Scientific Reasoning on GPQA (Acc, Out Tok, Delta CoT)
61.19AccuracyRoT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RoTModel=Qwen3-4B2025.09 | 61.19 | 4,773.43 | 12.3 | |
| RoTModel=Qwen3-14B2025.09 | 59.62 | 3,336.25 | 16.6 | |
| CoTModel=Qwen3-4B2025.09 | 58.21 | 5,443.16 | — | |
| CoTModel=Qwen3-14B2025.09 | 57.69 | 3,999.04 | — | |
| RoTModel=Qwen3-8B2025.09 | 57.14 | 4,165.19 | 19.4 | |
| CoTModel=Qwen3-8B2025.09 | 53.97 | 5,169.84 | — | |
| RoT+TIModel=Qwen3-4B2025.09 | 53.73 | 3,021.03 | 44.5 | |
| RoT+TIModel=Qwen3-8B2025.09 | 47.62 | 1,646.27 | 68.2 | |
| GACmode=+ Token-φ2026.05 | 43.5 | — | — | |
| GACmode=w/o φ (plain SFT)2026.05 | 42.4 | — | — | |
| RoT+TIModel=Qwen3-14B2025.09 | 42.31 | 1,954.94 | 51.1 | |
| HPT2026.05 | 40.4 | — | — | |
| LUFFY2026.05 | 40.1 | — | — | |
| KL-ctrl2026.05 | 40 | — | — | |
| CHORDmode=Token-φ, fixed sched.2026.05 | 39.6 | — | — | |
| GradNorm-ctrl2026.05 | 39.5 | — | — | |
| SRFT2026.05 | 39.2 | — | — | |
| Nash-MTL2026.05 | 39.1 | — | — | |
| CAGrad2026.05 | 38.8 | — | — | |
| MGDA2026.05 | 38.4 | — | — | |
| SFT-best + RL2026.05 | 37.2 | — | — | |
| SFT-best2026.05 | 34.8 | — | — | |
| Qwen2.5-7B-Inst.2026.05 | 31.5 | — | — | |
| CoTModel=Qwen3-1.7B2025.09 | 30.23 | 5,615.28 | — | |
| RoT+TIModel=Qwen3-1.7B2025.09 | 30.23 | 1,448.26 | 74.2 | |
| RoTModel=Qwen3-0.6B2025.09 | 29.41 | 4,158.04 | 2.9 | |
| CoTModel=Qwen3-0.6B2025.09 | 27.94 | 4,039.28 | — | |
| RoTModel=Qwen3-1.7B2025.09 | 25.58 | 5,496.93 | 2.1 | |
| RoT+TIModel=Qwen3-0.6B2025.09 | 20.59 | 797.34 | 80.3 |