List your company
ResearchFormerly WizWand

First-error detection on ProcessBench

68.7AccuracyTeacher
22.62834.58946.5558.511May 13, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Teacherdeployment=non-deployable2026.0568.7
TL-Entropy2026.0546.3
LLM-Checkmechanism=attention2026.0543.8
TL-Perplexity2026.0543.2
Studentdeployment=deployable2026.0534.4
Linear Probe2026.0524.4