AEWM: A World Model That Edits an Agent's Bad Decisions
Renmin University researchers built a world model that judges an agent's steps and rewrites noisy ones instead of predicting tool output. It improved scores by 3.2 to 6.7 points across six benchmarks.
TL;DR
- Renmin University of China researchers propose the Agent-Editing World Model (AEWM), which judges an agent's decisions and edits noisy ones instead of predicting environment output [^1].
- Paired with real execution, it improved average scores by 3.2 to 6.7 points over the strongest baseline across six benchmarks and three agent models [^1].
Background
World models for language agents usually predict environment observations, such as the output of a tool call [^1]. The authors argue that reconstructing such unpredictable tool responses has limited value when real feedback is available anyway. Agents also suffer from what they call task-state contamination: unsupported assumptions and outdated plans linger in the history and distort later decisions [^1].
What happened
AEWM models how an agent's reasoning and actions shape future task progress [^1]. It combines an Action Judge, which classifies decisions as critical, exploratory or noisy, with State Revision, which edits noisy reasoning-and-action continuations from the same history [^1]. A system called EditAct combines both with real execution, changing the state that later decisions are based on rather than only offering critiques [^1]. The team trained AEWM on search, terminal and software engineering tasks [^1].
AEWM reached 70.5 percent macro-F1 on the authors' Action Judge benchmark, 10.6 points above the strongest frontier model baseline [^1]. EditAct improved average scores by 3.2 to 6.7 points across six benchmarks and three agent backbones [^1]. Fine-tuning agents on verified EditAct trajectories, a variant the authors call AEWM-RFT, improved over self-training by 2.2 to 2.6 points without needing AEWM at run time [^1].
Why it matters
Long-running agents fail less from single bad steps than from bad steps that stay in their history and shape everything afterward. A component that recognizes which steps were noise and rewrites them attacks that problem directly. The fact that the gains carried over to agents trained on the cleaned trajectories, without the world model running, suggests the approach can improve agents themselves rather than just supervise them [^1].
Practical example
Say a coding agent assumes early on that a bug is in the database layer, which is wrong, and keeps reading database files. In an EditAct-style setup, the judge would flag that step as noisy and the revision step would rewrite it, so later decisions are no longer anchored to the wrong assumption [^1].
Related gear
We recommend this classic text because it provides the foundational understanding of structural logic required to build agents that operate reliably within complex system architectures.