Agent-Editing World Model: Rethinking World Modeling for LLM Agents
AEWM edits the agent’s task state instead of trying to predict tool outputs.
The paper argues that simulating high-entropy tool responses is less useful when real execution feedback is already available. Its Agent-Editing World Model classifies decisions as critical, exploratory, or noisy, then revises noisy reasoning-action continuations from the same observed history. Integrated as EditAct, it changes the state used for later decisions rather than just critiquing them. The authors report gains across Search, Terminal, and Software Engineering benchmarks, including a 70.5% macro-F1 Action Judge result and 3.2–6.7 point average score improvements over the strongest baseline. HF Daily Papers' note
The paper argues that simulating high-entropy tool responses is less useful when real execution feedback is already available. Its Agent-Editing World Model classifies decisions as critical, exploratory, or noisy, then revises noisy reasoning-action continuations from the same observed history. Integrated as EditAct, it changes the state used for later decisions rather than just critiquing them. The authors report gains across Search, Terminal, and Software Engineering benchmarks, including a 70.5% macro-F1 Action Judge result and 3.2–6.7 point average score improvements over the strongest baseline. HF Daily Papers' note
score 5