Online Skill Learning for Web Agents via State-Grounded Dynamic Retrieval
Quick Answer
This paper shows that The State-Grounded Dynamic Retrieval (SGDR) method enhances online skill learning for web agents, enabling dynamic skill reuse based on current webpage states.
Quick Take
Experiments on WebArena show SGDR achieves 37.5% success with GPT-4.1 and 24.3% with Qwen3-4B, outperforming existing methods by 10.6% and 10.0%, respectively.
Key Points
- SGDR enables stepwise skill reuse tailored to current webpage states.
- It utilizes a sliding-window extraction process for reusable sub-procedures.
- The method connects skill retrieval with executable actions via dual text-code representation.
- SGDR outperforms strong baselines in five domains on WebArena.
- Code for SGDR is publicly available on GitHub.
Paper Resources
Source Excerpt
From the original publisher, up to about 700 charactersarXiv:2606. 04391v1 Announce Type: new Abstract: Language agents increasingly rely on reusable skills to improve multi-step web automation across related tasks. A growing line of work studies online skill learning, where agents continually induce skills from previous task trajectories and reuse them in future tasks on the fly. However, existing methods mainly reuse skills at the task-level: a fixed set of skills is retrieved based on the initial task instruction and then held fixed throughout execution. …
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →Automatic Ordinary Differential Equations Discovery For Biological Systems Using Powered Agentic System
The MEDA system utilizes large language models and symbolic regression to autonomously discover ordinary differential equations for biological systems, achieving strong structural recovery and biologically plausible models. It outperforms existing methods by integrating domain knowledge and mechanistic constraints, demonstrating effective retrieval and extrapolation capabilities.