Mathematical Reasoning on GSM8K (Drop in Utility)
36.7Drop in UtilityKL
Evaluation Results
| Method | Links | |
|---|---|---|
| KLAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | 36.7 | |
| Latent Similarity- Layer 12Alignment Strategy=Deliberative Alignment (GRPO)2026.04 | 5.3 | |
| RandomAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | 1.9 | |
| PerplexityAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -6.2 | |
| Self CertaintyAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -8.8 | |
| Latent Similarity- Best LayerAlignment Strategy=Deliberative Alignment (GRPO)2026.04 | -12.7 |