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, et al.
cs.CL · cs.AI · cs.LG
Quick verdict
Proposes AEWM: judge and edit agent decisions instead of predicting tool outputs; strong gains but no significance tests and possible label circularity.
Summary
The paper proposes the Agent-Editing World Model (AEWM), which rethinks language world modeling for LLM agents: instead of predicting environment observations or tool responses, AEWM predicts how a proposed reasoning-action pair will affect task progress and edits it when predicted to be unproductive. AEWM combines an Action Judge that classifies decisions as Critical, Exploratory, or Noisy with a State Revision module that rewrites noisy continuations from the same history; EditAct integrates both with real execution, replacing the agent's proposed state before acting. A 35B model is trained via mid-training on ~52B tokens of synthesized judgment/revision data plus trajectories, then SFT on 120K curated examples across Search, Terminal, and SWE domains. On a self-constructed 3,000-decision Action Judge benchmark, AEWM reports 70.5% macro-F1, 10.6 points above the strongest frontier baseline (DeepSeek-V4-Pro, 59.9%). With EditAct, three Qwen3.5 backbones improve average scores by 3.2-6.7 points across six benchmarks over the strongest baseline (35B agent: Terminal-Bench 2.0 +11.2, SWE-Bench Pro +3.2 over agent-only). Rejection-sampling fine-tuning on EditAct trajectories (AEWM-RFT) adds 2.2-2.6 points over Self-RFT without online guidance. Main limitations: no variance or significance testing despite averaged runs, benchmark labels generated by the same frontier models used as baselines and training-data sources, partly internal training data, and unquantified inference overhead.
Review complete
Assessment at a glance
4
risk flags identified
● high 1
● medium 2
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Modeling decision effects and editing agent state is a more valuable world-modeling target than predicting environment observations, because observations are high-entropy, execution-dependent, and redundant given real tool feedback.
- AEWM achieves 70.5% macro-F1 on its 3,000-decision Action Judge benchmark, exceeding the strongest frontier baseline (DeepSeek-V4-Pro, 59.9%) by 10.6 points.
- EditAct improves average scores over the strongest baseline by 6.7, 5.2, and 3.2 points for Qwen3.5-4B, 9B, and 35B-A3B respectively across six benchmarks and three backbones.
- AEWM-RFT internalizes AEWM-guided decision patterns, surpassing Self-RFT by 2.2-2.6 points on BrowseComp, Terminal-Bench 2.0, and Doc2Repo without online AEWM guidance at inference.
- Direct replacement of noisy reasoning-action continuations outperforms resampling and critique-style guidance, and joint reasoning-action editing beats editing either component alone (Sec. 4.5 ablations).
Strengths
Well-motivated reframing with a named failure mechanism
The shift from predicting high-entropy, execution-dependent observations to modeling decision effects (Sec. 1-2.1) is argued concretely via task-state contamination, and Appendix D case studies (e.g., the IPv6 socket misdiagnosis; the ICS timestamp regex) show the mechanism operating on real trajectories rather than toy examples.
Ablation suite that actually tests the design, not just the whole system
Sec. 4.5 isolates learned gating (Random Gate), direct replacement vs regeneration vs critique (Agent Resampling, AEWM Hint), joint vs single-component editing (Reasoning-only / Action-only Revision), and world-model identity (Self-WM, DeepSeek-V4-Pro WM); the training-stage ablation quantifies mid-training complementarity (+0.8/+6.4/+4.3 on BC/TB2/Doc2Repo over SFT-only).
Unusually transparent data construction
Honest class-level reporting
Behavioral evidence beyond task scores
Weaknesses
No variance, confidence intervals, or significance tests anywhere
Benchmark-label circularity
Retrospective labels from successful trajectories only
Author-built scaffolds and in-family training data cloud generalization claims
Inference cost of EditAct is unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
AEWM reframes language world modeling for LLM agents. Rather than training a model to predict environment observations or tool responses - which the authors argue are high-entropy, execution-dependent, and redundant given real feedback - AEWM (a Qwen3.5-35B-A3B backbone) models how a proposed reasoning-action pair will shape task progress. Two capabilities: **Action Judge** classifies the pre-execution state as CRITICAL/EXPLORATORY/NOISY; **State Revision** rewrites noisy continuations from the same history. **EditAct** deploys both at inference: noisy proposals are replaced before real execution, so the intervention changes the history that future decisions condition on, rather than merely critiquing. The motivating failure mode is task-state contamination (Sec. 1-2.1; case analyses in Sec. 5.2 and App. D). Three headline claims: (i) 70.5% macro-F1 on a self-built 3,000-decision Action Judge benchmark, +10.6 over the strongest frontier baseline (Table 11); (ii) EditAct lifts average scores over the strongest baseline by 6.7/5.2/3.2 points for Qwen3.5-4B/9B/35B-A3B across six benchmarks (Sec. 1); (iii) AEWM-RFT beats Self-RFT by 2.2-2.6 points on BrowseComp, Terminal-Bench 2.0, and Doc2Repo without online guidance (Sec. 4.4).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the Agent-Editing World Model (AEWM)?
AEWM is a trained 35B model that acts as a world model for LLM agents: instead of simulating what tools will return, it judges whether the agent's next reasoning-action step is critical, exploratory, or noisy, and rewrites noisy steps before they execute. The paper reports 70.5% macro-F1 on a 3,000-decision judgment benchmark and average agent gains of 3.2-6.7 points across six benchmarks and three model sizes.
What is task-state contamination in LLM agents?
It is the paper's name for a failure mode where unsupported assumptions, outdated plans, or prematurely treated progress persist in the agent's interaction history and distort later decisions - for example, chasing a search candidate that evidence already contradicted, or building on unverified code. AEWM's State Revision edits the pending reasoning-action pair in history to remove such contamination before execution.
How much does EditAct improve agent performance and on which benchmarks?
Across six benchmarks (BrowseComp, DeepSearchQA, Terminal-Bench 2.0, Doc2Repo, NL2Repo, SWE-Bench Pro) and three Qwen3.5 backbones (4B/9B/35B-A3B), EditAct improves average scores by 6.7, 5.2, and 3.2 points respectively over the strongest baseline. For the 35B agent, per-benchmark gains over the agent-only setup range from +2.9 (NL2Repo) to +11.2 (Terminal-Bench 2.0). Results are means of three runs; the paper reports no variance or significance tests.
Is the code or the AEWM model checkpoint publicly available?
The paper sidebar references a dataset GitHub repository, but the extracted text does not state that training code, the AEWM checkpoint, or the Action Judge benchmark data are publicly released. Full reproduction would also require proprietary frontier models (DeepSeek-V4-Pro, GLM-5, GPT-5.6-Sol) for data synthesis and auditing, plus partially internal training data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The main results section (Tables 1-2, covering per-backbone benchmark scores for Sec. 4.2-4.4) was omitted from the extracted text; backbone-level numbers beyond Figure 1's 35B-agent values could not be verified.
- Table 3 (inference/training ablations) content was not recoverable from the extracted text; ablation conclusions rely on the prose summary in Sec. 4.5.
- Figures are available only as embedded text; visual details of Figures 2-5 could not be independently analyzed.
- Appendix E prompt templates are truncated mid-way, so some annotation boundary rules may be incompletely captured.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.