PreprintNot peer reviewedFrom Capital OneA preprint reports an adaptive guardrail model that reasons only on tricky safety casesOriginal title: AdaGuard: Enhancing Safety and Policy Compliance with Reasoning-Enabled LLM-As-A-Judge GuardrailsResearchers describe AdaGuard, a guardrail model that learns to decide per-policy when to reason, cutting latency while keeping accuracy.Added 23 Oct 2026
PreprintNot peer reviewedFrom AppleResearchers report a model can improve its own safety alignment without external supervisionOriginal title: SIGMA: Self-Improving Alignment Generalization from a Model SpecA preprint describes SIGMA, a pipeline letting a model generate and judge its own alignment training tasks from a Model Spec.Added 22 Oct 2026
PreprintNot peer reviewedA preprint reports a learned controller that picks agent configurations per queryOriginal title: Learning to Configure Agentic AI SystemsResearchers describe ARC, a system that learns to pick workflows, tools and budgets for each query rather than using fixed templates.Aditya Taparia et al. · Added 22 Oct 2026
PreprintNot peer reviewedResearchers propose token-level routing to combine LLM reasoning with precise computationOriginal title: PiERN: Token-Level Routing for Integrating High-Precision Computation and ReasoningA preprint describes PiERN, an architecture that routes individual tokens between language reasoning and isolated high-precision computation experts.Jingyuan Fan et al. · Added 22 Oct 2026
PreprintNot peer reviewedResearchers report LLM-agent design search for edge AI chip mapping beats standard optimisersOriginal title: Agentic Design Space Exploration for Joint Hardware Configuration Selection and Mapping of AI Inference Workloads on Heterogeneous Edge SoCsA preprint reports an LLM-driven proposer-critic system that uses execution traces to speed up hardware design search for edge AI chips.Geetha Prasuna Yarramneni et al. · Added 22 Oct 2026
PreprintNot peer reviewedAPEX guards LLM agents by checking actions right before they happenOriginal title: APEX: Active Protection at Execution Boundaries for LLM AgentsA preprint describes APEX, a defence that checks AI agent actions at the point of execution rather than scanning for attack patterns.Xinran Zheng et al. · Added 22 Oct 2026
PreprintNot peer reviewedPreprint proposes patching model weights to offset KV cache compression lossesOriginal title: PatchKV: Weight-Space Compensation of KV CacheA preprint describes PatchKV, a training-free method that patches LLM weights to recover context lost from aggressive KV cache compression.Chanryeol Lee et al. · Added 22 Oct 2026
PreprintNot peer reviewedResearchers report alignment training raises confident but wrong LLM answersOriginal title: The Alignment Paradox: How Post-Training Amplifies Confident Hallucinations in Language ModelsA preprint argues that post-training alignment, not missing data, is the main cause of confidently wrong answers in language models.Qingjia Huang et al. · Added 22 Oct 2026
PreprintNot peer reviewedFrom BYDReserved chat-template tokens drive much of prompt-injection success, preprint findsOriginal title: Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt InjectionA preprint reports that reserved control tokens, not just template text, account for most of chat-template prompt-injection success in LLM agents.Added 22 Oct 2026
PreprintNot peer reviewedPreprint merges policy and value learning in one model for agent RL trainingOriginal title: SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement LearningA preprint reports a single-model, single-rollout method that matches PPO's performance while cutting memory and runtime.Dayang Liang et al. · Added 22 Oct 2026
PreprintNot peer reviewedFrom Meta, LightOnResearchers report steering how much AI assistants trust user claimsOriginal title: TrustMI: Causally controlling how assistants trust their usersA preprint reports that adjusting internal activations can make language model assistants more or less willing to rely on unverifiable user claims.Added 22 Oct 2026
PreprintNot peer reviewedStudy finds LLMs struggle to spot deliberately disguised malicious smart contractsOriginal title: The Deception Delta: Adversarial Evaluation of LLM-Based Smart Contract Bytecode ForensicsA preprint reports that adversarial camouflage in smart contract bytecode sharply cuts large language models' ability to detect hidden drains.Timo Schefold · Added 22 Oct 2026
PreprintNot peer reviewedResearchers propose Patcher, a defence against full-parameter malicious finetuning of open LLMsOriginal title: Defending Against Malicious Finetuning by Scaling Train-time Adversarial AttacksA preprint presents Patcher, an adversarial training method that reuses simulated attack vectors to defend LLMs against malicious finetuning.Haoming Wen et al. · Added 22 Oct 2026
PreprintNot peer reviewedFrom MicrosoftResearchers report RL framework boosts LLM creativity in scientific proposal writingOriginal title: Reinforcing Agentic Creativity in Scientific Ideation with Night ScienceA preprint reports that reinforcement learning taught language models to deviate from predictable reasoning, producing more varied and original research proposals.Added 22 Oct 2026
PreprintNot peer reviewedResearchers report runtime that shares cloud compute across edge-cloud LLM splitsOriginal title: DySCo: Dynamic Sharding for Collaborative Edge-Cloud LLM Inference with Depth-Synchronized BatchingA preprint describes DySCo, a runtime that lets edge devices share partial LLM inference with the cloud more efficiently via depth-synchronized batching.Jingpo Xu et al. · Added 22 Oct 2026
PreprintNot peer reviewedPreprint describes an AI agent that tunes RAG pipelines by diagnosing failuresOriginal title: Agentic AutoRAG: RAG Pipeline Optimization through Reasoning-Driven AgentsResearchers report an LLM-agent optimizer that diagnoses retrieval versus generation failures to tune RAG pipelines more cheaply and accurately.Lasse B. Strand et al. · Added 22 Oct 2026
PreprintNot peer reviewedFrom ZTEA preprint describes G2PTQ, a refreshed gradient-and-Hessian method for compressing LLMsOriginal title: G$^2$PTQ: Improving LLM Post-Training Quantization with Generalized Gradient CompensationResearchers propose a post-training quantization method that updates gradient and Hessian estimates per block to reduce LLM compression errors.Added 22 Oct 2026
PreprintNot peer reviewedFrom SK hynixResearchers release trace dataset and simulator to study complex agentic AI systemsOriginal title: Characterizing How Complex Agentic AI Systems Handle General Tasks: A Trace-Based Simulation StudyA preprint introduces a token-level trace dataset and simulator for studying how multi-agent AI systems behave on general tasks.Added 22 Oct 2026
PreprintNot peer reviewedA preprint reports agents self-onboarded onto tasks cut costs sharply, but with variable resultsOriginal title: AutoOnboarder: Autonomously Onboarding LLM AgentsA preprint describes AutoOnboarder, a system that retrains and reconfigures agents for a task, reporting large cost cuts but variable outcomes.Tanay Arora et al. · Added 19 Oct 2026
PreprintNot peer reviewedPreprint reports self-improving agent framework that adapts its own upgrade workflowsOriginal title: RSIFlow: Adaptive Self-Improvement Workflows with an Evolving Meta-AgentA preprint describes RSIFlow, a framework where a meta-agent diagnoses an AI agent's weak points and evolves its own improvement process.Huan Zhang et al. · Added 19 Oct 2026
