Agent-Editing World Model: Rethinking World Modeling for LLM Agents
AEWM edits contaminated agent state and improves six benchmarks by 3.2 to 6.7 points over the strongest baseline.
The Agent-Editing World Model predicts how reasoning and actions change future task progress instead of reconstructing high-entropy tool responses. It pairs an Action Judge, which labels Critical, Exploratory, and Noisy decisions, with State Revision that edits noisy continuations from the same history. Trained on Search, Terminal, and Software Engineering, it scores 70.5% macro-F1, 10.6 points above the strongest frontier baseline. EditAct then improves average scores by 3.2 to 6.7 points across six benchmarks and three agent backbones.
- AEWM models task progress rather than reconstructing tool responses.
- Action Judge separates Critical, Exploratory, and Noisy decisions.
- Action Judge reaches 70.5% macro-F1, 10.6 points above the best frontier baseline.
- EditAct improves average scores by 3.2 to 6.7 points across six benchmarks.
- AEWM-RFT beats Self-RFT by 2.2 to 2.6 points without online guidance.
Full article200 words · extracted from huggingface.co · click to collapse
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 Meanwhile, agents suffer from task-state contamination, where unsupported assumptions and outdated plans persist in history and distort subsequent decisions.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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.28416