PreprintNot peer reviewedFrom Ant GroupA preprint proposes letting frozen LLMs learn orchestration skills without retrainingOriginal title: Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent SystemsResearchers report a method that lets frozen large language models improve multi-agent orchestration by refining a written 'Meta-Skill' rather than updating model weights.Added 6 Nov 2026
PreprintNot peer reviewedA preprint maps which quantization choices make LLMs practical to prove with zero-knowledge proofsOriginal title: Bits Under ZK-LLM: Evaluating Zero-Knowledge-Friendly Quantization for Verifiable Private LLM InferenceResearchers report that activation precision and lookup-table design, not just bit-width, drive utility loss and proving cost in zero-knowledge LLM inference.Taeung Yoon et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom NVIDIANVIDIA reports open MoE hybrid model aimed at long agentic tasksOriginal title: Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic ReasoningNVIDIA describes Nemotron 3 Ultra, a large mixture-of-experts hybrid model the authors say boosts inference speed while matching rival accuracy.Added 6 Nov 2026 · Released Jun 2026
PreprintNot peer reviewedPreprint finds multi-turn chat badly degrades retrieval-augmented generation accuracyOriginal title: Lost in Conversation or Lost in Translation? Diagnosing Multi-Turn Degradation in RAGA preprint reports that spreading questions across conversation turns makes RAG systems both less accurate and less reliable.Pranav Handa et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom Anyscale / SnowflakeResearchers report a CPU offload method that cuts GPU idle time in LLM fine-tuningOriginal title: ZenFlow: Enabling Stall-Free Offloading for LLM TrainingA preprint describes ZenFlow, which prioritises important gradients to reduce GPU stalls during offloaded LLM training.Added 6 Nov 2026
PreprintNot peer reviewedPreprint proposes a universal action layer to help AI agents handle messy tool ecosystemsOriginal title: AnyAct: Universal Action for Self-Evolving AgentsA preprint reports a unified action layer, AnyAct, that filters and ranks tools dynamically to help agents cope with huge, changing tool ecosystems.Lingrui Xu et al. · Added 6 Nov 2026
PreprintNot peer reviewedResearchers test LLMs’ social intelligence using real Reddit persuasion debatesOriginal title: PerSo: Investigating Facets of LLMs' Social Intelligence in Real-World Persuasion ScenariosA preprint introduces PerSo, a benchmark using real ChangeMyView debates to test whether LLMs can read minds and pick persuasive arguments.Ali Zamani et al. · Added 6 Nov 2026
PreprintNot peer reviewedResearchers report back-translation boosts LLM translation for low-resource languagesOriginal title: Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMsA preprint reports that back-translating monolingual text improves LLM translation into low-resource languages more than extra pretraining or parallel data.Divyanshu Bathla et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom Maro Systems / AMDResearchers propose frozen-backbone method to curb forgetting in continual LLM fine-tuningOriginal title: CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual LearningA preprint describes CRAFT, which adapts language models to new tasks via representation edits rather than weight changes, aiming to reduce forgetting.Added 6 Nov 2026
PreprintNot peer reviewedResearchers report a training method to curb context-driven factual errors in LLM reasoningOriginal title: Reasoning Error from Known Fact: Step-Level Self-Consistency Group Relative Policy Optimization for LLMsA preprint presents SSC-GRPO, which uses self-consistency across rollouts to reward correct reasoning steps and reduce hallucinations.Xiaomeng Hu et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom FactiverseStudy finds five LLM factuality alignment methods fail to beat base modelOriginal title: Do Factuality Methods Make LLMs More Factual? Reproducibility and Fair ComparisonA preprint reports that none of five tested factuality alignment methods for LLMs reliably beat the untrained base model on a shared benchmark.Added 6 Nov 2026
PreprintNot peer reviewedResearchers report a mixed-precision method compressing LLM weights to 1.6 bits eachOriginal title: Squeeze10-LLM: Squeezing LLMs' Weights by 10 Times via a Staged Mixed-Precision Quantization MethodA preprint describes Squeeze10-LLM, a staged quantization method claimed to compress LLM weights tenfold with reduced accuracy loss.Qingcheng Zhu et al. · Added 6 Nov 2026
PreprintNot peer reviewedResearchers report internal model signals can detect training data in fine-tuned LLMsOriginal title: inLoRA-MINT: Internal Representations for Auditing Training Data in Domain-adapted LLMsA preprint reports a method using internal model activations to audit whether specific text was used to fine-tune LLMs with LoRA.Gonzalo Mancera et al. · Added 6 Nov 2026
PreprintNot peer reviewedPreprint proposes meta-cognitive LLM framework for budget-aware reasoning across problemsOriginal title: ROI-Reasoning: Rational Optimization for Inference via Pre-Computation Meta-CognitionA preprint presents ROI-Reasoning, a two-stage method letting language models judge problem difficulty and allocate a shared token budget across tasks.Muyang Zhao et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom Microsoft ResearchPreprint proposes bias control and injection to improve LLM multi-agent answer aggregationOriginal title: Bias Control and Injection for Multi-Agent AggregationA preprint describes a training-free framework that normalises reasoning style and then stress-tests answers to pick better final outputs from multiple LLM agents.Added 6 Nov 2026
PreprintNot peer reviewedA preprint reports an LLM taught to pick and interpret EHR data for ED riskOriginal title: Incentivizing Tool-Use Reinforcement Learning for Post-Arrival Emergency Care PredictionResearchers report a 4B-parameter LLM trained with tool-use reinforcement learning matches larger models at predicting emergency department outcomes.Pengfei Hu et al. · Added 6 Nov 2026
PreprintNot peer reviewedOracle-based process rewards boost agentic LLM training across three puzzle gamesOriginal title: Verifiable Process Rewards for Agentic ReasoningA preprint reports that turning task-solving oracles into dense, action-level rewards improves agentic reasoning training over outcome-only methods.Huining Yuan et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom BaiduResearchers report a lightweight framework for agentic reinforcement learning training of LLMsOriginal title: A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated RecipesA preprint describes LoongSage, an open-source agentic RL framework designed to cut overhead and keep training and inference in sync.Added 6 Nov 2026
PreprintNot peer reviewedResearchers report training LLMs on real user behaviour logs boosts user-task performanceOriginal title: Vela: User Behavior as Supervision Sparks LLM IntelligenceA preprint describes Vela, a framework that turns anonymised user behaviour data into training tasks, reportedly improving both user-related and general LLM performance.Yang Xiao et al. · Added 6 Nov 2026
PreprintNot peer reviewedFrom DialpadResearchers report GPT-5.4 leads new benchmark for LLM customer-support skill executionOriginal title: Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support SkillsA preprint evaluates 17 LLMs as customer-support workflow executors, finding judge disagreement is a significant factor in results.Added 6 Nov 2026
