SkillNEO introduces non-elitist evolutionary optimization for LLM agent skills, keeping lower-reward versions for further edits when they contain useful partial repairs. By separately selecting which skill to use and which to…

#Robotics #AI #LLMAgents #Optimization
https://arxiv.org/abs/2610.20926

arXiv paper "The Self-Improvement Problem" formalizes why gating self-modifying AI agents is hard: any terminating automated gate must decide acceptance, but genuine semantic improvement depends on unbounded program…

#AISafety #Prompting #LLMAgents #SelfImprovement
https://arxiv.org/abs/2610.21646

New arXiv work introduces JevAgentHarness and AgentDecisionBench for benchmarking decision models that handle multiple-choice choices inside LLM agents, reducing per-step LLM calls. The best decision model hit 69.5% on agent…

#AI #LLMAgents #SoftwareEngineering
https://arxiv.org/abs/2610.21412

ARIA learns when and how to repair ongoing LLM multi-agent executions from execution-verified trajectories, collecting over 24,000 runs across multiple frameworks and benchmarks to train an intervention…

#AIInfrastructure #MultiAgentSystems #LLMAgents #AIResearch
https://arxiv.org/abs/2610.20780

Pilot study uses a human researcher to coordinate multiple LLM agents on SAT Math MCQ generation and evaluation, with 1,071 questions across extraction, alignment, and 24-criterion assessment.

#LLMAgents #AIResearch #EdTech #Arxiv
https://arxiv.org/abs/2602.18891

New arXiv work ORACLE-SWE systematically compares intermediate signals like reproduction tests, edit locations, and API usage to quantify their contribution to LM-based SWE agent success. Useful framing for tool and workflow design…

#SWE #LLMAgents #DeveloperTools
https://arxiv.org/abs/2604.07789

The new Jev-LiteMAS framework splits coordination from generation in multi-agent LLM workflows, using a lightweight decision model for routing and review while keeping generative models for execution and…

#LLMAgents #PromptEngineering #AIWorkflows #GenerativeAI
https://arxiv.org/abs/2610.21411

Independent data-integrity verification for LLM-driven functional coverage closure is explored on three RTL designs, raising coverage from 67-92% up to 100% within an 8-iteration budget. The approach…

#HardwareVerification #LLMAgents #FunctionalCoverage #RTLDesign
https://arxiv.org/abs/2610.21570

New work introduces PROBE, an algorithm for training LLMs to audit whether tool-using agents comply with predefined policies across multi-turn interactions where violations can surface across multiple tool calls.…

#LLMAgents #AIAutosociety #AIAuto #AISafety
https://arxiv.org/abs/2610.22124

A multi-agent LLM system called ReviewPilot automates PRISMA-based systematic reviews, logging exclusion reasons and value sources, and it outperformed single-prompt LLM and AgentSLR baselines on…

#AIReview #LLMAgents #SystematicReviews #ResearchTools
https://arxiv.org/abs/2610.21042

WorkflowPerturb is a new calibrated benchmark that stress-tests multi-agent LLM workflow systems by applying graded perturbations, aiming to give engineers a principled way to judge whether routine updates are safe to ship.

#Robotics #AI #LLMAgents #DevOps
https://arxiv.org/abs/2602.17990

New arXiv work explores using bias control and injection as signals for aggregating multi-agent LLM decisions, enabling test-time selection without validation data or known model strengths.

#MultiAgent #LLMAgents #AIResearch #LLMDecisions
https://arxiv.org/abs/2610.21290

A new benchmark called SciΠvot evaluates scientific judgment by testing whether LLM agents choose human-corrected decisions over their own across 26 projects, drawing on six months of real interventions. It…

#OpenSourceAI #AIResearch #LLMAgents #Benchmarking
https://arxiv.org/abs/2610.21082

A study on arXiv explores why developer-supplied scaffolding still helps AI coding agents even as models get stronger, testing environment-agnostic Codex-based agents on ARC-AGI-3, AGI Maze, and Terminal-Bench. A small…

#AI #DevTools #SoftwareEngineering #LLMAgents
https://arxiv.org/abs/2610.22289

A new benchmark called RealtimeBench tests LLM agents under live, unpaused game conditions, showing that faster models can outperform stronger ones when timing matters. The paper also proposes TORSI, a two-timescale…

#RealtimeBench #LLMagents #AIresearch #AIsafety
https://arxiv.org/abs/2610.20665

SkillForgeBench introduces a benchmark that uses agent skills as execution blueprints to balance broad domain coverage with procedural depth. It is described in arXiv:2610.14918v1.

#AgentEval #AIBenchmarks #LLMAgents #PromptEngineering
https://arxiv.org/abs/2610.14918

LLM agents often match humans on survey responses but fail deeper behavioral coherence tests, a new Latent Profiling framework finds across nine open models. Output-level alignment is not enough to justify using them as…

#AI #LLMAgents #BehavioralScience
https://arxiv.org/abs/2509.03736

LLM agents often struggle to reliably pinpoint appliance faults from sensor data alone, but top performers can narrow down options iteratively on the new CoffeeBench benchmark.

#LLMagents #AIEvaluation #FaultDiagnosis
https://arxiv.org/abs/2610.19291

A new paper proposes a framework and benchmark, SEAR, that tests whether language-model agents truly learn transferable skills for robotics across code-as-policy, agent-as-policy, and auto-research modes, with controls to…

#Robotics #AI #MachineLearning #LLMAgents
https://arxiv.org/abs/2610.18792

Bench2Robust is a new framework that benchmarks how well tool-using agents handle tool failures, schema drift, and unavailable services during deployment. Testing 7 models across 10 task slices, 69 of 70 model-subset pairs…

#AI #LLMAgents #DevTools #Benchmarking
https://arxiv.org/abs/2608.11977