Agent-Editing World Model: Rethinking World Modeling for LLM Agents
Shuang Sun, Guoxin Chen, Fanzhe Meng, Jia Deng, Huatong Song, Jinhao Jiang, Wayne Xin Zhao, Hongteng Xu and 1 others
Renmin University of China
Published September 23, 2026 · on the feed September 25, 2026 · ▲ 24 · arXiv 2609.28416
The abstract
Recent advances in large language models (LLMs) have enabled agents to tackle long-horizon tasks across diverse environments. To further improve agent performance, existing language world models typically predict environment observations, yet reconstructing high-entropy, execution-dependent tool responses offers limited value when real feedback is available. Meanwhile, agents suffer from task-state contamination, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions. We propose the Agent-Editing World Model (AEWM), which models how reasoning and actions shape future task progress rather than simulating tool responses. AEWM combines Action Judge to distinguish Critical, Exploratory, and Noisy decisions with State Revision to edit noisy reasoning--action continuations from the same observed history. EditAct integrates these capabilities with real execution, directly changing the state underlying subsequent decisions rather than merely providing critiques. We train AEWM across Search, Terminal, and Software Engineering through mid-training and supervised fine-tuning. AEWM achieves 70.5\% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones, EditAct improves average scores by 3.2--6.7 points over the strongest baseline. Furthermore, rejection sampling fine-tuning on verified EditAct trajectories, termed AEWM-RFT, improves over Self-RFT by 2.2--2.6 points across three domains without online AEWM guidance.
The Brief
2:35A short podcast summary of the paper. Written and voiced by AI.
Transcript
Agent-Editing World Model. Rethinking World Modeling for LLM Agents. By Shuang Sun and eight others, from Renmin University of China. Published on the twenty-third of September. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.
Agents built on large language models now take on long-horizon tasks across many environments, and one common way to push them further is a language world model that predicts what the environment will return. The paper questions the value of that. Tool responses are high in entropy and depend on execution, so reconstructing them offers limited value when the real feedback is available anyway. A second problem sits inside the agent's own history. The authors call it task-state contamination: unsupported assumptions and outdated plans persist there, and distort the decisions that follow.
The Agent-Editing World Model, AEWM, changes what is being modeled. Instead of simulating tool responses, it models how the agent's reasoning and actions shape future progress on the task. It has two capabilities. Action Judge sorts decisions into three kinds: Critical, Exploratory and Noisy. State Revision edits noisy continuations of reasoning and action, working from the same observed history.
EditAct is what puts both to work alongside real execution. The distinction the authors draw is between a critique and an edit. EditAct does not merely comment on what the agent did. It directly changes the state that underlies the agent's subsequent decisions. AEWM itself is trained across three domains, Search, Terminal and Software Engineering, through mid-training and supervised fine-tuning.
On the authors' own Action Judge benchmark, AEWM reaches seventy point five percent macro-F1, which is ten point six points above the strongest frontier baseline. Across six benchmarks and three agent backbones, EditAct raises average scores by between three point two and six point seven points over the strongest baseline.
A further result takes the world model out of the loop. Rejection sampling fine-tuning on verified EditAct trajectories, called AEWM-RFT, improves over Self-RFT by two point two to two point six points across three domains, with no online AEWM guidance at all.
Seen whole, the proposal moves a world model's job from forecasting the environment to looking after the agent's own state. Predicting what a tool will say is left to the tool. What the model watches is whether the history the agent reasons from still holds up, and when it does not, the model fixes that history rather than annotating it. The last result adds that some of the gain survives even when the world model is no longer running online.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
An agent's worst enemy can be its own history. Here is the wall.
You build an agent that searches the web, works in a terminal and fixes code over long tasks. To help it plan, a common idea is a helper model that predicts what each tool will return.
The road ahead
- 1 · The problem ~7 min A helper that predicts tool outputs guesses what the real tool will soon say anyway, while bad assumptions in the agent's history go unfixed.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Instead of predicting tool outputs, AEWM predicts how each proposed step will affect task progress, rewrites the unproductive ones, and lets real tools answer.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Train one model to judge and revise, use it to filter and edit each step before real execution, then fine-tune agents on the verified results.
- 6 · The experiment ~6 min You’ll guess: On the authors' Action Judge benchmark, by how many macro-F1 points did AEWM beat the strongest frontier baseline?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.