VisSearch-Bench: Evaluating Long Horizon Visual Search in Multimodal Agents
The best of 16 multimodal models fully solved just 26.47% of the visual search tasks, even with search tools.
VisSearch-Bench contains 204 human-written tasks across seven main domains and 25 subdomains. The tasks require agents to connect visual details, object appearances and spatial relations with external information across search steps. Each includes a reference answer, traceable evidence and scoring rules. Separate accuracy and process scores distinguish complete task success from correct intermediate results.
HF Daily Papers' note
VisSearch-Bench contains 204 human-written tasks across seven main domains and 25 subdomains. The tasks require agents to connect visual details, object appearances and spatial relations with external information across search steps. Each includes a reference answer, traceable evidence and scoring rules. Separate accuracy and process scores distinguish complete task success from correct intermediate results.
HF Daily Papers' note
score 4