Step-level correctness prediction on GTA
92.17AUROCPAIR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PAIRProbing Category=Prefix-Aware Internal Reward2026.05 | 92.17 | 11.98 | |
| Multi-layerProbing Category=Hidden state probes2026.05 | 90.82 | 13.42 | |
| Multi-AttnProbing Category=Attention probes2026.05 | 90.34 | 13.56 | |
| Hidden + AttnProbing Category=Attention probes2026.05 | 89.78 | 13.78 | |
| Last-tokenProbing Category=Hidden state probes2026.05 | 89.67 | 12.94 | |
| Mean-PooledProbing Category=Hidden state probes2026.05 | 88.51 | 15.21 | |
| Head EntropyProbing Category=Unsupervised internal signal methods2026.05 | 88.21 | 14.32 | |
| Lookback LensProbing Category=Unsupervised internal signal methods2026.05 | 87.45 | 11.82 | |
| AttentionProbing Category=Attention probes2026.05 | 87.09 | 9.54 | |
| CoE-CProbing Category=Unsupervised internal signal methods2026.05 | 74.91 | — |