AEWM: editing agent task states beats simulating tool responses

RUC · hf · 2026-09-25

Existing language world models predict environment observations, but reconstructing high-entropy, execution-dependent tool responses adds little value when real feedback exists — and agents suffer from task-state contamination, where unsupported assumptions and outdated plans distort later decisions.

RUC's Agent-Editing World Model (AEWM) instead models how reasoning and actions shape future task progress. Key components: an Action Judge distinguishing Critical/Exploratory/Noisy decisions, State Revision to edit noisy reasoning-action continuations, and EditAct which integrates these with real execution to directly change the state underlying subsequent decisions.

Trained across Search, Terminal, and Software Engineering: 70.5% macro-F1 on their Action Judge benchmark (+10.6 over the strongest frontier baseline); EditAct improves average scores by 3.2–6.7 points across six benchmarks and three agent backbones; rejection-sampling fine-tuning (AEWM-RFT) adds another 2.2–2.6 points over Self-RFT.

Original post →

More from coding & agent

coding & agent channel →