SCAFFOLD: Self-Improving Web Agents Through Recursive Parametric Skill Abstraction

Summary

SCAFFOLD is a self-improving framework for visual web agents designed for long-horizon tasks on changing, visually rich websites. The authors argue that many existing agents learn tasks in isolation and discard procedural knowledge, while newer skill-augmented systems keep libraries that are flat or only weakly structured. SCAFFOLD induces parametric, executable skills from successful trajectories under a multi-instance abstraction constraint, so the resulting skills are intended to capture reusable procedures rather than one-off traces. It organizes these skills into a recursive hierarchy, allowing higher-level skills to invoke lower-level ones. The framework also compacts its library with a minimum-description-length criterion and behavioral-equivalence checks to reduce redundancy. At regular intervals, skill-augmented trajectories are distilled back into the model weights so that useful abstractions can become internalized rather than remaining only in an external library. Across WebArena, VisualWebArena, and a held-out Online-Mind2Web split, SCAFFOLD improves success rates by 11.1 to 17.2 absolute percentage points over the strongest skill-augmented baseline. Performance rises monotonically over five self-improvement iterations, and the authors report no library collapse. Code and documentation are released in a GitHub repository.