List your company

Step-level Correctness Prediction on ToolBench

0.8129AUROCPAIR
0.5865960.6453480.70410.762852May 18, 2026

Evaluation results

MethodLinks
PAIRProbing Category=Prefix-Aware Internal Reward2026.050.812922.36
Multi-layerProbing Category=Hidden state probes2026.050.782124.76
Multi-AttnProbing Category=Attention probes2026.050.776519.61
Last-tokenProbing Category=Hidden state probes2026.050.765425.62
Hidden + AttnProbing Category=Attention probes2026.050.759826.83
Head EntropyProbing Category=Unsupervised internal signal methods2026.050.756123.29
Mean-PooledProbing Category=Hidden state probes2026.050.737228.71
Lookback LensProbing Category=Unsupervised internal signal methods2026.050.729516.85
AttentionProbing Category=Attention probes2026.050.684719.98
CoE-CProbing Category=Unsupervised internal signal methods2026.050.5953—