CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling
CLIFT turns judge feedback into a reusable verifier for web-agent training and judge-free test-time selection.
The method has the agent answer natural-language checks about its own browser rollouts, then keeps only verification signals that conform to a training-time judge. At inference, the certified question bank is frozen and used to compare a greedy rollout against retries without calling an external judge. The paper reports state-of-the-art results for open-source agents on WebArena Infinity, transfer gains on VisualWebArena with GPT-5.5, and zero-shot improvement on Online Mind2Web. ArXiv · AI/CL/LG's note
The method has the agent answer natural-language checks about its own browser rollouts, then keeps only verification signals that conform to a training-time judge. At inference, the certified question bank is frozen and used to compare a greedy rollout against retries without calling an external judge. The paper reports state-of-the-art results for open-source agents on WebArena Infinity, transfer gains on VisualWebArena with GPT-5.5, and zero-shot improvement on Online Mind2Web. ArXiv · AI/CL/LG's note
score 5