Mathematical Reasoning on GSM8K (ACC, A.Tok, ε³)
89.16AccuracyVPG-EA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| VPG-EABackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 89.16 | 677.82 | 11.73 | |
| Original ModelBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 87.87 | 877 | 8.8 | |
| Manifold SteeringBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 82.8 | 542 | 12.65 |