Researchers propose a failure-driven self-improvement loop for computer-use agents that utilizes failed trajectories to enhance model performance. The method employs an LLM to diagnose failure modes and generate code patches, which are lightly verified by humans to upgrade the agent.
- The approach targets multimodal large language models (MLLMs) operating in verifiable environments.
- It complements standard success-based pipelines by extracting information from discarded failed trajectories.
- Validated on the OSWorld benchmark using the OpenCUA-72B model, the method improved the success rate from 42.3% to 48.9%.
- The improvement represents a gain of 6.6 percentage points with no additional training cost and only modest inference overhead.
This work demonstrates that failure-driven self-improvement is a viable complement to existing pipelines, enabling more efficient agent improvement.