Research and Analysis of Latency Diagnosis and Structured Prompt Optimization for High-Difficulty Mathematical Reasoning Based on DeepSeek-R1

In view of the extremely high reasoning delay caused by “excessive thinking” in large models such as DeepSeek-R1, this paper built a pollution-free data set based on the real math problems of the postgraduate entrance examination, and innovatively introduced the “Time-Token” two-dimensional evaluation system. The experiment shows that using the “Analysis-Derivation-Conclusion” three-stage structured prompt and single sample guidance, the model significantly reduces the invalid verification while maintaining 100% accuracy, and the time consumption of complex problems drops from 408 seconds to 55 seconds (86.6% faster). In addition, token analysis objectively reveals the computational cost of “Token Inflation” inevitably caused by the typesetting of high-dimensional matrix LaTeX. The research confirms that external structural constraints can effectively balance the logical rigor and computational cost of large-scale models. In the future, this paper will further explore the dynamic prompt routing mechanism to achieve a dynamic balance between the computational efficiency of the large model and divergent innovation.