EMMA - CoT: leaderboard
Multimodal models solve math, physics, chemistry and coding problems that need reasoning over the image (items taken from older benchmarks were dropped when text-only models given image captions solved them; 1,796 were written new with domain experts); the score is accuracy on full EMMA (2,788 questions) with chain-of-thought prompting, run by the maintainers, as the maintainers' leaderboard prints it (the paper's rows plus one later run whose released result file matches the printed score). Unlike MathVista or MMMU, it filters out items a caption makes solvable. One response per question, scored by rule; no judge model.
Metric: Accuracy (%). Source: emma-benchmark.github.io. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 33.72 | #755 |
| 2 | Gemini 2.0 Flash (Preview) | 29.12 | #742 |
| 3 | Qwen 2 VL 72B Instruct | 27.12 | #1382 |
| 4 | InternVL2.5-78B | 27.08 | #1036 |
| 5 | LLaVA-OneVision-72B | 23.82 | #1776 |
| 6 | InternVL2-Llama3-76B | 23.35 | #2021 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=emma-cot · How It Works · Data refreshed daily, snapshot 2026-11-17.