Code Generation on LiveCodeBench (@1, @32, @128, AUC, Tokens)
75.2Rate @32 ScoreDFPO (Orth-Proj)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DFPO (Orth-Proj)Backbone=Qwen3-Next-80B-A3B-Thinking2026.04 | 75.2 | — | — | — | — | |
| DFPO (Min-Replace)Backbone=Qwen3-Next-80B-A3B-Thinking2026.04 | 75.1 | — | — | — | — | |
| GRPO-fixBackbone=Qwen3-Next-80B-A3B-Thinking2026.04 | 73.9 | — | — | — | — | |
| DFPO (Min-Replace)Backbone=Qwen3-32B2026.04 | 71.6 | — | — | — | — | |
| DFPO (Orth-Proj)Backbone=Qwen3-32B2026.04 | 71.6 | — | — | — | — | |
| GSPOBackbone=Qwen3-Next-80B-A3B-Thinking2026.04 | 71 | — | — | — | — | |
| GRPOBackbone=Qwen3-Next-80B-A3B-Thinking2026.04 | 70.9 | — | — | — | — | |
| GRPO-fixBackbone=Qwen3-32B2026.04 | 69.1 | — | — | — | — | |
| GSPOBackbone=Qwen3-32B2026.04 | 64.7 | — | — | — | — | |
| GRPOBackbone=Qwen3-32B2026.04 | 64.5 | — | — | — | — | |
| Dele-SimKOBackbone=Qwen3-8B2026.01 | 47.5 | 37.2 | 49.8 | 44.1 | 5.8 | |
| Dele-GRPOBackbone=Qwen3-8B2026.01 | 46.8 | 36.4 | 49.1 | 43.4 | 5.9 | |
| LCoT-SimKOBackbone=Qwen3-8B2026.01 | 46.5 | 36.5 | 48.8 | 43.2 | 6 | |
| TGR-TokenBackbone=Qwen3-8B2026.01 | 45.5 | 35.1 | 47.8 | 42 | 4.4 | |
| LCoT-GRPOBackbone=Qwen3-8B2026.01 | 44.5 | 35.5 | 46.8 | 41.5 | 6 | |
| PSamplingBackbone=Qwen3-8B2026.01 | 42.1 | 32.4 | 44.1 | 38.8 | 3.6 | |
| Base ModelBackbone=Qwen3-8B2026.01 | 36.8 | 26.8 | 38.9 | 33.5 | 1.6 | |
| TGR-LatentBackbone=Qwen3-8B2026.01 | — | 36.8 | 50.2 | 44.4 | 4.6 |