Navers lab
← Research

BrowserBC

Scalable behaviour cloning for browser use via skill distillation

BrowserBC distills successful human browser trajectories into reusable skills, helping web agents solve tasks with higher success rates and fewer interactions.

Navers Lab · Einsia.AI · 2026

arXiv Code Case demos Contact

Core Idea

Human browsing trajectories contain more than clicks. They encode efficient paths, site-specific logic, and practical decisions that are hard for agents to infer from the current page alone.

BrowserBC turns these trajectories into reusable skills. Those skills give agents prior knowledge for acting under incomplete information, moving them from simply operating websites to operating them efficiently.

Method

BrowserBC extracts task evidence from human runs, summarizes it into skills, and retrieves relevant skills during new browser tasks.

BrowserBC method overview from raw trajectories to evidence extraction, skill distillation, and skill graph management.

Results

BrowserBC improves both success rate and interaction efficiency across WebArena-Hard and ClawBench.

WebArena-Hard · 258 tasks
Group Skill-off BrowserBC Gain
Overall 60.5 81.4 +20.9
GitLab 64.9 86.0 +21.1
Shopping 60.7 89.3 +28.6
Shopping admin 56.4 70.9 +14.5
Reddit 78.6 85.7 +7.1
Multi-site 43.8 75.0 +31.2
ClawBench · 152 tasks
Group Skill-off BrowserBC Gain
Overall 32.9 68.4 +35.5
Daily 24.6 64.9 +40.3
Finance 50.0 100.0 +50.0
Work 47.1 76.5 +29.4
Dev 33.3 66.7 +33.4
Academic 50.0 78.6 +28.6
Travel 38.5 76.9 +38.4
Social 25.0 56.2 +31.2
Pets 27.3 54.5 +27.2

Efficiency

Mean WebArena-Hard tool calls drop from 31.2 to 22.7, median calls drop from 24 to 16, and Sonnet-distilled skills lift Qwen from 53% to 77%.

BrowserBC improves interaction efficiency by reducing average and median steps across WebArena-Hard task groups.
Skill distillation transfers across executors, with Sonnet-distilled skills improving both Qwen and Sonnet agents.

Live Case Demos

Watch skill-guided agents complete real browser tasks. Across 9 demos, BrowserBC reduces actions by 53% and tokens by 28% on average.

Open demos

Citation

@misc{yang2026scalablebehaviourcloningbrowser,
      title={Scalable Behaviour Cloning on Browser Using via Skill Distillation},
      author={Kaisen Yang and Zheng Jiang and Yuzhao Peng and Houde Qian and Boshi Zhang and Youjie Zheng and Shijin Hong and Qingle Liu and Ruoyu Han and Bohan Lyu and Bingxiang He and Eren Cai and Calvin Xiao and Qinhuai Na},
      year={2026},
      eprint={2606.32014},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.32014},
}
← Back to Research