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

Forum topic · 小凯 · 2026-10-07

Summary

CLIFT is a training and test-time scaling framework for open-source web agents built around conformal self-verification, presented in arXiv paper 2610.06829. During reinforcement learning training, the agent answers natural-language verification questions about its own rollouts. A Compositional Conformal Certifier keeps only question signals whose URL-conditioned evidence agrees with a judge from training, assigns signed trust weights via polarity-aware boosting, and blends verifier scores into per-step rewards without degrading the judge baseline. At test time, the certified question bank is frozen and reused as structured evidence for Conformal Trajectory Selection (CTS): the agent samples a greedy rollout plus diverse retries, summarizes each URL-level trajectory with self-verifiers, and a conservative majority-vote rule selects the better trajectory without any external judge. On WebArena Infinity, CLIFT achieves state-of-the-art performance among open-source web agents. On VisualWebArena, a library trained with open models transfers to GPT-5.5 at test time, reaching SOTA under standard evaluation. On Online Mind2Web, a translated certified question bank improves live web agents in zero-shot evaluation.

Paper Overview

Research Area: LLM Agents Authors: Yifan Zhang, Yutong Dai, Viraj Prabhu, Zhiyuan Hu, Ran Xu, Zeyuan Chen Published: 2026-10-05 arXiv: 2610.06829

Summary

Open-source web agents are now capable of performing real browser tasks, but training them with reinforcement learning still relies on weak supervision: binary task-success signals are too sparse for credit assignment, while invoking frontier LLM judges at every step is prohibitively expensive and cannot be assumed available at deployment time.

CLIFT addresses this by building conformal self-verification into both training and test-time scaling:

During training: the agent answers natural-language verification questions about its own rollouts. A Compositional Conformal Certifier retains only question signals whose URL-conditioned evidence agrees with the judge from the training period, assigns signed trust weights through polarity-aware boosting, and blends verifier scores into per-step rewards in a way that does not degrade the judge baseline.
At test time: the same certified question bank is frozen and reused as structured evidence for Conformal Trajectory Selection (CTS). The agent samples one greedy rollout plus one or more diverse retries, summarizes each URL-level trajectory with the self-verifiers, and a conservative majority-vote rule decides whether to switch away from the current best trajectory — all without calling any external judge.

Results

WebArena Infinity: CLIFT achieves state-of-the-art performance among open-source web agents.
VisualWebArena: a library trained with open-source models transfers at test time to GPT-5.5, reaching SOTA under the standard evaluation.
Online Mind2Web: a translated certified question bank improves live web agents in zero-shot evaluation.

Original Abstract

CLIFT builds conformal self-verification into web agent training and test-time scaling. A Compositional Conformal Certifier filters verification signals, blends them into per-step rewards, and enables judge-free Conformal Trajectory Selection at test time.

Link: https://arxiv.org/abs/2610.06829

--- *Auto-collected on 2026-10-07*

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

Summary

Paper Overview

Summary

Results

Original Abstract

Tags