Salesforce AI Research 论文：31B 开源 Gemma-4 Web Agent 在 WebArena Infinity 达 74.6%
DAIR.AI 转发 Salesforce AI Research 的论文，一个 31B 开源 Gemma-4 web agent 在 9-app WebArena Infinity 上取得 74.6%，高于 Gemini 3 Flash browser use 的 70.1%，且无需在每步或部署时调用前沿裁判。
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: arxiv.org/abs/2610.06829
Chat with Paper: academy.dair.ai/papers/clift…
来源：dair_ai · x.com