PreprintNot peer reviewedPreprint proposes method to infer an LLM’s pretraining data mix from its output textOriginal title: LLMSurgeon: Diagnosing Data Mixture of Large Language ModelsA preprint describes LLMSurgeon, which estimates a language model's pretraining domain mixture using only generated text.Yaxin Luo et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom EvenUpA preprint reports an agent that lets small models handle long document QAOriginal title: DocPilot: Interactive Agent Orchestration and Training for Long Visual Document UnderstandingThe authors describe DocPilot, a trainable document agent that pairs adaptive evidence-seeking with curriculum training to help compact models rival larger VLMs.Added 3 Nov 2026
PreprintNot peer reviewedResearchers report adversarial self-attack improves AI-based vulnerability patchingOriginal title: P2Patch: Exploit-Guided Vulnerability RepairA preprint describes P2Patch, which has an AI exploiter attack AI-written patches, finding new bypasses in 41% of runs.Mohsen Iranmanesh et al. · Added 3 Nov 2026
PreprintNot peer reviewedA preprint finds multimodal AI models rarely admit when none of the options fitOriginal title: Measuring Epistemic Humility in Multimodal Large Language ModelsA new benchmark finds that even top multimodal models seldom pick 'none of the above' when it is the correct answer.Bingkui Tong et al. · Added 3 Nov 2026
PreprintNot peer reviewedPreprint reframes a reported AI agent ‘escape’ as a reproducible access-control bugOriginal title: From Training to Escalation: Forensic Analysis and Local Recreation of RL/RLHF-Induced Privilege Escalation in AI Agents (CVE-2026-65616)A preprint argues that a widely reported AI agent breakout was really a reproducible token-forgery flaw in misconfigured repository servers.Alicia Colmenero-Fernández · Added 3 Nov 2026
PreprintNot peer reviewedFramework tags LLM reasoning steps to test strategy-reuse and RL mode-seeking claimsOriginal title: SHAPE of Thinking: Mathematical Heuristics for Analyzing and Improving LLM ReasoningA preprint introduces SHAPE, a heuristic-tagging method for chain-of-thought, and uses it to probe reasoning shifts and RL training behaviour.Jonghyun Song et al. · Added 3 Nov 2026
PreprintNot peer reviewedResearchers report local LLM apps leak prompts despite on-device inferenceOriginal title: The Illusion of Local Privacy: Confidentiality Boundary Failures in Consumer LLM Serving SystemsA preprint reports that running LLMs locally does not by itself keep prompts confidential, citing memory, storage and server-isolation gaps.Youssef Hamdi Zafan Ibrahim et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom Yuanhui AIA preprint reports LLMs often judge their own outputs inconsistentlyOriginal title: VALCON: Benchmarking and Enhancing Value Consistency in Large Language Models via Dual-Role VariationA preprint finds that large language models frequently fail to judge their own generated answers the way they produced them.Added 3 Nov 2026
PreprintNot peer reviewedFrom Li AutoResearchers report a unified vision-language model for embodied and agent tasks, tuned for edge chipsOriginal title: ME-VLM: A Unified VLM for Embodied Cognition and Agent CoordinationA preprint describes ME-VLM, a vision-language model merging embodied perception and agent skills, with edge-optimised deployment on Li Auto's M100 chip.Added 3 Nov 2026
PreprintNot peer reviewedFrom Oracle AI, Turing EnterpriseNew benchmark exposes gaps in how AI models find and read document evidenceOriginal title: VELD-QA: A Large-Scale Dataset for Visual Evidence Grounding and Question Answering in Long DocumentsA preprint introduces a large human-authored benchmark testing whether vision-language models can locate and interpret visual evidence in long documents.Added 3 Nov 2026
PreprintNot peer reviewedFrom GoogleResearchers report a cheap decision model can take over routine agent choicesOriginal title: JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMsA preprint finds a small decision model, trained on an agent's own states, can safely replace many LLM calls to cut cost and latency.Added 3 Nov 2026
PreprintNot peer reviewedPreprint benchmark finds VLMs struggle to justify actions in first-person videoOriginal title: NoRA: Evaluating Grounded Reasonableness in Visual First-person Normative Action ReasoningA preprint introduces NoRA, a benchmark testing whether vision-language models can propose and justify reasonable next actions from first-person video.Sichao Li et al. · Added 3 Nov 2026
PreprintNot peer reviewedFrom Kuaishou TechnologyPreprint proposes low-cost way to decide when LLM agents should use learned skillsOriginal title: SkillScaffold: Scaffolding Agentic Reinforcement Learning with Policy-Adaptive Skill Credit Assignment and Selective InjectionA preprint introduces SkillScaffold, which estimates skill usefulness cheaply and only injects skills during training when they help.Added 3 Nov 2026
PreprintNot peer reviewedA preprint reports Agent-Reflex trains language agents to recover from noisy tool feedbackOriginal title: Agent-Reflex: Teaching Language Agents to Act ReflectivelyA preprint describes Agent-Reflex, a training framework that teaches language agents to reflect and recover when tools fail or users intervene.Guanting Dong et al. · Added 2 Nov 2026
PreprintNot peer reviewedA preprint benchmarks when fast ‘System-1’ models can replace reasoning LLMsOriginal title: System1-Arena: Which Decisions Can Be Left to System 1?A new benchmark finds fast, single-pass decision models can match reasoning LLMs on many tasks but lag on forecasting and long-horizon agent planning.Tianyu Chen et al. · Added 2 Nov 2026
PreprintNot peer reviewedFrom Telefonica Innovacion Digital, LogitechResearchers report a benchmark for ‘procedural hallucination’ in chat agentsOriginal title: DialFork: Going Beyond Factuality to Evaluate Procedural Hallucination in Conversational AgentsA preprint introduces DialFork, a benchmark testing whether conversational agents stay on process rather than just staying truthful.Added 2 Nov 2026
PreprintNot peer reviewedA preprint reports gradient geometry shifts during RLVR, motivating a rank-adaptive optimizerOriginal title: Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable RewardsResearchers propose Ares, an optimizer that adapts its update rank to the shrinking gradient subspace seen during RLVR training of language models.Zhipeng Chen et al. · Added 2 Nov 2026
PreprintNot peer reviewedResearchers report a debate protocol that hunts down hallucinating AI agents via image testsOriginal title: Multi-agent Undercover Gaming: Hallucination Removal via Counterfactual Test for Multimodal ReasoningA preprint describes MUG, a multi-agent debate protocol that spots hallucinating agents using modified images, reporting gains on reasoning benchmarks.Dayong Liang et al. · Added 2 Nov 2026
PreprintNot peer reviewedResearchers report prompt-security results depend heavily on how they are measuredOriginal title: SoK: Systematizing LLM Prompt Security: Taxonomies, Datasets, and Unified Evaluation of Attacks and DefensesA preprint argues that jailbreak attack and defence results for large language models shift sharply with evaluation configuration.Hanbin Hong et al. · Added 1 Nov 2026
PreprintNot peer reviewedFrom OPPOResearchers propose MARFT, a reinforcement fine-tuning framework for multi-agent LLM systemsOriginal title: MARFT: Multi-Agent Reinforcement Fine-TuningA preprint presents MARFT, a framework for fine-tuning teams of LLM agents with reinforcement learning, reporting gains over a baseline method.Added 1 Nov 2026
