PreprintNot peer reviewedResearchers propose attack-chain approach to catch email prompt injection attemptsOriginal title: Prompt Injection Detection for Email Agents Through Attack Chain ModelingA preprint models indirect prompt injection in email agents as a multi-stage attack chain rather than a single text classification task.Ahmad Hashmi et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom IntuitResearchers report offline fine-tuning method that treats feedback scores as goalsOriginal title: Goal-Conditioned Supervised Learning for LLM Fine-TuningA preprint proposes training LLMs via supervised learning to exceed quality thresholds, reportedly beating standard offline fine-tuning methods.Added 1 Nov 2026
PreprintNot peer reviewedFrom EnhansResearchers report multi-agent prompt optimizer outperforms prior methods on HelpSteer benchmarksOriginal title: MA-SAPO: Multi-Agent Reasoning for Score-Aware Prompt OptimizationA preprint describes MA-SAPO, a multi-agent system that turns evaluation scores into reusable reasoning assets to guide prompt edits.Added 1 Nov 2026
PreprintNot peer reviewedPreprint proposes routing questions to specialist medical LLMs for better QA accuracyOriginal title: MedRouter: Demystifying Knowledge Differences Across Medical LLMs for Routing-Based ReasoningA preprint reports that routing medical questions to specialist LLMs and combining their answers outperforms picking a single best model.Lang Cao et al. · Added 1 Nov 2026
PreprintNot peer reviewedPreprint proposes combining prompt evolution with RL to fix stuck reward signals in LLM reasoningOriginal title: P^2O: Joint Policy and Prompt OptimizationA preprint reports a method joining prompt evolution and policy updates to tackle vanishing reward signals on hard reasoning problems.Xinyu Lu et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom NVIDIAResearchers report a tuning-free method to stretch LLM context windows without retrainingOriginal title: Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPEA preprint describes Jet-Long, a zero-shot context-extension technique that reportedly outperforms existing baselines while keeping most of FlashAttention-3's speed.Added 1 Nov 2026
PreprintNot peer reviewedPreprint reports narrow fine-tuning can misalign vision-language models broadlyOriginal title: Narrow Multimodal Fine-Tuning Can Induce Emergent MisalignmentA preprint reports that narrow multimodal fine-tuning can produce broad, unrelated misaligned behaviour across vision-language models.Shunchang Liu et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom ForSight RoboticsResearchers report using vision-language models to prioritise RL replay buffer experiencesOriginal title: VLM-Guided Experience ReplayA preprint reports that scoring replay buffer clips with a frozen vision-language model boosts reinforcement learning success rates and sample efficiency.Added 1 Nov 2026
PreprintNot peer reviewedA preprint proposes a fairness-aware training method for multi-step AI agentsOriginal title: Agentic Multi-Turn Reasoning: A Fairness ApproachResearchers describe Phi-MPO, a preference-learning method aimed at improving long multi-turn AI agent reasoning and reducing data imbalance effects.Thanh-Dat Truong et al. · Added 1 Nov 2026
PreprintNot peer reviewedPreprint explores compact per-turn memory instead of full chat history for LLM agentsOriginal title: Continuous Context ManagementA preprint reports that updating a short memory every turn cuts context size but affects task success differently across models.William Hoy et al. · Added 1 Nov 2026
PreprintNot peer reviewedA preprint surveys how RAG and reasoning methods curb LLM hallucinationsOriginal title: Mitigating Hallucination in Large Language Models: A Capability-Oriented Survey on RAG, Reasoning, and Agentic SystemsA preprint proposes a framework linking hallucination types to retrieval, reasoning and agentic mitigation methods.Yihan Li et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom ByteDanceA preprint reports a likelihood-free RL method for diffusion language modelsOriginal title: LFPO: Likelihood-Free Policy Optimization for Masked Diffusion ModelsResearchers describe LFPO, a training approach for diffusion language models that avoids estimating intractable sequence likelihoods during reinforcement learning.Added 1 Nov 2026
PreprintNot peer reviewedA preprint proposes a diffusion-based reward model for LLM alignmentOriginal title: Diffusion Reward ModelsResearchers report a diffusion reward head that models reward distributions instead of a single score, matching or beating baselines on five benchmarks.Xiangyang Wang et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom Qualcomm AI ResearchPreprint proposes looped transformer design that keeps memory constant as reasoning depth growsOriginal title: Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language ModelsA preprint describes MELT, a looped transformer that shares one KV cache across reasoning loops to avoid linear memory growth.Added 1 Nov 2026
PreprintNot peer reviewedFrom Alibaba GroupResearchers report faster on-device LLM responses via KV cache reuse on mobile NPUsOriginal title: Dynamic Flow, Static Graph: KV Cache Reuse for Efficient LLM Serving on Mobile NPUsA preprint describes a compute-storage system that reuses KV cache on mobile NPUs to cut first-token latency.Added 1 Nov 2026
PreprintNot peer reviewedReview maps environmental costs of generative and agentic AI, authors reportOriginal title: Environmental Impact of Generative and Agentic AI: An in-Depth Analysis and Green SolutionsA preprint reviews energy, carbon, water and e-waste impacts across the AI lifecycle, flagging agentic workflows as a major multiplier.Abderaouf Bahi et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom NAVER Cloud, Toss, A2SYSResearchers propose shared prefill module to cut redundancy in multi-model LLM servingOriginal title: PrefillShare: A Shared Prefill Module for KV Reuse in Multi-LLM Disaggregated ServingA preprint describes PrefillShare, which lets multiple fine-tuned LLMs reuse one prefill computation and KV cache, cutting latency and boosting throughput.Added 1 Nov 2026
PreprintNot peer reviewedPreprint measures energy costs of 25 open LLMs for code completionOriginal title: Green AI: Cost of LLM-Based Code CompletionA preprint reports that energy use in LLM code completion depends on task type, with smaller quantized models often matching larger ones efficiently.Negar Alizadeh et al. · Added 1 Nov 2026
PreprintNot peer reviewedResearchers propose unlearning for agent action trajectories, not just factsOriginal title: Trajectory Unlearning on LLM-based AgentsA preprint introduces trajectory-level unlearning for LLM agents and a method called GiRPO that forgets bad action sequences while keeping task performance.Yingdan Shi et al. · Added 1 Nov 2026
PreprintNot peer reviewedResearchers propose Theory of Scene to fix coordination failures in homogeneous LLM agent teamsOriginal title: Theory of Scene: Breaking the Symmetry Trap in Multi-Agent LLM CoordinationA preprint reports a training-free reasoning method that helps identical LLM agents split or combine tasks without talking to each other.Liangqi Yuan et al. · Added 1 Nov 2026
