PreprintNot peer reviewedResearchers report an open-weight AI agent that improves formal maths proof translationOriginal title: AIProver: Agentic Auto-Formalization of Mathematical Research via Certificate-Driven Evolving HarnessA preprint describes AIProver, an agentic system that jointly trains a model and evolves its tool harness to auto-formalise research-level maths proofs in Lean.Prithwish Jana et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom GoogleResearchers report insight injection lifts agentic ML-engineering results by 123.9%Original title: Assembling Insights for Agentic Machine Learning Engineering SystemsA preprint finds that feeding assembled task insights into an agentic machine-learning engineering system sharply boosts its benchmark scores.Added 25 Oct 2026
PreprintNot peer reviewedPreprint finds tool-hallucination fixes can backfire when tool setups changeOriginal title: Understanding and Mitigating Hallucination Escape in Tool-Using LLM AgentsA preprint reports that existing LLM tool-hallucination fixes can increase errors under other tool configurations, and proposes a method to curb this.Peigui Qi et al. · Added 25 Oct 2026
PreprintNot peer reviewedFrom MetaA preprint proposes finer credit assignment to boost LLM reasoning via RLOriginal title: Hierarchical Credit Assignment for RLVR on Fused Gromov-Wasserstein GeometryResearchers report a new method that rewards semantically novel reasoning steps in LLM training, improving results on several math benchmarks.Added 25 Oct 2026
PreprintNot peer reviewedPreprint models why self-distillation resists forgetting while fine-tuning does notOriginal title: Learning without Overwriting: A Theory of Self-Distillation and Supervised Fine-Tuning in Continual ReasoningA preprint offers a theoretical account of why on-policy self-distillation preserves earlier skills better than standard fine-tuning during continual reasoning training.Shinichi Uemura et al. · Added 25 Oct 2026
PreprintNot peer reviewedResearchers report hardware counters can reveal if data trained an LLMOriginal title: TranScope: What the Software Hides About LLM Training Data, the Hardware Reveals at Scale, and Accelerators MagnifyA preprint describes TranScope, a tool using CPU performance counters to detect whether specific data was used to train a language model.Joshua Kalyanapu et al. · Added 24 Oct 2026
PreprintNot peer reviewedResearchers report code comments can steer LLMs to write vulnerable codeOriginal title: Hidden in the Comments: A Context-Injection Attack Surface in Code LLMsA preprint finds that insecure instructions hidden in code comments push Code LLMs toward vulnerable completions far more often than benign tasks.Noor Munir et al. · Added 24 Oct 2026
PreprintNot peer reviewedResearchers report vision models can boost weaker audio and video MLLMs via mergingOriginal title: Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMsA preprint reports that merging a well-aligned vision model into audio or video multi-modal LLMs improves their target-modality performance without extra training.Hoigi Seo et al. · Added 24 Oct 2026
PreprintNot peer reviewedA preprint reports a jailbreak defence that extracts intent from LLM outputsOriginal title: Reactivating Alignment: Defending LLMs from Jailbreaks via Intention-Aware Input-Output MatchingA preprint describes SENTINEL, a generation-time defence that matches LLM input and output text to spot and block jailbreak attempts.Luoyu Chen et al. · Added 24 Oct 2026
PreprintNot peer reviewedA preprint reframes LLM agent design using deep learning concepts like layers and NASOriginal title: Building LLM Agent Systems the Deep Learning Way: From Modular Design to Architecture SearchA preprint proposes building LLM agent systems as modular architectures, borrowing ideas like backpropagation and neural architecture search.Tao Feng et al. · Added 24 Oct 2026
PreprintNot peer reviewedFrom Dell TechnologiesResearchers propose analytical model for choosing LLM inference parallelism strategiesOriginal title: Characterizing Parallelism Strategies in LLM Inference: Fundamental Compute-Communication Trade-offsA preprint presents a compute-communication model explaining when tensor versus pipeline parallelism works best for LLM inference.Added 24 Oct 2026
PreprintNot peer reviewedPreprint proposes training LLMs to internalise a safe persona, not just refusal phrasesOriginal title: Beyond Refusal Patterns: Safe-Role Internalization for Robust and Generalizable LLM Safety AlignmentResearchers report a fine-tuning method that teaches models safety-oriented behaviour rather than refusal patterns, improving robustness and reducing over-refusal.Jinghao Pang et al. · Added 24 Oct 2026
PreprintNot peer reviewedFrom SGIT AI LabPreprint reports difficulty-aware evolution strategy improves LLM fine-tuning over standard ESOriginal title: DART-ES: Difficulty-Aware Reweighting and Targeted Replay for Fine-Tuning LLMs with Evolution StrategiesA preprint describes DART-ES, a method that reweights training problems by difficulty to improve evolution-strategy fine-tuning of language models.Added 24 Oct 2026
PreprintNot peer reviewedFrom Alibaba GroupA preprint reports FP4 rollout training method for MoE language models matching BF16 performanceOriginal title: TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language ModelsResearchers propose TRACE, an FP4 quantization scheme for RL training of MoE language models, reporting BF16-level performance with faster rollout.Added 24 Oct 2026
PreprintNot peer reviewedFramework aligns federated mixture-of-experts fine-tuning across heterogeneous mobile devicesOriginal title: Federated Mixture-of-Experts Alignment on Mobile Edge Networks under Data HeterogeneityA preprint proposes FedAlign-MoE, which aligns gating and expert behaviour across devices to improve federated LLM fine-tuning on non-IID data.Zihan Fang et al. · Added 24 Oct 2026
PreprintNot peer reviewedFrom NEC Labs AmericaPreprint proposes workflow-affinity embeddings to tailor agentic LLM workflows per queryOriginal title: RA-MoWE: Workflow-Affinity Embeddings for Query Clustering and Agentic Workflow GenerationA preprint reports that clustering queries by how they respond to reference workflows can boost task scores while cutting LLM calls.Added 24 Oct 2026
PreprintNot peer reviewedPreprint proposes lightweight controllers to cut multi-agent LLM coordination costsOriginal title: Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of LaborA preprint reports that routing coordination decisions to small models instead of large LLMs cuts token use and latency while keeping accuracy high.Zihan Zhou et al. · Added 24 Oct 2026
PreprintNot peer reviewedResearchers link LLMs’ behaviour in cooperation games to multi-agent science task performanceOriginal title: Cooperative Profiles Predict Multi-Agent LLM Team Performance in AI for Science WorkflowsA preprint reports that how LLMs behave in behavioural economics games predicts their performance as AI-for-Science collaborators.Shivani Kumar et al. · Added 24 Oct 2026
PreprintNot peer reviewedFrom NVIDIAA preprint reports frozen medical AI models can improve via external expertise, not retrainingOriginal title: Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AIResearchers report a framework letting frozen language and vision-language models improve in medicine by storing skills, facts and images rather than retraining.Added 24 Oct 2026
PreprintNot peer reviewedA preprint reports multi-agent framework with causal graph boosts vulnerability detection accuracyOriginal title: CLEAR: Causal Context-Based Agentic Reasoning for Vulnerability DetectionResearchers propose CLEAR, a multi-agent system using a causal knowledge graph to better distinguish vulnerable from benign code.Sungju Yun et al. · Added 24 Oct 2026
