HAERAE-Vision - Science: leaderboard
The 81 science questions of HAERAE-Vision (What Users Leave Unsaid, Choi et al., ACL 2026 Findings): real Korean community questions on physics and other sciences, many of them problems shown in images, in the users' original under-specified wording. A GPT-5-mini judge scores each answer against a 1-5 item checklist (met 1, partly met 0.5, not met 0); the score is the mean checklist credit per question, in percent, averaged over three runs; higher is better.
Metric: Checklist Score (%; original under-specified queries, GPT-5-mini judge, mean of 3 runs). Source: arxiv.org. Saturation forecast: Around February 2027. 45 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 52.49 | #349 |
| 2 | Qwen 3 VL 32B (Thinking) | 51.94 | #557 |
| 3 | Qwen 3 VL 235B A22B Instruct | 51.51 | #521 |
| 4 | Qwen 3 VL 32B Instruct | 51.39 | #582 |
| 5 | Qwen 3 VL 30B A3B (Thinking) | 49.92 | #710 |
| 6 | Qwen 3 VL 235B A22B (Thinking) | 49.19 | #451 |
| 7 | Qwen 3 VL 8B (Thinking) | 42.07 | #887 |
| 8 | Gemini 2.5 Flash Lite | 41.73 | #958 |
| 9 | Grok 4 | 40.7 | #306 |
| 10 | Qwen 3 VL 30B A3B Instruct | 40.4 | #870 |
| 11 | Gemini 2.5 Pro | 39.93 | #262 |
| 12 | Gemini 2.5 Flash | 39.62 | #452 |
| 13 | Qwen 3 VL 4B (Thinking) | 38.53 | #1303 |
| 14 | GPT-5 | 36.31 | #163 |
| 15 | Gemma 3 27B | 34.93 | #1638 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=haerae-vision-science · How It Works · Data refreshed daily, snapshot 2026-11-12.