General Science Reasoning on GPQA Diamond (Accuracy, Length)
57.68AccuracyDDPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DDPOBackbone=Qwen3-14B2026.03 | 57.68 | 5,683 | |
| L1Backbone=Qwen3-14B2026.03 | 57.26 | 5,543 | |
| GRPOBackbone=Qwen3-14B2026.03 | 57.11 | 5,728 | |
| SHORTER BETTERBackbone=Qwen3-14B2026.03 | 56.88 | 4,416 | |
| GRPOBackbone=Qwen3-8B2026.03 | 56.76 | 7,379 | |
| A-DLPBackbone=Qwen3-14B2026.03 | 56.41 | 4,986 | |
| DDPOBackbone=Qwen3-8B2026.03 | 56.35 | 6,363 | |
| GRPO-LEADBackbone=Qwen3-14B2026.03 | 55.01 | 7,147 | |
| DDPOBackbone=Qwen3-4B2026.03 | 51.84 | 6,252 | |
| GRPOBackbone=Qwen3-4B2026.03 | 51.02 | 7,423 |