Mathematical Reasoning on AIME24 (ACC, A.Tok, ε³)
50ACCManifold Steering
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Manifold SteeringBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 50 | 9,457 | 26 | |
| R1-VeriThinkerBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 46.9 | 11,285 | 19 | |
| VPG-EABackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 46.7 | 8,225.07 | 27 | |
| ACPOBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 43.3 | 7,405 | 25 | |
| Original ModelBackbone=DeepSeek-R1-Distill-Llama-8B2026.05 | 33.33 | 12,977.53 | 9 |