Logical Reasoning on BBH (Component Accuracies)
66.5Logical Deduction AccuracyGAC
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| GACvariant=Token-φ2026.05 | 66.5 | 70.9 | 59.7 | 65.7 | |
| GACvariant=plain SFT2026.05 | 65.5 | 69.9 | 58.4 | 64.6 | |
| HPT2026.05 | 63.2 | 68.1 | 56.4 | 62.6 | |
| KL-ctrl2026.05 | 63 | 67.8 | 56.2 | 62.3 | |
| LUFFY2026.05 | 62.9 | 67.8 | 56 | 62.2 | |
| CHORDvariant=Token-φ, schedule=fixed2026.05 | 62.7 | 67.5 | 55.8 | 62 | |
| Nash-MTL2026.05 | 62.1 | 67 | 55.4 | 61.5 | |
| SRFT2026.05 | 62 | 66.9 | 55.2 | 61.4 | |
| SFT-best + RL2026.05 | 59.4 | 64.8 | 52.3 | 58.8 | |
| SFT-best2026.05 | 56.8 | 62.1 | 49.6 | 56.2 | |
| Qwen2.5-7B-Inst.2026.05 | 52.4 | 58.7 | 45.2 | 52.1 |