RVMS-Bench (RACLO) - Video Retrieval (Global): leaderboard
Metric: Retrieval recall (%) on the 160 queries giving the global impression: the target video counts as found when a retrieved URL matches it exactly or, failing that, the backbone model accepts a candidate against the global impression (64 frames and the audio) and Gemini 2.5 Pro agrees; RVMS-Bench's real-world YouTube video search queries built from fuzzy memory cues (global impression G, key moment K, temporal context T, auditory memory A; 160 queries per cue combination, 20 topics, four duration bands), each model run as the backbone of the RACLO open-web search agent (two search queries, two candidate URLs each); higher is better. Source: arxiv.org. Saturation forecast: Around September 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 45.6 | #77 |
| 2 | Gemini 2.5 Pro | 36.2 | #145 |
| 3 | Qwen 3 VL 235B A22B Instruct | 32.5 | #264 |
| 4 | GPT-5 Mini | 31.2 | #176 |
| 5 | Qwen 2.5 VL 72B Instruct | 29.4 | #364 |
| 6 | GPT-5.2 | 26.2 | #105 |
| 7 | GPT-4o | 18.8 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=rvms-bench-raclo-video-retrieval-global · How It Works · Data refreshed daily, snapshot 2026-10-11.