Graduate-level Science QA on GPQA (Accuracy)
52.02AccuracyParallel-Synthesis
Evaluation Results
| Method | Links | |
|---|---|---|
| Parallel-SynthesisBackbone=Qwen3-14B, Workers=3, Merging=Weighted averaging (lambda=0.5)2026.06 | 52.02 | |
| Text-SerializationBackbone=Qwen3-14B, Workers=32026.06 | 50 | |
| VotingBackbone=Qwen3-14B, Workers=32026.06 | 48.99 | |
| KVLINKBackbone=Qwen3-14B, Workers=32026.06 | 35.35 | |
| SingleBackbone=Qwen3-14B, Workers=12026.06 | 31.31 | |
| Qwen2.5-3B-InstructBackbone=Qwen2.5-3B, Training Protocol=Instruct2026.05 | 30.8 | |
| Qwen2.5-3B-RLVRBackbone=Qwen2.5-3B, Training Protocol=RLVR2026.05 | 29.3 | |
| Qwen2.5-3B-GRLOBackbone=Qwen2.5-3B, Training Protocol=GRLO2026.05 | 29.3 | |
| Qwen2.5-3B-GRLO+RLVRBackbone=Qwen2.5-3B, Training Protocol=GRLO+RLVR2026.05 | 27.8 | |
| Qwen2.5-3B-BaseBackbone=Qwen2.5-3B, Training Protocol=Base2026.05 | 22.2 | |
| CacheBlendBackbone=Qwen3-14B, Workers=32026.06 | 12.63 | |
| APEBackbone=Qwen3-14B, Workers=32026.06 | 10.1 | |
| Qwen2.5-3B-MathSFTBackbone=Qwen2.5-3B, Training Protocol=MathSFT2026.05 | 4.6 |