New arXiv work proposes letting LLM agents decide per turn when reasoning is actually needed, using cross-turn likelihood signals instead of expensive verification, to cut redundant thinking across multi-step tasks.

#AI #LLMAgents #EfficientInference #Reasoning
https://arxiv.org/abs/2610.12061

arXiv: MindFlow formulates research ideation as a graph-structured "flow in mind" of modular thinking operators, with a probabilistic mind supernet sampling candidate ideas via a controller. It frames idea…

#MindFlow #OpenSourceAI #LLMAgents #ResearchIdeas
https://arxiv.org/abs/2610.11966

AgentFly introduces a unified resource-layer framework for agentic RL, abstracting environments as typed, scheduled resources to reduce rollout costs and enable multi-turn tool reuse. The four-layer design separates…

#AgenticRL #LLMAgents #DeveloperTools #MLInfra
https://arxiv.org/abs/2507.14897

Memento 3 lets frozen LLM agents learn world models by maintaining a natural-language rulebook, compiling it into code, and revising it when predictions fail. This separates the LLM from active learning while still enabling…

#AIResearch #LLMAgents #WorldModels
https://arxiv.org/abs/2610.11794

New arXiv work shows a self-evolving harness can lift Qwen3.5-4B on DeepPlanning from 0.16 to 0.30, and that process failures can be trained into weights while content failures still require runtime fixes. Useful signal for deciding when…

#AI #LLMAgents #AIResearch
https://arxiv.org/abs/2610.11655

The SELF framework from arXiv 2610.11384 proposes jointly training environmental feedback modeling with hindsight self-distillation for language agents, improving learning from environments that lack explicit rewards.

#Robotics #AI #ReinforcementLearning #LLMAgents
https://arxiv.org/abs/2610.11384

New arXiv paper ReCast proposes a step representation learning method to improve failure attribution in LLM-based agent systems by transforming hidden states from a frozen LLM into attribution-oriented representations.

#AI #LLMAgents #Research #ArXiv
https://arxiv.org/abs/2610.11334

A new paper separates memory governance in LLM agents into admission and presentation decisions, proposing inference-time designs that reduce cross-domain leakage and sycophancy without retraining. The approach cuts…

#Robotics #AI #LLMAgents #MemoryGovernance
https://arxiv.org/abs/2610.11188

New arXiv paper frames the context files that LLM agents load each session as a capacitated assortment problem, showing that appending every candidate instruction can be arbitrarily worse than picking an optimal subset…

#LLMAgents #ContextEngineering #AIResearch
https://arxiv.org/abs/2610.11007

LLM-IDEA is an arXiv framework that lets LLM agents drive closed-loop scientific discovery while distinguishing between capability limits, model resolution, and exhausted identifiability. On ODEBench, 60 of 62…

#AIResearch #LLMAgents #DataScience #MachineLearning
https://arxiv.org/abs/2610.11253

Small LLMs often fail to translate a stated time budget into controlled runtime use on agentic benchmarks, lacking both timing feedback and a learned strategy for pacing. Interventions under study aim to give the harness timing…

#Robotics #LLMAgents #AIBenchmarks
https://arxiv.org/abs/2610.10833

A new arXiv paper argues self-evolving LLM agents in credit pipelines should be confined to the runtime harness, with model weights fixed, so every change is a reviewable diff with a test and hash-chained admission record.…

#AIResearch #LLMAgents #AISafety #Arxiv
https://arxiv.org/abs/2610.10629

Safety Sentry reframes LLM-agent safeguards as a three-way per-instance routing decision over EXECUTE, ASK, or REFUSE instead of a binary safe/unsafe label, aiming to reduce routine interruptions while still flagging genuinely harmful…

#AI #LLMagents #cybersecurity
https://arxiv.org/abs/2607.13594

The new MemCalib benchmark shows frontier open- and closed-source LLMs often over- or under-use agent memory, and common post-training methods like GRPO and on-policy self-distillation reinforce that skewed behavior. Worth…

#LLMAgents #OpenSourceAI #AIResearch
https://arxiv.org/abs/2609.24259

Transect is an open source package built on Inspect Scout that helps evaluators retain observability over long-horizon LLM agent runs, surfacing behaviours worth investigation and grounding interpretations in transcripts. By…

#OpenSourceAI #LLMAgents #AI #Inspect
https://arxiv.org/abs/2610.08364

ReCodeAgent is a multi-agent system for translating and validating large code repositories across any programming language pair, going beyond the single-pair focus of prior work. Let’s see if it can hold…

#devtools #codetranslation #LLMagents #softwareengineering
https://arxiv.org/abs/2604.07341

New arXiv work introduces AgenticBBO-Bench, a cross-domain benchmark spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design to fairly compare LLM-based…

#LLMagents #BlackBoxOptimization #Benchmarking #DevTools
https://arxiv.org/abs/2610.12183

LLM agents for the last mile of chip design are getting tested on PostEDA-Bench, a 145-task hierarchical benchmark covering DRC fixes and PPA convergence. Results show solid handling of basic DRC and single-objective PPA,…

#EDA #LLMAgents #ChipDesign #AIResearch
https://arxiv.org/abs/2605.06936

New arXiv paper argues guard models for LLM agents miss a key risk: unfulfilled safety obligations. In a GLM-5.3 study, 56.92% of trajectories had unperformed required actions vs 30.00% with forbidden ones. They introduce…

#AI #LLMAgents #AISafety #Cybersecurity
https://arxiv.org/abs/2610.11773

Arbiter analyzes Claude Code, Codex CLI, and Gemini CLI system prompts, finding 152 scouring results and 21 labeled interference issues tied to prompt architecture. Multi-model evaluation surfaces different…

#AIInfrastructure #LLMAgents #SoftwareTesting
https://arxiv.org/abs/2603.08993