1
Certify the questions
The Compositional Conformal Certifier checks agreement with judge feedback at URL-path, host, and global tiers. Reliable questions receive polarity-aware trust weights.
Salesforce AI Research · October 2026
Conformal Self-Verification for Web Agent Training and Test-Time Scaling
Turn judge feedback into a reusable bank of certified questions that helps web agents learn and choose better trajectories.
Learn what to verifyReuse at test time
WebArena Infinity
74.6%
+12.8 pp vs. base Gemma-4 · K = 3
VisualWebArena
53.7%
Bank transfer to GPT-5.5 · K = 4
Online Mind2Web
61.0%
+11.3 pp vs. GPT-5.5 high effort · K = 4
01 / THE METHOD
Feedback that lasts
beyond training.
Binary outcomes give sparse training signals. Language-model judges provide richer feedback, but their supervision is costly. CLIFT converts that feedback into URL-scoped verification questions with calibrated trust weights, then reuses the same bank during training and inference.
2
Improve credit assignment
Certified evidence adds a non-negative bonus to per-step judge rewards. URL-stratified advantages guide the GRPO policy update.
3
Select with repeated evidence
CTS compares completed trajectories using the frozen bank. A retry replaces the incumbent only after a majority of verifier votes supports the swap; ties keep the incumbent.
02 / THE EVIDENCE
Three settings.
One reusable bank.
Train an open web agent, transfer verification knowledge to a closed model, and adapt the bank to live websites. Every CTS result emits one selected trajectory per task; K is the candidate rollout budget.
Stronger training on an open backbone
Task-weighted success rate on the public 9-app comparison set. CLIFT trains Gemma-4 31B with LoRA; CTS adds inference-time selection.
Success rate (%) · axis range 50–80% · higher is better
Source: arXiv v1, Table 6. Training rows share the comparative judge and browser-use backbone. Training contributes +10.8 pp; CTS adds +2.0 pp.
Transfer the bank to a closed model
Environment-size-weighted success rate across Classifieds, Shopping, and Reddit. GPT-5.5 weights remain frozen; the bank learned during open-model training is reused at inference.
Success rate (%) · higher is better
Source: arXiv v1, Table 4. The final CTS row uses K = 4 and returns one trajectory per task. Published references are separate systems as reported in the paper.
Adapt verification to the live web
Realised success on 300 Online Mind2Web tasks. The question bank is translated, expanded, and recalibrated on OM2W rollouts; policy weights receive no OM2W-specific fine-tuning.
Success rate (%) · higher is better
Source: arXiv v1, §4.3 and Figure 5. The final GPT-5.5 row uses K = 4. Bank adaptation includes OM2W-specific questions and held-out rollout recalibration; no policy training on OM2W.
+12.8 pp
CLIFT + CTS improves Gemma-4 over its base WAI success rate. The gain combines stronger policy training with conservative trajectory selection.
The core idea
A trajectory score is used once. A certified question bank can be inspected, recalibrated, and reused. CLIFT makes that bank the shared interface between expensive judge supervision and an agent’s own rollouts: it supplies dense evidence during reinforcement learning and structured evidence for test-time selection. Across the evaluated settings, the bank supports open-model training, transfer to a frozen closed model, and adaptation to live websites.
Cite CLIFT
@misc{zhang2026clift,
title={CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling},
author={Yifan Zhang and Yutong Dai and Viraj Prabhu and Zhiyuan Hu and Ran Xu and Zeyuan Chen},
year={2026},
eprint={2610.06829},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.06829}
}