PC-DrDPO - Avg@32 - M500: leaderboard
PC-Dr.DPO v1 Table1. Test-time adaptation accuracy (%) under Avg@32 on AIME2024/AIME2025/AMC/MATH500/GPQA-Diamond/MMLU-Pro Math (source labels A24/A25/AMC/M500/GPQA-D/MMLU-M). Each backbone and Base/RENT/TTRL/Co-Reward/PC-Dr.DPO method stays a separate run; online-adapted states are not released checkpoint aliases. Star denotes training collapse and is missing, never zero. Question metrics and source-native method settings preserved.
Metric: Accuracy / success (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen2.5-7B-Instruct (PC-DrDPO Run: method=PC-Dr.DPO;evaluation=Avg@32;effort=unreported) | 80.3 | |
| 2 | Qwen2.5-7B-Instruct (PC-DrDPO Run: method=TTRL;evaluation=Avg@32;effort=unreported) | 79.5 | |
| 3 | Qwen2.5-Math-1.5B-Instruct (PC-DrDPO Run: method=PC-Dr.DPO;evaluation=Avg@32;effort=unreported) | 78.2 | |
| 4 | Qwen2.5-Math-1.5B-Instruct (PC-DrDPO Run: method=TTRL;evaluation=Avg@32;effort=unreported) | 77.9 | |
| 5 | Qwen2.5-Math-1.5B-Instruct (PC-DrDPO Run: method=Co-Reward;evaluation=Avg@32;effort=unreported) | 76.9 | |
| 6 | Qwen2.5-7B-Instruct (PC-DrDPO Run: method=Co-Reward;evaluation=Avg@32;effort=unreported) | 75.2 | |
| 7 | Qwen2.5-7B-Instruct (PC-DrDPO Run: method=Base;evaluation=Avg@32;effort=unreported) | 74 | |
| 8 | Qwen2.5-3B-Instruct (PC-DrDPO Run: method=PC-Dr.DPO;evaluation=Avg@32;effort=unreported) | 73.3 | |
| 9 | Qwen2.5-Math-1.5B-Instruct (PC-DrDPO Run: method=Base;evaluation=Avg@32;effort=unreported) | 73.2 | |
| 10 | Qwen2.5-3B-Instruct (PC-DrDPO Run: method=TTRL;evaluation=Avg@32;effort=unreported) | 70.9 |
Interactive version: theaggregate.ai/benchmark?slug=pc-drdpo-avg-32-m500 · How It Works · Data refreshed daily, snapshot 2026-11-12.