An agent world model edits contaminated reasoning instead of predicting tool output

What's this about?

AEWM identifies noisy or exploratory decisions and revises the reasoning state that future actions consume, while still using real tool execution for actual observations.

What exactly is task-state contamination?

Wrong assumptions and outdated plans linger in an agent's conversation history. Later steps treat that stale text as truth, so one early mistake keeps steering the task off course.

Why does this matter?

1. Its action judge reaches 70.5% macro-F1, 10.6 points above the strongest tested frontier baseline.
2. EditAct adds 3.2–6.7 points across six benchmarks and three backbones.
3. Editing history can also erase useful uncertainty, so reliability depends on the judge.