DiffCap-Bench - Hallucination Rate: leaderboard
Metric: Hallucination rate (hallucinated difference items per 100 ground-truth differences, sample-level mean; can exceed 100) on the 1,075 image pairs of DiffCap-Bench (6,713 human-verified difference items), one fixed captioning prompt, open models at temperature 0.1 and proprietary models at their defaults, a Gemini 2.5 Pro judge matching each described difference against the key difference list; lower is better. Source: arxiv.org. Saturation forecast: Around May 2027. 16 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 2.6 |
| 2 | Grok 4.20 | 10.5 |
| 3 | GPT-5.2 | 12 |
| 4 | Kimi K2.5 (Thinking) | 18.9 |
| 5 | Qwen 3 VL 8B (Thinking) | 20.2 |
| 6 | Step3 VL 10B | 22.8 |
| 7 | Qwen 3 VL 32B (Thinking) | 25.8 |
| 8 | Qwen 3 VL 8B Instruct | 26.1 |
| 9 | Qwen 3 VL 32B Instruct | 31.3 |
| 10 | Kimi K2.5 (Non-reasoning) | 37.7 |
Interactive version: theaggregate.ai/benchmark?slug=diffcap-bench-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-10-07.