CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling

Yifan Zhang, Yutong Dai, Viraj Prabhu, Zhiyuan Hu, Ran Xu, Zeyuan Chen

arXiv: 2610.06829Published: October 5, 20262 upvotes

AI summary

TL;DR

CLIFT introduces conformal self-verification for web agent training and test-time scaling, achieving state-of-the-art performance on WebArena Infinity and transferring to other models and benchmarks. This method enables reusable training signals and judge-free test-time scaling. CLIFT improves performance without relying on external judges.

Why It Matters

CLIFT's approach reduces the need for costly judge feedback, making it more practical for training and deploying web agents. This can lead to more efficient and effective web agent development. The method's ability to transfer to other models and benchmarks also increases its potential impact.

Key Findings

state-of-the-art performance on WebArena Infinity
successful transfer to GPT-5.5 on VisualWebArena
improved zero-shot evaluation on Online Mind2Web

Topics

conformal

self-verification

reinforcement learning

web agents

natural language processing

Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.