TrustRL-Eval: An Interpretable Reinforcement Learning Framework for Predicting and Optimizing Large Language Model Response Quality in Multi-Step Reasoning Tasks. JDIAS 2026, 1 (1).