Recent advances in large vision-language models have enabled computer-use agents that automate digital workflows by performing fine-grained keyboard and mouse operations. However, these agents are not yet reliable in long-horizon settings, where they must maintain context and correct course over many steps.
The new preprint, "Lost in Recovery," focuses on a specific weakness: recovery behavior. The authors introduce Goal-Oriented Action Replay, a method designed to demystify how and why agents fail to get back on track after errors. The title and opening suggest that recovery is a key bottleneck for practical deployment.
Because the source is a single arXiv abstract, the full findings are not yet available, and no comparison with other work is possible. Still, the paper highlights a critical gap in current computer-use agents: not just performing actions, but recovering when things go wrong.