LLM feedback signalsOriginal title: Unbiased Reward Modeling from Implicit Feedback for LLM AlignmentA preprint describes ImplicitRM, a method for building reward models from passive signals like clicks and copies rather than explicit ratings.Added 8 Oct 2026
