Investment Rating - The industry rating is "Overweight" indicating that the industry stock index is expected to outperform the benchmark [22] Core Insights - The introduction of the Self-Principle Critique Tuning (SPCT) method by the DeepSeek team enhances the efficiency and performance of generalist reward modeling during the inference phase, suggesting a new scaling method for inference [2][3] - The DeepSeek GRM model, with 27 billion parameters, achieves performance comparable to the existing R1 model with 671 billion parameters, indicating significant advancements in model efficiency [4] - The SPCT method improves model generation quality and scalability, outperforming existing models in benchmark tests, and demonstrates that inference phase scaling strategies are more advantageous than merely increasing model parameters during training [4][5] - The GRM model reduces hardware requirements significantly, with training costs being only one-sixth of the R1 model, and inference energy consumption at approximately 17% of the R1 model, making it favorable for edge deployment [5] - The upcoming release of the DeepSeek R2 model is anticipated within 1-2 months, with the GRM model serving as a precursor to further algorithmic innovations [6] Summary by Sections Inference Scaling - SPCT enhances the adaptability and scalability of models during inference, addressing challenges in obtaining accurate reward signals in general domains [3] - The new method provides insights for further iterations of large model algorithms [3] Model Performance - DeepSeek GRM-27B outperforms existing models, achieving results comparable to R1 and GPT-4o, while utilizing a dual-loop structure for real-time evaluation and correction [4] - The research indicates that new exploration in the inference phase can expand model boundaries despite a slowdown in scaling laws during pre-training [4] Hardware Efficiency - The GRM model's hardware requirements are significantly lower, allowing for potential deployment on consumer-grade GPUs, thus expanding the performance-cost boundary [5] Future Developments - The anticipated release of the DeepSeek R2 model is expected to bring further algorithmic innovations, with a focus on optimizing training and inference efficiency [6]