AI 代理與瀏覽器自動化
Wuying Browser Agent Trains on Error-Recovery Data for Long-Horizon Tasks; 27B Model Reaches 65.1% on BrowserBench
The Alibaba Cloud team integrates browser-agent error recovery, complex UI interaction, and long-horizon credit assignment into a unified training pipeline. Its 27B model achieves 65.1% Pass@1 on the newly created bilingual Chinese-English BrowserBench, though the public availability of the model weights and evaluation data remains to be confirmed.
Alibaba Cloud’s AI Model Application & Engineering team has introduced Wuying-Browser-Agent. Rather than simply scaling up the model, the project redesigns execution, data, and reinforcement learning around the failure modes of long-horizon workflows on real-world websites. Its browser harness provides 24 structured actions and reconstructs the decision context at each step using the current DOM, action history, and environment feedback, preventing stale page states from accumulating in the conversation history. All training and evaluation run in isolated AgentBay browser environments.
The first stage, RUIC-SFT, combines standard successful trajectories, data involving complex components such as date pickers and rich-text editors, and “make an error—detect it—reflect—correct it” trajectories. According to the paper, a model trained only on standard supervised data recovers from detected errors just 8.5% of the time. In the second stage, DAO-GRPO supplements sparse task-completion scores with progress rewards. It also compares divergence points across multiple rollouts for the same task, assigning greater weight to branches that genuinely determine success or failure instead of applying uniform updates across the entire shared prefix.
The team also created BrowserBench, comprising 350 bilingual Chinese-English tasks on real-world websites across 254 sites. Tasks require an average of 37.9 steps, with a maximum of 100 steps allowed. Wuying-Browser-Agent-27B reports success rates of 80.6% on WebVoyager, 66.7% on Online-Mind2Web, and 65.1% on BrowserBench. When 500 trajectories were manually audited, the judge used to estimate reported performance agreed with human assessments 96.4% of the time.
The engineering value lies in treating recovery capability as a trainable objective and allowing SFT, online RL, and deployment to share the same tool protocol. However, nearly half of BrowserBench’s tasks involve e-commerce information, and success evaluation partly relies on a judge from the same family as the one used for training rewards. Continuous changes to websites may also affect rerun results. The paper states that 4B, 9B, and 27B models will be released, but as of publication, it remains necessary to confirm whether the weights, data, and complete evaluation code are publicly available before the claim of “best open-source performance” can be independently verified.