C3-Bench - Reversibility: leaderboard

Metric: Reversibility (%): share of C3-Bench image pairs for which a GPT-5.2 judge finds that the model's before-to-after and after-to-before captions describe the same changes in opposite directions; mean of three judge runs; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 26 models tracked.

Top models

#ModelScore
1Gemini 3 Pro73
2GPT-5.163
3GPT-5.262
4Gemini 3 Flash62
5Gemini 2.5 Pro52
6GPT-4o49
7Gemini 2.5 Flash48
8Qwen 3 VL 32B Instruct47
9InternVL3-38B46
10GPT-5 Mini45
11InternVL3-78B45
12Llama 4 Scout Instruct44
13Qwen 3 VL 8B Instruct40
14Qwen 2.5 VL 32B Instruct40
15Qwen 2.5 VL 72B Instruct38

Interactive version: theaggregate.ai/benchmark?slug=c3-bench-reversibility · How It Works · Data refreshed daily, snapshot 2026-09-29.