Truncated CoT Answering on BBEH
0.665AOCFaithfulness Only
Evaluation Results
| Method | Links | |
|---|---|---|
| Faithfulness OnlyBackbone=Qwen3-14B2026.02 | 0.665 | |
| MAT-SteerBackbone=Qwen3-14B2026.02 | 0.644 | |
| REMULBackbone=Qwen3-14B2026.02 | 0.626 | |
| OriginalBackbone=Qwen3-14B2026.02 | 0.58 | |
| Balanced RewardsBackbone=Qwen3-14B2026.02 | 0.574 | |
| Correctness OnlyBackbone=Qwen3-14B2026.02 | 0.518 | |
| Hint OptimizedBackbone=Qwen3-14B2026.02 | 0.503 |