CMI-RewardBench - CMI-Pref Alignment (Text and Lyrics): leaderboard
Metric: Pairwise accuracy (%) against the human instruction-alignment preference on the 250 CMI-Pref test pairs whose instruction is text alone or text plus lyrics (the mean of the two 125-pair subsets); the model is given the two clips (with the instruction) and must select the preferred one; the zero-shot audio LLMs are prompted to choose, the reward models compare their scalar scores; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 67.2 | #145 |
| 2 | Gemini 3 Pro | 64 | #77 |
| 3 | Gemini 2.5 Flash | 60.8 | #237 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=cmi-rewardbench-cmi-pref-alignment-text-and-lyrics · How It Works · Data refreshed daily, snapshot 2026-10-11.