AEWM: World Model That Judges & Corrects Agent Decisions Pre-Execution

Introduction

Researchers at Renmin University's Gaoling School of Artificial Intelligence have open-sourced a new world model for LLM agents called Agent-Editing World Model (AEWM) . The paper, titled "Agent-Editing World Model: Rethinking World Modeling for LLM Agents," addresses a core failure mode in long-horizon agents: task-state contamination , where unverified assumptions, outdated plans, and erroneous progress judgments persist in the interaction history and continuously mislead subsequent reasoning and actions.

Problem: Task-State Contamination

Agents often receive real tool feedback but still persist in wrong directions. The authors observe that each step may look reasonable locally, yet the overall trajectory drifts further from the goal. The key issue is not whether feedback is obtained, but how the agent interprets that feedback and decides what to do next.

AEWM: From Environment Prediction to Decision Judgment

Existing language world models for agents typically predict environment observations (search results, terminal output, test results). However, such observations are highly uncertain and context-dependent. Since agents can call real tools to get reliable feedback, spending model capacity on simulating the environment is not the most valuable use.

AEWM changes the modeling target: instead of reconstructing "what the tool will return," it predicts "what role the current decision will play in advancing the task." A single AEWM model learns two capabilities:

Action Judge: Classifying Decision Impact

Before an action is executed, the model predicts which of three categories the candidate reasoning+action belongs to:

Critical : Key advances — obtaining necessary evidence, resolving critical blockers, or effecting required state changes.

Exploratory : Valuable exploration — testing reasonable hypotheses, eliminating candidate branches, or reducing task-relevant uncertainty.

Noisy : Ineffective actions — repeating known information, deviating from task requirements, or continuing along unsupported directions.

Importantly, not immediately solving the problem does not mean the step lacks value . A troubleshooting step that reliably rules out a reasonable branch is still Exploratory and worth keeping. AEWM aims to reduce ineffective repetition, not necessary exploration.

State Revision: Rewriting the Next Step

If a candidate decision is judged Noisy, AEWM generates an alternative reasoning+action from the same observed history: re-examining faulty assumptions, handling contradictory evidence, or re-incorporating missed requirements. The revised step directly becomes the next execution point and enters the subsequent history.

EditAct Framework: Judge, Revise, Then Execute

To integrate AEWM into agent execution, the authors propose EditAct . For regular tool interactions, the flow becomes:

Agent proposes decision → AEWM judges → Retain Critical/Exploratory or revise Noisy → Real execution → Update interaction history

Critical and Exploratory decisions are kept; only Noisy decisions trigger revision. The adopted reasoning, action, and real tool feedback all enter the history for subsequent steps.

Case Study: gRPC Service IPv4 vs IPv6

In a paper example, an agent must implement and run a gRPC service. Logs show startup success, but the target port is absent from the IPv4 listening list. The agent continues troubleshooting assuming the service isn't listening properly.

AEWM notices the code actually binds to an IPv6 address while the agent checked IPv4. It revises the next step to check IPv6 listeners; real execution confirms the port is listening. A further revision shifts repeated port checks to actual RPC read/write verification, and the task passes all 7 tests.

The revision's value is not having the world model declare "service is normal" out of thin air, but correcting the agent's understanding of existing evidence and choosing an action that verifies that judgment .

Training Methodology

Using Qwen3.5-35B-A3B as the base, the team trained a unified world model across Search, Terminal, and Software Engineering (SWE) domains. Training data centered on two questions:

What contribution did this step make to the task? For Action Judge, an annotation model combined real tool feedback and subsequent trajectory to label each decision's contribution type. Future information was used only for offline labeling, not as input to AEWM's pre-execution judgment; reasoning traces were filtered to avoid leaking information not yet visible at decision time.

Is the revised decision actually more useful? For State Revision, a revision model proposed alternatives from the same history, then executed the revised action. Only samples that gained real execution feedback, subsequent trajectory support, and substantial task progress were retained.

AEWM underwent 52.16B tokens of mid-training followed by supervised fine-tuning on 120k curated samples (60k Action Judge, 60k State Revision).

Experimental Results

1. Pre-execution Decision Judgment (Action Judge Benchmark)

A benchmark of 3,000 decision samples (1,000 each from Search, Terminal, SWE) was constructed. Models only see pre-execution information, not the current action's tool output or future trajectory. AEWM achieves Macro-F1 70.5% , surpassing the strongest baseline DeepSeek-V4-Pro (59.9%) by 10.6 percentage points . Domain Macro-F1 scores: Search 60.9%, Terminal 72.1%, SWE 77.8%, all above respective strongest baselines.

2. End-to-End Task Performance (EditAct)

Six benchmarks: BrowseComp, DeepSearchQA, Terminal-Bench 2.0, Doc2Repo, NL2Repo, SWE-Bench Pro. Baselines: ReAct, Best@3 over three candidate decisions, Best@3 over three candidate trajectories. Across 18 combinations (6 benchmarks × 3 agent bases), EditAct outperforms all baselines.

For Qwen3.5-4B, Qwen3.5-9B, Qwen3.5-35B-A3B, EditAct's six-benchmark average scores reach 41.8, 44.1, and 48.8, improving over respective strongest baselines by 6.7, 5.2, and 3.2 percentage points.

3. Ablation: Is the Gain Just from Extra Generation?

Comparisons include random revision triggering, agent self-regeneration, providing AEWM as a prompt, and revising only reasoning or only action. On Terminal-Bench 2.0 with Qwen3.5-35B-A3B, self-regeneration scores 40.4%, prompt-only 39.3%, while full EditAct reaches 48.3%. The gain is not merely from giving the agent another generation chance; when to intervene and how to jointly revise reasoning and action both matter .

4. Distilling Corrections: AEWM-RFT

To test whether agents retain improvements without online AEWM, the authors propose AEWM-RFT : collect quality-filtered EditAct trajectories supported by real feedback, then apply rejection sampling fine-tuning. Control Self-RFT uses the agent's own trajectories. All settings (tasks, trajectory count, filtering, fine-tuning config) are identical; only trajectory generation differs. Evaluation is performed without AEWM.

AEWM-RFT outperforms Self-RFT on BrowseComp (+2.2), Terminal-Bench 2.0 (+2.6), and Doc2Repo (+2.5) percentage points.

This shows that revised interaction trajectories are not just one-off external help; they can become training supervision for the agent's subsequent learning.

Conclusion

For long-horizon agents, calling tools and receiving real feedback is only part of reliable task completion. More critical is the agent's ability to continuously distinguish verified facts from guesses, and solved problems from unmet requirements.

AEWM explores a world-model path for long-horizon agents: from "predicting what the tool returns" to "judging whether the current decision is useful" to "directly revising the next step and learning from it." Its goal is not to make agents do more or uniformly do less, but to make every step better grounded. Let the real environment provide facts; let the world model help the agent take a better next step based on those facts.