Most LLM-based agents try to predict environment or tool outputs, but those high-entropy, execution-dependent predictions often add little value when true feedback is available. Worse, agents accumulate "task-state contamination": stale assumptions and unsupported plans that linger in history and distort later decisions. This paper flips the world-modeling goal from reconstructing outputs to editing the agent's internal state and future reasoning traces so that subsequent choices are grounded in verifiable progress.
Key Findings
-
Action-oriented editing beats next-observation simulation: AEWM trains an Action Judge to label decisions as Critical, Exploratory, or Noisy, then applies State Revision to rewrite noisy continuations. So what? Agents exposed to these edits make fewer misinformed commitments and recover from bad plans faster.
-
Integration with execution improves outcomes: EditAct ties edits to real executions rather than only issuing critiques, directly changing the state that subsequent reasoning uses. So what? This produces measurable downstream gains—EditAct raises average benchmark scores by about 3.2–6.7 points across three agent backbones.
-
Better supervision and fine-tuning recipes: The paper introduces AEWM-RFT, a rejection-sampling fine-tuning on verified EditAct trajectories that improves over Self-RFT by ~2.2–2.6 points without needing online AEWM guidance. So what? You can obtain most benefits offline through curated, edited trajectories.
-
Robust Action Judge performance: AEWM attains 70.5% macro-F1 on the introduced Action Judge benchmark, outperforming the strongest baseline by 10.6 points. So what? Reliable classification of decision types is central to effective state edits.
Who it's for and trade-offs
Great fit if you develop or evaluate LLM-based agents that rely on long-horizon planning, tool use, web search, or software engineering workflows and find that hallucinated or stale reasoning history harms performance. The approach is particularly relevant when real execution feedback is intermittently available and you prefer editing agent state over simulating noisy outputs.
Look elsewhere if your setting provides perfect, low-latency environment simulators or you cannot alter an agent's decision history at inference time—AEWM's gains depend on the ability to integrate edits into the agent's subsequent reasoning and, in some variants, on curated fine-tuning data.
Where it fits
AEWM sits between simulation-heavy world models and purely critique-based repair: instead of predicting what tools will return, it judges which decisions matter, edits noisy continuations, and commits those edits into the agent's state so planning proceeds from a corrected history. This reframing is especially useful for mitigating "task-state contamination" in long-horizon, multi-tool interactions.