RGDT-Bench: Benchmarking LLM Reasoning for Rule-Governed Decisions and Their Justifications
A simple reward model with warrant supervision is trained, training a simple reward model that outperforms both outcome-supervised baselines across nearly all response-selection comparisons, supporting RGDT-Bench's warrant supervision for RGDT reasoning.
Jian-Peng Zhao, Hai-Hua Xu, Hao-Yan Zhang et al.
· 0 citations