CapRiCorn-1K - Referential Consistency: leaderboard
Subject referential consistency of audiovisual captions on the 1,000 CapRiCorn-1K videos (Chen et al., 2026): the share of pairs of descriptions of the same subject that a GPT-4.1 judge groups as one referent, in percent. It rewards captions that keep track of who is who across a long video; higher is better.
Metric: Subject referential consistency (%; share of same-subject description pairs a GPT-4.1 judge clusters as one referent, over all pairs of that subject's keypoints, audiovisual captions of 1,000 videos). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 39.6 | #159 |
| 2 | Gemini 3.1 Pro (Preview) | 39.1 | #91 |
| 3 | Qwen3 Omni 30B A3B Instruct | 1.6 | #902 |
| 4 | Qwen2.5-Omni-7B | 0.6 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=capricorn-1k-referential-consistency · How It Works · Data refreshed daily, snapshot 2026-11-12.