Fixing bad reasoning before it compounds matters more than simulating tool responses

Most language world models for LLM agents are built around a single intuition borrowed from embodied AI: predict what the environment will look like after an action. The Agent-Editing World Model (AEWM) argues this is the wrong target. Tool responses in real agent tasks—search rankings, terminal outputs, test results—are high-entropy and execution-dependent. Reconstructing them is hard, and when real tools are available, simulating them adds fabricated evidence without improving the agent's interpretation of what it already knows.

The actual failure mode AEWM identifies is subtler: task-state contamination. Agents exploring unfamiliar environments accumulate unsupported assumptions, retain outdated plans despite contradictory feedback, and mistake partial progress for completion. These errors persist in history and compound through locally plausible decisions even when the observations themselves are accurate. The problem is not bad observations—it is bad reasoning that survives them.

AEWM addresses this by modeling decision effects rather than environment states. Its Action Judge component classifies each proposed reasoning–action pair, before execution, as Critical, Exploratory, or Noisy. The three-way distinction matters: Exploratory preserves useful uncertainty reduction that a binary good/bad label would discard. When a proposal is judged Noisy, State Revision replaces both the reasoning and the action from the same observed history—not just the action, and not a critique handed back to the agent. The revised action then runs in the real environment. This is EditAct: selective pre-execution editing that changes what enters the agent's future history, not just what happens next.

The numbers are specific. On a 3,000-decision cross-domain benchmark, AEWM reaches 70.5% macro-F1, beating the strongest frontier baseline by 10.6 points. Across six benchmarks and three agent backbones—Qwen3.5-4B, Qwen3.5-9B, and the mixture-of-experts Qwen3.5-35B-A3B—EditAct improves average scores by 3.2 to 6.7 points over the strongest baseline. Qwen3.5-9B with EditAct outperforms Qwen3.5-35B-A3B with ReAct (44.1 vs. 42.2), a meaningful compression of the capability gap. The ablations are thorough: random intervention at the same rate as learned judgment underperforms, agent resampling after a Noisy label underperforms, and providing AEWM's reasoning as a hint rather than committing it directly underperforms. Direct state replacement beats all three alternatives.

AEWM-RFT closes the loop. Verified EditAct trajectories—with revisions included—are used to fine-tune the agent through rejection sampling. At inference time, the fine-tuned agent runs standard ReAct with no world-model guidance, yet still beats Self-RFT by 2.2 to 2.6 points across three domains. Average turn counts drop on search and terminal tasks, suggesting the agent internalizes more efficient decision patterns rather than just succeeding more often by chance.

The training scale is substantial: mid-training runs on roughly 52 billion tokens combining raw trajectories with synthesized judgment and revision data, followed by supervised fine-tuning on 120K curated examples evenly split between Action Judge and State Revision across three domains. The two-stage recipe matters—ablations show that SFT alone leaves significant performance on the table, particularly on terminal and software engineering tasks where broad interaction knowledge from mid-training provides a foundation that targeted fine-tuning then sharpens.