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

Researchers introduced LearnWeak, an annotation-free framework for specializing small computer-use agents.

It uses a stronger reference agent to spot student weaknesses, generate targeted tasks, and train with an error-aware objective.

On OSWorld, LearnWeak improved average performance by 11.6 points over EvoCUA-8B and 11.1 points over OpenCUA-7B across eight domains.

The approach offers a cheaper path to domain-specific gains without relying on human-labeled data or large expert models.