AI

WebFovea: When the Model Is Right but the Click Is Wrong -- Reliable Round Trips for Vision-Based Web Agents on Live Websites

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 02:18:26 UTC.

Briefings

WebFovea is a vision-based web agent that achieved a final score of 57.0 out of 100 in the WebRetriever Challenge 2026, placing second overall. The system operates by having an LLM generate actions, execute them via a harness, and report results back to the model, requiring four stages to succeed: parsing the model's reply, executing the action, reporting the result, and providing necessary context. Failures were found at these stages rather than in the model's reasoning, with specific issues including coordinate-space mismatches, silent failures on native dropdowns and iframes, and 4.9% of episodes contaminated by self-generated chat-template tokens. The official hidden-set score increased from 31.0 to 57.0 due to harness changes and run-to-run variance on live sites.

Key points

• WebFovea achieved a final score of 57.0 out of 100 in the WebRetriever Challenge 2026, placing second overall.
• The system requires four stages to succeed: parsing the model's reply, executing the action, reporting the result, and providing necessary context.

Context

The WebRetriever Challenge evaluates agents end-to-end on Protocol III of the WebRetriever benchmark, requiring them to operate a live website's interface and return a verifiable answer. This challenge highlights the critical need for multimodal large language models (LLMs) to bridge the gap between model decisions and browser actions, as native UI interactions often fail silently or require precise coordinate alignment.

Limitations

Evidence is limited to the WebRetriever Challenge 2026 results and does not include performance on other platforms or long-term stability. The official hidden-set score increased from 31.0 to 57.0 due to harness changes and run-to-run variance on live sites, indicating that results may vary across different environments and configurations.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 08:13:45 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeig5v5zbwlz56cuc2hqhd4t2t52jpwrlxpbeuzeztqceweficdzosm