DMC-CF: leaderboard

DMC-CF (2026) tests multimodal causal reasoning with yes/no counterfactual questions about real-world videos of causal events: does an outcome still hold under one or two hypothetical interventions? Accuracy in percent over all questions; higher is better.

Metric: Accuracy (%) over all DMC-CF counterfactual QA (the L1 and L2 sets pooled), yes/no counterfactual questions about 1,614 real-world causal event videos (videos given natively to Gemini and Qwen3-VL, as at most 16 frames to GPT and Claude); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)79.72#99
2Qwen 3 VL 235B A22B (Thinking)73.69#459
3DMC-CF Claude-4 (checkpoint unspecified)72.34
4GPT-5.268.9#185
5Qwen 3 VL 235B A22B Instruct61.53#532

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=dmc-cf · How It Works · Data refreshed daily, snapshot 2026-11-21.