WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites
WebFovea placed second in WebRetriever 2026 after harness fixes lifted its score from 31.0 to 57.0.
WebFovea is a vision-based web agent that finished second in the WebRetriever Challenge 2026 with a score of 57.0 out of 100 on Protocol III. The benchmark requires an agent to start from an entry URL on a live site, operate the site interface, and return a verifiable answer. The authors argue many failures occur in the harness—parsing actions, applying them, reporting results, and showing the model the right information—rather than in model reasoning. Using the same model, harness changes raised the official hidden-set score from 31.0 to 57.0; observed bugs included clicks at three-quarters of intended coordinates and chat-template tokens contaminating 4.9% of episodes.
- Finished second in WebRetriever Challenge 2026 with a score of 57.0.
- Harness fixes raised the hidden-set score from 31.0 to 57.0 using the same model.
- Failures included a three-quarter coordinate mismatch, silent UI actions, and template-token contamination.
- Self-generated chat-template tokens contaminated 4.9 percent of task episodes.
From the original article
1 min read at huggingface.co
We present WebFovea, a vision-based web agent that placed 2nd in the WebRetriever Challenge 2026 with a final score of 57. 0 out of 100. The challenge evaluates agents end to end on Protocol III of the WebRetriever benchmark (arXiv:2607.
By Jiangang Han · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.