Step-level Correctness Prediction on ToolBench
0.8129AUROCPAIR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PAIRProbing Category=Prefix-Aware Internal Reward2026.05 | 0.8129 | 22.36 | |
| Multi-layerProbing Category=Hidden state probes2026.05 | 0.7821 | 24.76 | |
| Multi-AttnProbing Category=Attention probes2026.05 | 0.7765 | 19.61 | |
| Last-tokenProbing Category=Hidden state probes2026.05 | 0.7654 | 25.62 | |
| Hidden + AttnProbing Category=Attention probes2026.05 | 0.7598 | 26.83 | |
| Head EntropyProbing Category=Unsupervised internal signal methods2026.05 | 0.7561 | 23.29 | |
| Mean-PooledProbing Category=Hidden state probes2026.05 | 0.7372 | 28.71 | |
| Lookback LensProbing Category=Unsupervised internal signal methods2026.05 | 0.7295 | 16.85 | |
| AttentionProbing Category=Attention probes2026.05 | 0.6847 | 19.98 | |
| CoE-CProbing Category=Unsupervised internal signal methods2026.05 | 0.5953 | — |