High-priority alert based on source significance and keyword relevance.
Context
What happened: Researchers introduced VisSearch-Bench, a benchmark of 204 human-written tasks testing how multimodal agents manage visual context during long-horizon web search across five search categories.
Key numbers: 204 tasks, 7 main domains, 25 subdomains, 16 multimodal models evaluated, best system fully solves 26.47% of tasks.
Why it matters: The results highlight the challenge of linking visual evidence with search results to produce complete answers, even when agents have search tools.