M3R-Bench - Explanation Rubric: leaderboard
M3R-Bench (Jiang et al., 2026) contains 1,000 human-annotated image-text pairs from social media and advertisements for multimodal metaphor understanding, where each model jointly predicts metaphor presence, target and source concepts and sentiment. This board scores the stage-wise explanation: a Gemini-3.1-Pro judge rates visual evidence, mapping correctness and sentiment justification against gold rationales, and a stage whose prediction is wrong scores zero. Mean rubric score on a 0-100 scale; higher is better.
Metric: Rubric score (0-100): mean of visual evidence, mapping correctness and sentiment justification, each scored by Gemini-3.1-Pro against the gold stage-wise rationale and set to zero when the associated prediction is wrong. Source: arxiv.org. Saturation forecast: Around 2029. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 51.1 | #158 |
| 2 | Kimi K2.5 | 48.8 | #252 |
| 3 | GPT-4o | 40.71 | #741 |
| 4 | GPT-5.5 | 37.76 | #48 |
| 5 | Qwen 3 VL 32B Instruct | 37.72 | #582 |
| 6 | Qwen 3 VL 8B Instruct | 37.55 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=m3r-bench-explanation-rubric · How It Works · Data refreshed daily, snapshot 2026-11-12.