ReTrace is an evaluation framework for how computer-use agents recover from their own mistakes, and the paper argues that model-based recovery improves long-horizon reliability.
The criterion: Transition-Consistent Recovery, meaning identical downstream actions from the recovered and target states must produce identical outcomes.
The method: Goal-oriented action replay after controlled error injection, since environment resets are infeasible in proprietary settings and full action replay is expensive over long horizons.
The findings: Ideal recovery can unlock latent task completion, current agents fail to exploit their latent recovery capabilities, and accurate model-based recovery yields measurable end-to-end gains.
Resources: Code and dataset are released on GitHub.
Sign in to comment.
TL;DR of the paper, generated automatically.
ReTrace is an evaluation framework for how computer-use agents recover from their own mistakes, and the paper argues that model-based recovery improves long-horizon reliability.