CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling
CLIFT turns judge feedback into reusable rewards and judge-free test-time selection for web agents.
CLIFT trains web agents with conformal self-verification: the agent answers questions about its own rollouts, and a Compositional Conformal Certifier keeps only signals whose URL-conditional evidence agrees with a training-time judge, then blends signed trust weights into per-step rewards without subtracting from the judge baseline. At test time the frozen question bank drives Conformal Trajectory Selection among a greedy rollout and retries, with no external judge. The paper reports state-of-the-art open-source results on WebArena Infinity, transfer of an open-model bank to GPT-5.5 for state-of-the-art on VisualWebArena, and a zero-shot gain on Online Mind2Web.
- A conformal certifier keeps only self-verification signals that match a training-time judge
- Blended rewards never fall below the judge baseline
- The frozen question bank selects trajectories at test time without an external judge
- Reports open-source SOTA on WebArena Infinity and SOTA on VisualWebArena via GPT-5.5
- Tags
- web-agentsconformal-predictionreinforcement-learningwebarenavisualwebarenamind2webtest-time-scaling
- Products
- WebArenaVisualWebArenaMind2Web
- AI models
- GPT-5.5
From the original article
1 min read at huggingface.co
Open-source web agents are now strong enough to execute realistic browser tasks, but training them with reinforcement learning still depends on weak supervision: binary task success is too sparse for credit assignment, while frontier-language-model judges are too expensive to call at every step and cannot be assumed available at deployment.
By Yifan Zhang, Yutong Dai, Viraj Prabhu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.