Reasoning on GPQA D (Accuracy, Loss)
38.72AccuracyTIES
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TIESBackbone=Qwen2.5-7B-Base2026.06 | 38.72 | — | |
| ReasonerBackbone=Qwen2.5-7B-Base2026.06 | 35.35 | — | |
| TSV-MergeBackbone=Qwen2.5-7B-Base2026.06 | 35.02 | — | |
| RAMBackbone=Qwen2.5-7B-Base2026.06 | 34.51 | — | |
| RESMERGEBackbone=Qwen2.5-7B-Base, configuration=SRC-A+LHC2026.06 | 34.34 | — | |
| TABackbone=Qwen2.5-7B-Base2026.06 | 33.67 | — | |
| DARE + TIESBackbone=Qwen2.5-7B-Base2026.06 | 32.15 | — | |
| ISO-CTSBackbone=Qwen2.5-7B-Base2026.06 | 31.48 | — | |
| SimpleRLBackbone=Qwen2.5-7B-Base2026.06 | 30.81 | — | |
| ISO-CBackbone=Qwen2.5-7B-Base2026.06 | 30.47 | — | |
| ZeroBackbone=Qwen2.5-7B-Base2026.06 | 30.3 | — | |
| Qwen2.5-7BBackbone=Qwen2.5-7B-Base2026.06 | 30.13 | — | |
| MuonOptimizer=Muon, Model Size=3B2026.04 | 24.2 | 1.874 | |
| Adam+NexusOptimizer=Adam+Nexus, Model Size=3B2026.04 | 23.4 | 1.881 | |
| AdamWOptimizer=AdamW, Model Size=3B2026.04 | 22.6 | 1.91 |