Learn from Weaknesses: Automated Domain Specialization for Small Computer-Use Agents

Researchers introduce LearnWeak, a framework that improves small computer-use agents by having them learn from their own failures in specific domains rather than training on generic synthetic data. The approach achieves 11-12 percentage point improvements on benchmark tests, demonstrating that targeted, error-aware specialization is more efficient than broad data synthesis for adapting AI agents to particular software environments.

The development of computer-use agents capable of operating software autonomously represents a significant challenge in AI research, as deploying separate large expert models for each domain becomes prohibitively expensive. LearnWeak addresses a core inefficiency in current specialization approaches: the assumption that more training data automatically yields better performance. The research reveals that naive synthetic data generation provides only marginal gains, suggesting that data quality and relevance matter far more than quantity alone.

The framework's innovation lies in its student-aware methodology, where a stronger reference agent actively identifies where smaller agents fail within specific domains, then generates targeted training tasks that address these exact weaknesses. This mirrors human learning principles—struggling students benefit more from focused remedial instruction than from broadly comprehensive material. The error-aware specialization objective further refines the approach by distinguishing between planning errors (strategic mistakes) and execution errors (tactical failures), enabling more precise behavioral corrections.

For the AI industry, this research has meaningful implications for cost-effectiveness and accessibility. Small open-source models have practical advantages over massive proprietary systems—lower computational requirements, easier deployment, and faster inference. LearnWeak's 11.6 and 11.1 percentage point improvements over existing small models suggest a pathway toward making open-source agents genuinely competitive within specific domains without requiring prohibitive computational resources or proprietary data.

The work points toward a broader principle: specialization should be systematic and evidence-driven rather than brute-force. Future development will likely focus on automating domain identification, scaling this approach across more complex software environments, and determining whether error-aware training techniques transfer to other agent architectures and tasks.

• →LearnWeak improves small computer-use agents by identifying and targeting their specific failures rather than generating generic synthetic training data.
• →The framework achieves 11-12 percentage point performance gains by distinguishing between planning and execution errors during training.
• →Student-aware dataset generation outperforms existing autonomous trajectory generation methods, suggesting data relevance exceeds data volume.
• →The approach makes specialized AI agents more accessible by improving small open-source models rather than requiring separate large expert models.
• →Error-driven, domain-specific training represents a more principled path for efficient agent specialization across diverse software environments.