Agent-Editing World Model: Rethinking World Modeling for LLM Agents
The paper argues LLM agents need editable task state more than simulated tool output.
AEWM is built to identify whether an agent’s decisions are critical, exploratory, or noisy, then revise bad reasoning-action continuations from the same observed history. Its EditAct setup uses real execution feedback while changing the state that guides later decisions. The authors report 70.5% macro-F1 on their Action Judge benchmark, 10.6 points above the strongest frontier baseline. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2 to 6.7 points. ArXiv · AI/CL/LG's note
AEWM is built to identify whether an agent’s decisions are critical, exploratory, or noisy, then revise bad reasoning-action continuations from the same observed history. Its EditAct setup uses real execution feedback while changing the state that guides later decisions. The authors report 70.5% macro-F1 on their Action Judge benchmark, 10.6 points above the strongest frontier baseline. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2 to 6.7 points. ArXiv · AI/CL/LG's note
score 5