AEWM: Direct Agent Correction via Editing World Models

Renmin University proposes AEWM (Agent Editing World Model), a new paradigm that shifts the world model's focus from predicting environment responses to judging and correcting how each decision affects task progress. In long-horizon tasks, errors persist in the interaction history, polluting the task state and misleading subsequent decisions. AEWM addresses this by directly rewriting the current reasoning and action before execution.

EditAct: Judge, Revise, Execute

The EditAct framework operates in three stages:

Judge : An Action Judge classifies the proposed action as critical (advances the task), exploratory (reduces uncertainty), or noisy (detrimental). The judge sees only the history and current proposal — no future information.

Revise : Only for noisy actions, a State Revision module replaces both the current reasoning and the action. Prior history is preserved; tool feedback is never fabricated.

Execute : The real tool runs the original or revised action. The new reasoning, action, and actual feedback are appended to the history for the next step.

An important distinction: actions that do not directly solve the task but prune reasonable branches are labeled exploratory and retained. AEWM does not merely suggest; it rewrites the next step.

Training: Mid-Training and Supervised Fine-Tuning

AEWM is built on Qwen3.5-35B-A3B . It undergoes mid-training on 52.16B tokens across search, terminal, and software engineering domains, followed by supervised fine-tuning on 120,000 high-quality samples (half judgment, half revision). Judgment labels come from retrospective annotation of complete trajectories, but the model input excludes future steps. Revision samples are kept only if actual execution yields measurable task progress.

Experiments: Judgment Accuracy and End-to-End Gains

On a 3,000-decision Action Judge benchmark, AEWM achieves 70.5% macro-F1 , surpassing the strongest baseline by 10.6 percentage points .

Integrating EditAct improves average scores over the strongest baseline by 6.7 (4B), 5.2 (9B), and 3.2 (35B-A3B) points across six benchmarks. For the 35B-A3B agent, BrowseComp rises from 40.9 to 48.1 and Terminal-Bench 2.0 from 37.1 to 48.3 . These are system-level gains from plugging in an independent AEWM; online judgment and revision require extra model calls.

Ablation: Direct Editing Outperforms Prompting and Resampling

Ablation shows that directly editing the action beats two alternatives: (1) asking the agent to resample after a negative judgment, and (2) providing only a reasoning hint. Editing both reasoning and action together also outperforms editing only one component.

A concrete terminal example illustrates the mechanism: the agent fails to see the service port in an IPv4 listing and suspects startup failure. AEWM redirects the check to IPv6, after which a real RPC call verifies functionality. The problem was not that the service failed to start, but that the inspection method missed the information.

Search behavior also shifts: the mean longest consecutive search sequence drops from 16.9 to 13.2 , while distinct pages visited per question rises from 7.7 to 10.1 . Agents perform fewer repetitive keyword searches and read more pages.

AEWM-RFT: Internalizing Correction via Rejection Sampling Fine-Tuning

AEWM-RFT uses EditAct to generate and verify high-quality trajectories, then fine-tunes the base agent per domain. At test time, AEWM is not called. Compared to Self-RFT, scores improve by 2.2–2.6 percentage points across the three domains. Relative to the base agent, average interaction turns decrease by 30.0% on BrowseComp and 16.7% on Terminal-Bench 2.0, though Doc2Repo turns increase — not all tasks become shorter.

Conclusion and Resources

AEWM demonstrates that a world model need not replicate the environment; it can learn how decisions influence task progress and rewrite erroneous steps before they compound. Real tools supply evidence; the model judges and corrects.

Paper: https://arxiv.org/abs/2609.28416
Code: https://github.com/RUCAIBox/Agent-Editing-World-Model
Hugging Face: https://huggingface.co/collections/RUC-AIBOX/agent-editing-world-model