EVENT DOSSIER
In response to the dilemma faced by large language models in complex reasoning tasks between System 1 (high efficiency but low precision) and System 2 (high precision but high computational cost), the latest review article introduces the concept of “reasoning economy”. The study comprehensively analyzes the causes of efficiency bottlenecks in the post-training phase and the reasoning phase during testing, and examines the behavioral characteristics of different reasoning modes and potential optimization solutions, aiming to provide actionable insights for improving the reasoning economy of large language models. Additionally, the authors have established a public repository to continuously track the latest developments in this field.
该综述文章针对大语言模型（LLMs）在复杂推理任务中面临系统 1 效率高但性能低、系统 2 精度高但计算成本高的权衡问题，提出了“推理经济”概念。文章全面分析了后训练阶段和测试时推理阶段的推理不效率成因、不同推理模式的行为特征及潜在解决方案，旨在提供可操作的见解以优化大语言模型的推理经济性。此外，作者建立了公共仓库持续追踪该领域的最新发展。