CoD强化大模型不可能三角，加速端侧应用落地

Investment Rating - The industry investment rating is "Positive" [1][39] Core Insights - The report emphasizes the potential investment opportunities brought by "Artificial Intelligence+" in the electronics industry [4] - It highlights the importance of high-end computing chips and lower-computing edge chips from a chip perspective, and suggests focusing on intelligent terminals that incorporate computing chips and AI large models from an application perspective [32] Summary by Sections 1. Chain-of-Thought (CoT) - CoT is a mechanism that enhances the performance of large language models (LLMs) in solving complex reasoning tasks by emphasizing detailed, step-by-step reasoning processes [5][6] - The concept was introduced in a paper published in January 2022, which demonstrated that incorporating reasoning steps into prompts can improve accuracy in reasoning tasks [6] 2. Chain of Draft (CoD) - CoD simplifies intermediate reasoning steps, reducing token usage and reasoning delays while maintaining accuracy comparable to CoT [8][9] - A paper published in March 2025 introduced CoD, which encourages LLMs to generate concise, information-dense outputs, thus making them more efficient for practical applications [8][9] 3. The Impossible Triangle of Large Models - The report discusses the "impossible triangle" of cost, effect, and performance in the development of large models, indicating that there are multiple paths to develop large models [30][31] - It compares the performance of GPT-4 and DeepSeek-R1, noting that DeepSeek-R1 improves reasoning speed at a lower cost, while GPT-4 achieves high accuracy at a higher cost [30][31] 4. Investment Recommendations - The report recommends focusing on high-end computing chips and lower-computing edge chips, as well as intelligent terminals that utilize these chips and AI large models [32]