In an arXiv preprint, Salesforce AI Research authors report that a bank of page-specific yes-or-no checks raised a browser agent’s task success from 61.8% to 74.6% on a nine-app test set. The same bank supplied feedback during training and helped choose among complete task attempts later, so the reported gain combines both steps.
A pass-or-fail score at the end of a browser task does not show which steps helped the agent. CLIFT asks simple questions about what the agent did at each website address, such as whether it reached a page with useful content. During training, a language-model judge scores sample runs. The researchers keep a check only when its answers reliably track the judge’s scores for that site or page, then use the checks to give the agent feedback on individual steps.
At test time, CLIFT freezes the same question bank. The model compares the evidence and page sequences from an initial attempt and its retries several times. It replaces the first attempt only when most votes favor a retry; ties keep the first. This selection step does not call the external judge, although the judge is still used to train the agent and calibrate the checks.
On the public WebArena Infinity comparison across nine apps, a Gemma 4 31B agent scored 61.8% without fine-tuning, 64.7% after ordinary reinforcement learning, 72.6% with CLIFT’s step-level checks, and 74.6% after adding three-attempt selection. Most of the gain came during training; selection added 2.0 percentage points. In a separate 300-task live-web test, the researchers translated and recalibrated the checks without fine-tuning an agent on that benchmark. Selection raised GPT-5.5’s reported success from 49.7% to 61.0%, with an LLM rubric grading results.
The results come from one preprint and a single open training backbone. The judge remains part of training and check calibration, and the authors say the question bank may not transfer reliably to every site. The live-web result also depends on an LLM judge.
Sources
CLIFT paper (arXiv); full paper and results tables; Figure 2 by the authors, CC BY 4.0.