Agent-Editing World Model edits an agent's noisy reasoning mid-trajectory instead of simulating tool outputs

arXiv / HuggingFace Daily Papers·low signalXBlueskyLinkedInCopy link

arXiv 2609.28416 (23 Sep, Renmin University) argues that predicting high-entropy tool responses adds little when real feedback is available. The real problem it targets is task-state contamination: stale plans and unsupported assumptions that persist in the agent's history. AEWM classifies each decision as Critical, Exploratory or Noisy, then rewrites the noisy continuations, and it reaches 70.5% macro-F1 on its Action Judge benchmark, 10.6 points above the strongest frontier baseline. It is trained across search, terminal and software-engineering tasks, and code is on GitHub (RUCAIBox/Agent-Editing-World-Model).