Scientific Reasoning on GPQA
79.2Pass@1CoNL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| CoNLBackbone=Qwen3-8B, Training Required=true2026.01 | 79.2 | — | |
| Phi-4-mini + Mistral3-3BTraining Framework=CORE, Model Configuration=Team2026.01 | 70.18 | 77.34 | |
| Phi-4-mini-reasoningTraining Framework=CORE, Model Configuration=Single2026.01 | 67.76 | 75.26 | |
| Mistral3-3B-ReasoningTraining Framework=CORE, Model Configuration=Single2026.01 | 65.58 | 74.82 | |
| CoNLBackbone=Qwen3-4B-Instruct, Training Required=true2026.01 | 55.2 | — | |
| Phi-4-mini + Mistral3-3B + OracleTraining Framework=SD-E2, Model Configuration=Oracle2026.01 | 48.12 | 55.26 | |
| Phi-4-mini-reasoningTraining Framework=SD-E2, Model Configuration=Single2026.01 | 46.35 | 50.23 | |
| Mistral3-3B-ReasoningTraining Framework=SD-E2, Model Configuration=Single2026.01 | 42.13 | 44.38 | |
| CoNLBackbone=Llama-3.1-8B, Training Required=true2026.01 | 34 | — | |
| CoNLBackbone=Llama-3.2-3B, Training Required=true2026.01 | 30.5 | — | |
| Qwen2.5-3B + Qwen3-4B (Team)Paradigm=CORE2026.01 | 16.25 | 17.8 | |
| Qwen2.5-3B-InstructParadigm=CORE2026.01 | 11.9 | 13.05 | |
| Phi-4-mini + Mistral3-3B + OracleTraining Framework=Base, Model Configuration=Oracle2026.01 | 11.86 | 14.03 | |
| Qwen3-4B-InstructParadigm=CORE2026.01 | 10.75 | 12.1 | |
| Phi-4-mini-reasoningTraining Framework=Base, Model Configuration=Single2026.01 | 10.35 | 12.23 | |
| Qwen2.5-3B + Qwen3-4B + OracleParadigm=SD-E22026.01 | 10.2 | 11.95 | |
| Mistral3-3B-ReasoningTraining Framework=Base, Model Configuration=Single2026.01 | 9.14 | 11.2 | |
| Qwen3-4B-InstructParadigm=SD-E22026.01 | 8.55 | 10.1 | |
| Qwen2.5-3B-InstructParadigm=SD-E22026.01 | 8.1 | 9.7 | |
| Qwen2.5-3B + Qwen3-4B + OracleParadigm=Base2026.01 | 5.35 | 7.1 | |
| Qwen3-4B-InstructParadigm=Base2026.01 | 4.55 | 6.25 | |
| Qwen2.5-3B-InstructParadigm=Base2026.01 | 4.2 | 5.9 |