Truncated CoT Answering on MuSR
33.6AOCMAT-Steer
Evaluation Results
| Method | Links | |
|---|---|---|
| MAT-SteerBackbone=Qwen3-14B2026.02 | 33.6 | |
| OriginalBackbone=Qwen3-14B2026.02 | 33.2 | |
| Faithfulness OnlyBackbone=Qwen3-14B2026.02 | 33 | |
| Balanced RewardsBackbone=Qwen3-14B2026.02 | 32.8 | |
| REMULBackbone=Qwen3-14B2026.02 | 32.5 | |
| Correctness OnlyBackbone=Qwen3-14B2026.02 | 31.7 | |
| Hint OptimizedBackbone=Qwen3-14B2026.02 | 30.4 |