Reproduce “DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning” 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
Xirui Liu, Sihang Zhou, Yanning Hou, Rong Zhou, Haoyuan Chen, Maolin He, Siwei Wang, Hao Chen and 1 more
The DRIVE framework introduces a dual-level skill modeling approach for web agents, explicitly separating abstract reasoning skills from executable interaction skills. This method, combined with scene-aware coordination and continual refinement, improved task success rates by 7.3 percentage points on the WebArena benchmark, demonstrating more robust adaptation in dynamic web environments.
Key insights
- Explicitly separating reasoning and interaction knowledge resolves the abstraction-grounding mismatch, allowing agents to effectively manage both transferability and executability of experience.
- Continual learning is enhanced by tailoring skill induction and refinement strategies to the specific nature of reasoning and interaction knowledge, addressing distinct failure modes.
- Reasoning skills and interaction skills play complementary roles: reasoning skills guide 'what to do' (task understanding), while interaction skills define 'how to do it' (reliable execution).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.