Reinforcement Learning RL has become a widely used post-training method for LLMs, enhancing capabilities like human alignment, long-term reasoning, and adaptability. A major challenge, however, is generating accurate reward signals in broad, less structured domains, as current high-quality reward models are largely built on rule-based systems or verifiable tasks such as math and coding. In general applications, reward criteria are more diverse and subjective, lacking clear ground truths. To address this, generalist reward models (RMs) are being explored for broader applicability. However, these models must balance input flexibility and scalability during inference, particularly in producing reliable, high-quality rewards across varied…