PreprintNot peer reviewedFrom PleIAsResearchers release Common Corpus, a two trillion token open dataset for LLM trainingOriginal title: Common Corpus: The Largest Collection of Ethical Data for LLM Pre-TrainingA preprint presents Common Corpus, a large openly licensed dataset for training language models without copyright concerns.Added 8 Oct 2026
PreprintNot peer reviewedResearchers propose SEER framework to shorten AI chain-of-thought in coding tasksOriginal title: Reasoning Efficiently Through Adaptive Chain-of-Thought Compression: A Self-Optimizing FrameworkA preprint reports that trimming AI reasoning length can cut truncation and loops while keeping coding accuracy.Kerui Huang et al. · Added 8 Oct 2026 · Released Mar 2026
PreprintNot peer reviewedResearchers report diffusion language models beat autoregressive ones on constraint-heavy puzzlesOriginal title: Blackboard Intelligence Can Surpass Autoregressive on Globally Constrained ProblemsA preprint finds that letting a diffusion language model revise a whole solution canvas, guided by its own confidence, outperforms token-by-token generation on constrained puzzles.Woosang Jeon et al. · Added 8 Oct 2026
PreprintNot peer reviewedA preprint proposes hash retrieval method to speed up long-context LLM inferenceOriginal title: HHR: Hierarchical Hash Retrieval for Efficient LLM GenerationResearchers report a two-stage hashing approach that improves retrieval accuracy and speeds up long-context LLM decoding.Lianjun Liu et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom NVIDIAResearchers report self-reflection training cuts prompt-injection success in LLM agentsOriginal title: Self-Reflection Fine-Tuning: Enhancing Agent Security against Prompt Injection Attacks from Failure ExperienceA preprint describes a fine-tuning method that teaches LLM agents to learn from failed attacks, sharply lowering injection success rates.Added 8 Oct 2026
PreprintNot peer reviewedFrom Fidelity InvestmentsA preprint reports a method to stop agent tasks from corrupting shared memoryOriginal title: Safe Context Switching for Agents in the Wild: Mitigating Subspace Interference via Orthogonal AdaptationResearchers propose AURA, a geometric method to stop sequential agent tasks from contaminating each other's residual activations.Added 8 Oct 2026
PreprintNot peer reviewedResearchers report visual tool-use harness boosts multi-image reasoning in MLLMsOriginal title: Rethinking Multi-Image Re-Representation in Multi-Image UnderstandingA preprint reports that letting models build visual intermediates, not just text, helps most on tasks needing precise visual evidence.Gengyuan Zhang et al. · Added 8 Oct 2026
PreprintNot peer reviewedPreprint reports a training-free sentence check for unsupported RAG contentOriginal title: Sentence-Level Context Sensitivity as a Training-Free Detector of Unsupported Content, Evaluated Against Trained VerifiersA preprint evaluates a cost-free likelihood-contrast signal for spotting unsupported sentences in RAG outputs against trained verifiers.Mohamed Aly Bouke · Added 8 Oct 2026
PreprintNot peer reviewedPreprint describes small model trained to manage its own long-context retrievalOriginal title: Trained Agentic Context ManagementA preprint reports that a fine-tuned small model with a tiny context window can match a frontier model's long-document performance.Bryce Sandlund · Added 8 Oct 2026
PreprintNot peer reviewedA preprint links weight averaging in LLM training to quantization robustnessOriginal title: Understanding the Weight Averaging Mechanism in LLM Training for Post-Training QuantizationA preprint proposes a theory explaining why averaging recent training weights helps post-training quantization, and tests it on LLMs.Hanzhang Wang et al. · Added 8 Oct 2026
PreprintNot peer reviewedA preprint proposes TACO, a sparse optimizer cutting LLM fine-tuning memoryOriginal title: TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-TuningResearchers describe a new optimizer that sharply reduces memory needed for full-parameter LLM fine-tuning while keeping accuracy competitive.Jichao Jiang et al. · Added 8 Oct 2026
Journal articleReview status unknownSurvey maps a gap between AI agents’ network access and the safety checks around themOriginal title: Large Language Models for Agentic NetOps and AIOps: Architectures, Evaluation, and SafetyA review of LLM-driven network and IT-operations agents finds assurance mechanisms lag behind the autonomy these systems are given.Muhammad Bilal et al. · Added 8 Oct 2026
PreprintNot peer reviewedPreprint finds LLM environment simulators collapse on multi-field state updatesOriginal title: EnvSimBench: A Benchmark for Evaluating and Improving LLM-Based Environment SimulationA preprint introduces EnvSimBench, finding that language models simulating environments fail sharply once several state variables must change at once.Yi Liu et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom Actava AIA preprint reports a healthcare model trained by recursive self-improvementOriginal title: Cura 1T: Healthcare Foundation Model via Recursive Self-ImprovementResearchers describe Cura 1T, a healthcare model refined through repeated self-evaluation cycles, reporting gains across clinical benchmarks without losing general skills.Added 8 Oct 2026
PreprintNot peer reviewedResearchers propose pivot-based reward method for training search agentsOriginal title: PiCA: Pivot-Based Credit Assignment for Search Agentic Reinforcement LearningA preprint introduces PiCA, a step-reward method for search agents that reportedly beats strong baselines on seven QA benchmarks.Dongyi Liu et al. · Added 8 Oct 2026
PreprintNot peer reviewedPreprint describes table-based pipeline that boosts LLM long-context reasoningOriginal title: $π^2$: Structure-Originated Reasoning Data Improves Long-Context Reasoning Ability of Large Language ModelsA preprint presents a QA pipeline using Wikipedia tables that reportedly improves long-context reasoning accuracy in fine-tuned LLMs.Quyet V. Do et al. · Added 8 Oct 2026
PreprintNot peer reviewedA preprint proposes a causal metric to check whether LLM unlearning really erases knowledgeOriginal title: Measuring the Depth of LLM Unlearning via Activation PatchingResearchers introduce the Unlearning Depth Score, which uses activation patching to test whether supposedly unlearned knowledge still lurks inside a model.Jaeung Lee et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom NVIDIAPreprint finds agentic retrieval beats standard search but at high compute costOriginal title: Beyond Semantic Similarity: Performance and Costs of Agentic Retrieval for Complex TasksA preprint reports that combining LLM reasoning with dense retrieval lifts search accuracy but takes far more time and tokens.Added 8 Oct 2026
PreprintNot peer reviewedResearchers propose structured multimodal memory for medical AI agentsOriginal title: MSM-Mem: A Universal Medical Structured Multimodal Memory Framework for Medical AI AgentsA preprint describes MSM-Mem, a memory framework letting medical AI agents learn from accumulated clinical cases over time.Md Asaduzzaman Jabin et al. · Added 8 Oct 2026
PreprintNot peer reviewedFrom Ant GroupResearchers split LoRA adapters to personalise federated multimodal modelsOriginal title: PFAdapter: Hierarchical LoRA Decomposition for Personalized Federated MLLMsA preprint proposes splitting adapter layers so federated multimodal models share some parameters but keep others private per device.Added 8 Oct 2026
