Truncated CoT Answering on ZebraLogicBench (ZLB)
58.7AOCFaithfulness Only
Evaluation Results
| Method | Links | |
|---|---|---|
| Faithfulness OnlyBackbone=Qwen3-14B2026.02 | 58.7 | |
| REMULBackbone=Qwen3-14B2026.02 | 56.6 | |
| MAT-SteerBackbone=Qwen3-14B2026.02 | 52.7 | |
| OriginalBackbone=Qwen3-14B2026.02 | 52 | |
| Balanced RewardsBackbone=Qwen3-14B2026.02 | 51.2 | |
| Hint OptimizedBackbone=Qwen3-14B2026.02 | 48.1 | |
| Correctness OnlyBackbone=Qwen3-14B2026.02 | 47.4 |