WorldSense (caption QA, Gemini-2.5-Pro reader): leaderboard
AVoCaDO paper Table 2, the producing source of the baseline rows later re-reported by DiaDem and EVA-Cap: WorldSense questions; the listed model writes an audiovisual caption and Gemini-2.5-Pro answers the benchmark's questions from that caption alone, told to abstain when the caption lacks the information (abstentions count as wrong); accuracy (%). Results reproduced with each benchmark's official code. InternVL3.5-8B and Qwen2.5-VL-7B receive the visual stream only (no audio); every other model receives audio and video. AVoCaDO is the paper's own 7B captioner. Metric: Accuracy (%). Higher is better. Source: https://arxiv.org/html/2510.10395v1
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around January 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 33.8 | #151 |
| 2 | Gemini 2.5 Flash | 31 | #242 |
| 3 | Qwen3 Omni 30B A3B Instruct | 12.7 | #379 |
| 4 | Qwen2.5-Omni-7B | 8.6 | #622 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=worldsense-caption-qa-gemini-2-5-pro-reader · How It Works · Data refreshed daily, snapshot 2026-10-18.