MUN-lang (5-Shot): leaderboard
Abductive commonsense reasoning from the MUN-lang task of Multimodal UNcommonsense (Son et al., 2026): given an ordinary-looking image and an unusual outcome, the model explains the hidden circumstances behind it. Win rate in percent against human-written explanations refined by GPT-4o, judged pairwise by GPT-4o in the AlpacaEval style, with five randomly chosen in-context examples; higher is better.
Metric: Win rate (%; share of MUN-lang items where a GPT-4o judge, AlpacaEval-style, prefers the model's explanation (explain how an ordinary-looking image leads to an unusual outcome) over the human-written, GPT-4o-refined reference; five randomly chosen in-context examples). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phi-4 Multimodal Instruct | 65.1 | #3914 |
| 2 | Gemma 3 4B (IT) | 49.8 | #3848 |
| 3 | Qwen 2.5 VL 7B Instruct | 41 | #1842 |
| 4 | Qwen 2 VL 7B Instruct | 36.5 | #2569 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mun-lang-5-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.