Salesforce CLIFT 方法让 31B Gemma-4 网页智能体在 WebArena 超过 Gemini 3 Flash
Salesforce 的新论文 CLIFT，31B 开源模型打网页任务居然赢了 Gemini 3 Flash，关键是部署时不用裁判模型，省成本。
Salesforce AI Research 提出 CLIFT 方法，让 31B 开源 Gemma-4 网页智能体在 9 个应用的 WebArena Infinity 上拿到 74.6%，高于使用 browser use 的 Gemini 3 Flash 的 70.1%。方法核心是让智能体回答关于自身 rollout 的验证问题，再用保形认证器筛选与训练时裁判一致的问题并加权加入每步奖励。测试时只用冻结题库在贪婪 rollout 和少量重试之间选择，无需外部裁判。该智能体比基础模型提升 12.8 分，在 9 个应用中赢下 7 个，题库还能迁移到 GPT-5.5（VisualWebArena）和 Online Mind2Web 的实时网页智能体上。
Another great paper from Salesforce AI Research.
The finding is that a 31B open Gemma-4 web agent scores 74.6% on the 9-app WebArena Infinity set, above Gemini 3 Flash with browser use at 70.1%.
They got there without calling a frontier judge at every step or at deployment.
CLIFT has the agent answer verification questions about its own rollouts.
A conformal certifier keeps only the questions whose answers agree with a training-time judge, weights them by how much they can be trusted, and adds the result to per-step rewards.
At test time, the same frozen question bank picks between a greedy rollout and a few retries, with no external judge.
The trained agent improves 12.8 points over its base model and wins 7 of 9 apps. The question bank also transfers to GPT-5.5 at test time on VisualWebArena, and a translated bank improves a live-web agent on Online Mind2Web without any training on that benchmark.
Paper: https://t.co/kzHeT4CYhW
Chat with Paper: https://t.co/SoMbjIxAuK
