CLIFT self-verifies web agents, claims open-source SOTA

Open-source web agents can execute realistic browser tasks, but training them with reinforcement learning is stuck between binary task-success rewards that are too sparse for credit assignment and frontier-model step judges that are too expensive to call every turn. CLIFT, posted to arXiv on October 5, 2026 by Yifan Zhang, Yutong Dai, Viraj Prabhu, Zhiyuan Hu, Ran Xu and Zeyuan Chen, proposes a different plumbing: have the agent answer natural-language verification questions about its own rollouts, then filter those self-answers through what the authors call a "Compositional Conformal Certifier" that keeps only signals whose URL-conditional evidence agrees with a training-time judge.

The certified bank then does two jobs. During training it blends into per-step rewards "in a way that never subtracts from the judge baseline," per the abstract, which leaves the external judge as a floor and lets the self-verifier lift scores only where it has URL-conditional evidence. At test time the same bank is frozen and reused for Conformal Trajectory Selection: the agent samples a greedy rollout plus one or more diverse retries, the self-verifier summarises each URL trace, and "a conservative majority-vote rule chooses whether to swap away from the current incumbent without calling any external judge."

The authors report three outcomes from this one mechanism. On WebArena Infinity they claim state-of-the-art among open-source agents. On VisualWebArena, a bank trained with the open model is applied at test time to GPT-5.5 and "reaches state-of-the-art performance under the canonical harness." On Online Mind2Web, translating the question bank is said to improve a live-web agent in zero-shot evaluation, with no training on the benchmark.

The abstract publishes no accuracy numbers, no model-size disclosure, and no cost comparison against a per-step frontier judge; the state-of-the-art claims are summarised rather than quoted, so any reader auditing the headline needs the paper's tables, not the abstract.