Reasoning on SuperGPQA (Accuracy, Length)
7,262LengthBase
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BaseBackbone=Qwen3-8B2026.05 | 7,262 | 44.2 | |
| BaseBackbone=Qwen3-4B2026.05 | 6,858 | 37.2 | |
| BaseBackbone=DSQW-7B2026.05 | 6,238 | 30 | |
| +GRPOBackbone=DSQW-7B2026.05 | 5,452 | 31.8 | |
| +ICRBackbone=DSQW-7B2026.05 | 4,635 | 33.2 | |
| +GRPOBackbone=Qwen3-4B2026.05 | 4,143 | 41 | |
| +GRPOBackbone=Qwen3-8B2026.05 | 4,022 | 42.6 | |
| +LC-R1Backbone=DSQW-7B2026.05 | 3,832 | 32.6 | |
| +ICRBackbone=Qwen3-8B2026.05 | 3,498 | 44 | |
| +LP-FBackbone=Qwen3-8B2026.05 | 3,392 | 41.6 | |
| +LP-FBackbone=DSQW-7B2026.05 | 3,390 | 33.4 | |
| +LP-G (Kimi)Backbone=DSQW-7B2026.05 | 3,277 | 31.8 | |
| +LP-FBackbone=Qwen3-4B2026.05 | 3,189 | 38.2 | |
| +ICRBackbone=Qwen3-4B2026.05 | 3,082 | 41.2 | |
| +LC-R1Backbone=Qwen3-8B2026.05 | 3,079 | 40.4 | |
| +ShorterBetterBackbone=DSQW-7B2026.05 | 2,967 | 31.6 | |
| +ICR /w LP-FBackbone=DSQW-7B2026.05 | 2,957 | 35.2 | |
| +LC-R1Backbone=Qwen3-4B2026.05 | 2,884 | 36.2 | |
| +ICR /w LP-FBackbone=Qwen3-4B2026.05 | 2,392 | 39.6 | |
| +ICR /w LP-FBackbone=Qwen3-8B2026.05 | 2,312 | 41.6 | |
| +LP-G (Kimi)Backbone=Qwen3-8B2026.05 | 2,280 | 39.4 | |
| +ShorterBetterBackbone=Qwen3-8B2026.05 | 2,269 | 40.8 | |
| +ShorterBetterBackbone=Qwen3-4B2026.05 | 2,009 | 34.8 | |
| +LP-G (Kimi)Backbone=Qwen3-4B2026.05 | 1,300 | 37.2 |