VisSearch-Bench - ReAct - Accuracy: leaderboard

VisSearch-Bench (Tsinghua University and Ant Group, arXiv 2611.04894v1) gives multimodal models 204 human-written open-web visual search tasks in five search families. Unlike BrowseComp-VL or MM-BrowseComp, where an image mostly seeds a text search, its tasks require keeping several images, object identities and external facts linked over long searches: finding every item of a visually defined set, chaining dependent visual hops and telling near-identical candidates apart. Systems answer the English questions once per task (Pass@1); temperature 0 except GPT-5.5 and Claude Opus 4.8 (provider default); reasoning effort is not reported. Question access: the 204 scored tasks, input images, reference answers and scoring prompts are public on Hugging Face (antins-labs/VisSearch-Bench-v1). ReAct agent framework with Serper web and image search, SerpApi reverse-image search, a webpage reader and an image-crop tool; up to 30 search rounds and 60 evidence-tool calls, 20-image budget. Accuracy credits a task only when the whole final answer is correct and complete (%).

Metric: Complete-answer accuracy (%, 204 tasks). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=vissearch-bench-react-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.