Agent-Editing World Model: Rethinking World Modeling for LLM Agents

AEWM edits noisy LLM-agent reasoning instead of predicting tool responses: its Action Judge hits 70.5% macro-F1 (+10.6 over the strongest frontier baseline) and EditAct adds 3.2-6.7 points across six benchmarks and three agent backbones.

The Agent-Editing World Model (AEWM) reframes world modeling for long-horizon LLM agents: rather than reconstructing high-entropy, execution-dependent tool responses, it edits noisy reasoning and actions and models task progress. An Action Judge classifies decisions as Critical, Exploratory, or Noisy, scoring 70.5% macro-F1, 10.6 points above the strongest frontier baseline, while State Revision rewrites contaminated continuations from the same history. Trained on Search, Terminal, and Software Engineering, the EditAct method improves average scores by 3.2-6.7 points across six benchmarks and three agent backbones, and AEWM rejection-sampling fine-tuning (AEWM-RFT) adds 2.2-2.6 points over Self-RFT without online guidance. The two sources, the Hugging Face daily papers note (2026-09-22) and the arXiv listing (2026-09-23), agree on all figures; the arXiv report adds no conflicting details.

• Action Judge reaches 70.5% macro-F1, 10.6 points above the strongest frontier baseline.
• Action Judge classifies decisions as Critical, Exploratory, or Noisy; State Revision rewrites contaminated continuations from the same history.
• AEWM is trained on Search, Terminal, and Software Engineering environments.
• EditAct improves average scores by 3.2-6.7 points across six benchmarks and three agent backbones.
• AEWM-RFT beats Self-RFT by 2.2-2.6 points without online guidance.
• AEWM models task progress instead of reconstructing high-entropy tool responses.
• Both sources (Hugging Face, 2026-09-22; arXiv, 2026-09-23) report identical figures with no disagreements.