Mathematical Reasoning on GSM8K (Det., Cov@8, Cov@16)
44.5Detection AccuracyLatent-SFT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Latent-SFTBackbone=Llama-3.2-1B, Sampling=MC-dropout2025.10 | 44.5 | 58.5 | 64.9 | |
| RoT-4BBackbone=Qwen3-VL, Sampling=MC-dropout2025.10 | 37.5 | 39.4 | 39.7 | |
| RoT-2BBackbone=Qwen3-VL, Sampling=MC-dropout2025.10 | 23.3 | 25.7 | 25.7 |