Reproduce “Learning from Failure: Inference-Time Self-Improvement for Computer-Use Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xueqiao Sun, Xiaohan Wang, Ludwig Schmidt, Serena Yeung-Levy, Yuhui Zhang
Researchers at Stanford University developed a framework that improves computer-use agents by leveraging failed task attempts at inference time, using an LLM to diagnose errors and apply targeted code patches. This approach increased agent success rates on the OSWorld benchmark by 6.6 percentage points and generalized across various models and environments.
Key insights
- Failed trajectories, typically discarded, represent a structured source of supervision that can be systematically leveraged for agent refinement without additional training costs.
- Inference-time self-improvement, guided by an LLM, provides a cost-effective and scalable mechanism for continuous agent enhancement, reducing the reliance on extensive human annotations or iterative model retraining.
- The identified failure patterns and their corrective strategies are generalizable, acting as a model-agnostic amplifier of base agent capabilities across different models, scales, and heterogeneous task environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.