In the February 5, 2026 paper in collaboration between Qwen Team, Alibaba Group, Fudan University, Tsinghua University, the researcher introduce **Rationale Consistency**, a new metric and framework designed to evaluate whether **Generative Reward Models (GenRMs)** reach their conclusions using human-like logic rather than superficial shortcuts. Researchers identified a **"Deceptive Alignment Trap"** where models achieve high accuracy in predicting outcomes but rely on flawed or non-human reasoning, a gap that traditional **Outcome Accuracy** fails to detect. To resolve this, the authors developed **METAJUDGE**, a system that decomposes human feedback into **atomic rationales** to perform fine-grained semantic matching against model justifications. By implementing a **hybrid reward signal** that combines outcome correctness with logical consistency, they trained a version of **Qwen3** that achieves state-of-the-art performance across multiple benchmarks. This methodology effectively reverses **rationale degeneration**, ensuring that AI judges provide evidence-grounded evaluations rather than relying on generic or style-based heuristics. Ultimately, the research demonstrates that supervising the **reasoning process** is essential for building reward models that truly align with human values during reinforcement learning. Source: February 05 2026 Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models Qwen Team, Alibaba Group, Fudan University, Tsinghua University Binghai Wang, Yantao Liu, Yuxuan Liu, Tianyi Tang, Shenzhi Wang, Chang Gao, Chujie Zheng, Yichang Zhang, Le Yu, Shixuan Liu, Tao Gui, Qi Zhang, Xuanjing Huang, Bowen Yu, Fei Huang, Junyang Lin https://arxiv.org/pdf/2602.04649
