Researchers behind a new paper, the Agent-Editing World Model (AEWM), argue that language world models for LLM agents have been optimizing the wrong target. Existing approaches typically predict environment observations, but the authors contend that reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is already available.
The paper identifies a separate failure mode it calls task-state contamination, in which unsupported assumptions and outdated plans persist in an agent's history and distort later decisions. AEWM instead models how reasoning and actions shape future task progress rather than simulating tool responses.
The approach has two components: an Action Judge that distinguishes Critical, Exploratory, and Noisy decisions, and State Revision that edits noisy reasoning–action continuations from the same observed history. The inference framework, EditAct, integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques.
The authors report training AEWM across Search, Terminal, and Software Engineering via mid-training and supervised fine-tuning. On their Action Judge benchmark, AEWM reaches 70.5% macro-F1, which the paper says exceeds the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct is reported to improve average scores by 3.2–6.7 points over the strongest baseline.
The paper also introduces AEWM-RFT, a rejection sampling fine-tuning method on verified EditAct trajectories, which is reported to improve over Self-RFT by 2.2–2.6 points across three domains without online AEWM guidance.