EPIC RL - AIME 2025: leaderboard
EPIC: Qwen3-30B-A3B-Base descendants trained 300 steps on DAPO-Math-17K; each clipping method is a separate research checkpoint. Mean@8, temperature 1, top-p 0.7, maximum 16,384 response tokens, rule-based correctness. Descendants remain distinct from the original base and released instruct weights. Mean@8 Accuracy (%).
Metric: Mean@8 Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | EPIC Study Qwen3-30B-A3B EPIC | 25.28 | |
| 2 | EPIC Study Qwen3-30B-A3B DAPO | 24.03 | |
| 3 | EPIC Study Qwen3-30B-A3B RIPO | 21.81 | |
| 4 | EPIC Study Qwen3-30B-A3B CTPO | 20.97 | |
| 5 | EPIC Study Qwen3-30B-A3B GSPO | 20.56 | |
| 6 | EPIC Study Qwen3-30B-A3B GRPO | 16.81 |
Interactive version: theaggregate.ai/benchmark?slug=epic-rl-aime-2025 · How It Works · Data refreshed daily, snapshot 2026-10-23.