PreprintNot peer reviewedA preprint reports data mix, not reward tweaks, boosts long-context RLOriginal title: Beyond Reward Engineering: A Data Recipe for Long-Context Reinforcement LearningA preprint finds that curating diverse training data, rather than engineering complex rewards, drives most gains in long-context reinforcement learning.Xiaoyue Xu et al. · Added 6 Nov 2026
PreprintNot peer reviewedA preprint reports a benchmark exposing legal citation hallucinations in LLMsOriginal title: When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination DetectionA new benchmark tests whether LLMs correctly judge if legal claims are backed by the cited court opinion.M. Mikail Demir et al. · Added 6 Nov 2026
PreprintNot peer reviewedA preprint reports an LLM agent that evolves competitive CPU microarchitecture designsOriginal title: Agentic Architect: An Agentic AI Framework for Architecture Design Exploration and OptimizationResearchers describe an agentic framework where an LLM evolves prefetcher, cache replacement and branch-predictor designs under hardware budgets.Vasilis Kypriotis et al. · Added 6 Nov 2026
PreprintNot peer reviewedPreprint proposes continuous trajectory planning for LLM tool-use agentsOriginal title: Tools as Continuous Flow for Evolving Agentic ReasoningA preprint reports that treating tool planning as continuous trajectory generation improves long-horizon success and generalisation over step-wise baselines.Tairan Huang et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom MetaA preprint reports LLMs can cite sources from pretraining without retrievalOriginal title: Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language ModelsResearchers propose Active Indexing, a continual pretraining method letting language models cite their own training data without retrieval, reportedly beating a passive baseline.Added 6 Nov 2026
PreprintNot peer reviewedResearchers propose prompt-based skill method to cut PII sent to cloud LLMsOriginal title: P2Skill: Privacy Preserving Skill Distillation for Cloud-Local LLM Inference SystemsA preprint reports a prompt-only technique that lets a local small model withhold personal details before querying a cloud LLM, without extra training.Myunghoon Ryu et al. · Added 6 Nov 2026
PreprintNot peer reviewedResearchers report a method to stabilise asynchronous on-policy distillation of LLMsOriginal title: $\boldsymbol{f}$-OPD: Stabilizing Long-Horizon On-Policy Distillation with Freshness-Aware ControlA preprint proposes freshness-aware control to fix instability when distilling language models asynchronously, nearing synchronous quality with much of the speed gain.Xianwei Chen et al. · Added 5 Nov 2026
PreprintNot peer reviewedPreprint benchmarks ranking-manipulation attacks on LLM search and recommendation enginesOriginal title: GEO-Bench: Benchmarking Ranking Manipulation in Generative Engine OptimizationA preprint introduces GEO-Bench, testing attacks that manipulate LLM-based rankings under one shared protocol for effectiveness and detectability.Ojas Nimase et al. · Added 5 Nov 2026
PreprintNot peer reviewedA preprint proposes graph-based defence against backdoors in federated LLM fine-tuningOriginal title: FedGraph: Defending Federated Large Language Model Fine-Tuning Against Backdoor Attacks via Graph-Based AggregationResearchers report a graph-clustering defence that spots coordinated malicious updates in federated LLM fine-tuning, cutting attack success below 10%.Xi Chen et al. · Added 5 Nov 2026
PreprintNot peer reviewedResearchers report training LLM agents to call tools via code boosts long-horizon task successOriginal title: ProgAgent: Learning Tool Orchestration through Programmatic TrainingA preprint finds training language models to orchestrate tools through code, rather than direct calls, improves long-horizon task performance.Hao Li et al. · Added 5 Nov 2026
PreprintNot peer reviewedResearchers report a layered on-device cache that speeds up mobile RAG responsesOriginal title: PerCache: Predictive Hierarchical Cache for RAG Applications on Mobile DevicesA preprint describes PerCache, a predictive multi-layer cache designed to cut latency in on-device retrieval-augmented generation for mobile apps.Kaiwei Liu et al. · Added 5 Nov 2026
PreprintNot peer reviewedResearchers report action chunking cuts LLM calls in robot agentsOriginal title: ChunkAct: Action Chunking for Agentic RoboticsA preprint reports that chunking primitives into single LLM calls boosts robot task success while sharply cutting language model usage.Bingxin Xu et al. · Added 5 Nov 2026
PreprintNot peer reviewedA preprint reports an in-stream fix that repairs insecure code as LLMs draft itOriginal title: SecReason: Vulnerability Detection and In-Stream Self-Repair for LLM Code GenerationA preprint describes SecReason, a system that scans code mid-generation and prompts LLMs to repair vulnerabilities before delivery.Ruidong Han et al. · Added 5 Nov 2026
PreprintNot peer reviewedPreprint proposes reusable label knowledge to cut costs in zero-shot text classificationOriginal title: Label-Centric Knowledge Retrieval for Zero-Shot Text ClassificationA preprint describes SLKR, which pre-builds label knowledge with a flagship LLM so a lightweight model can classify text cheaply.Daoping Wu et al. · Added 5 Nov 2026
PreprintNot peer reviewedResearchers report multimodal upgrade to the open Apertus language modelsOriginal title: Apertus 1.5: From Open Text Models to Open Multimodal ReasonersA preprint describes how continued training gave the open Apertus 8B and 70B models image, audio, longer-context and reasoning abilities.Project Apertus: Yixuan Xu et al. · Added 5 Nov 2026
PreprintNot peer reviewedFrom Capital OnePreprint reports adding expert text critiques improves LLM agent trainingOriginal title: Beyond Corrected Actions: Training LLM Agents on Textual CritiquesA preprint finds that fine-tuning LLM agents on expert-written critiques, not just corrected actions, lifts task success.Added 5 Nov 2026
PreprintNot peer reviewedFrom BaiduResearchers propose adaptive sparse attention to speed up long-context LLM decodingOriginal title: Prism Attention: Proposal-Refined Index Sharing Mechanism for Efficient LLMs InferenceA preprint reports a sparse-attention method that adapts index selection across layers to cut long-context decoding costs.Added 5 Nov 2026
PreprintNot peer reviewedFrom AmazonPreprint explains why optimising pass@k can quietly hurt pass@1 accuracyOriginal title: Prompt Interference in LLM Post-Training: Why Pass@k Optimization Can Degrade Pass@1A preprint traces pass@1 drops under pass@k training to a population-level gradient conflict between prompts, not a per-prompt trade-off.Added 5 Nov 2026
PreprintNot peer reviewedPreprint proposes global budget allocation method for compressing LLM KV cachesOriginal title: Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache CompressionA preprint describes GraceKV, a training-free method that globally allocates cache budget across layers and heads to balance detail and coverage in long-context inference.Haolin Tian et al. · Added 5 Nov 2026
PreprintNot peer reviewedFrom Amazon ScienceA preprint reports method for safer use of process rewards in maths reasoning RLOriginal title: Process Rewards for Outcome-Guided Reasoning Steps in LLM Reinforcement LearningA preprint describes PROGRS, a training method that keeps process reward signals from overriding outcome correctness in maths reasoning RL.Added 5 Nov 2026
