AEWM: Stop Predicting Tool Output, Start Editing the Agent's Bad Ideas
Most world models for LLM agents try to predict what a tool will return. Agent-Editing World Model (arXiv 2609.28416, RUCAIBox) argues that is the wrong job. When the real tool response is one call away, guessing it is close to worthless. The thing actually hurting long-horizon agents is inside their own history: stale plans and unsupported assumptions that keep steering decisions long after they should have been dropped. The authors call it task-state contamination.
So AEWM models how the agent's reasoning and actions shape task progress, and it does two things. Action Judge sorts each decision into Critical, Exploratory or Noisy. State Revision rewrites the noisy continuations, editing the reasoning and action that came from the same observed history. EditAct wires this into real execution, so it changes the state the next decision is made from rather than just leaving a critique the agent might ignore.
Numbers: 70.5 percent macro-F1 on their Action Judge benchmark, 10.6 points above the strongest frontier baseline. Across six benchmarks in search, terminal and software engineering, and three agent backbones, EditAct adds 3.2 to 6.7 points over the strongest baseline. And rejection-sampling fine-tuning on verified EditAct trajectories beats plain self-RFT by 2.2 to 2.6 points with no online AEWM at inference, which means the edits teach something that sticks.
This lands next to yesterday's memory fight between write-time and read-time curation, and quietly adds a third answer. Do not just curate what gets remembered, edit what the agent already believes. Critique-based self-correction has always had the problem that the bad reasoning stays in context and keeps pulling. Deleting and rewriting it is blunter and, on these numbers, better. Code at github.com/RUCAIBox/Agent-Editing-World-Model.
← Back to all articles
So AEWM models how the agent's reasoning and actions shape task progress, and it does two things. Action Judge sorts each decision into Critical, Exploratory or Noisy. State Revision rewrites the noisy continuations, editing the reasoning and action that came from the same observed history. EditAct wires this into real execution, so it changes the state the next decision is made from rather than just leaving a critique the agent might ignore.
Numbers: 70.5 percent macro-F1 on their Action Judge benchmark, 10.6 points above the strongest frontier baseline. Across six benchmarks in search, terminal and software engineering, and three agent backbones, EditAct adds 3.2 to 6.7 points over the strongest baseline. And rejection-sampling fine-tuning on verified EditAct trajectories beats plain self-RFT by 2.2 to 2.6 points with no online AEWM at inference, which means the edits teach something that sticks.
This lands next to yesterday's memory fight between write-time and read-time curation, and quietly adds a third answer. Do not just curate what gets remembered, edit what the agent already believes. Critique-based self-correction has always had the problem that the bad reasoning stays in context and keeps pulling. Deleting and rewriting it is blunter and, on these numbers, better. Code at github.com/RUCAIBox/Agent-Editing-World-Model.
Comments