Paper Overview
- Field: Machine Learning
- Authors: Jiaxi Li, Ke Deng, Yun Wang
- Released: 2025-06-01
- arXiv: 2606.04391
Problem
Language agents increasingly rely on reusable skills to improve multi-step web automation across related tasks. In online skill learning, 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 held fixed throughout execution.
This static strategy is misaligned with web execution, where the appropriate next action depends not only on the task goal but also on the current webpage state — which often transitions into situations that the initial skills fail to cover.
Proposed Method: SGDR
State-Grounded Dynamic Retrieval (SGDR) is an online skill learning method enabling step-wise skill reuse for web agents. It consists of three components:
1. Sliding-window extraction process — converts completed trajectories into reusable subprograms callable at intermediate execution states. 2. Dual text-code representation — connects skill retrieval with executable actions. 3. State-grounded dynamic retrieval mechanism — matches skills against both the task goal and the current webpage state.
Results
Experiments on WebArena across five domains show SGDR consistently outperforms strong baselines:
| Backbone | Average Success Rate | Gain over strongest baseline | |---|---|---| | GPT-4.1 | 37.5% | +10.6% | | Qwen3-4B | 24.3% | +10.0% |
Paper: arxiv.org/abs/2606.04391