Darwin R3 Card - Held-out SuperGPQA - Mean of 4: leaderboard
Darwin R3 author comparison, Held-out SuperGPQA, 1000 questions, 4 samples per question, 16384 thinking tokens, temperature1.0. Single sample / mean-of-4 / majority-of-4 remain separate. R3 is trained from R1 on separate SuperGPQA practice problems; source says GPQA Diamond and the held-out set were never used in training or selection. R0 is Qwen3.8-Flash-Next; R1 and R3 are distinct merged trained checkpoints, not effort modes. Exact source table and immutable card revision pinned. No inference from the card’s mixed-protocol external leaderboard comparison.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Darwin-180B-RSI-R3 (Darwin R3 Card Run: task=Held-out SuperGPQA;metric=Mean of 4;draws=4;thinking_budget=16384;temperature=1.0) | 66.7 | |
| 2 | Darwin-180B-RSI (Darwin R3 Card Run: task=Held-out SuperGPQA;metric=Mean of 4;draws=4;thinking_budget=16384;temperature=1.0) | 65.67 |
Interactive version: theaggregate.ai/benchmark?slug=darwin-r3-card-held-out-supergpqa-mean-of-4 · How It Works · Data refreshed daily, snapshot 2026-11-12.