MMedBench - Rationale (BLEU-1): leaderboard
MMedBench rationale generation, measured by BLEU-1 on a 0–100 scale over human-verified reference rationales. Open-model rows use full fine-tuning on MMedBench and are separate from the zero-shot API models.
Metric: BLEU-1 (0-100). Source: henrychur.github.io. Saturation forecast: Around March 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 | 26.01 | #818 |
| 2 | Gemini 1.0 Pro | 7.28 | #933 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mmedbench-rationale-bleu-1 · How It Works · Data refreshed daily, snapshot 2026-10-20.