[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Towards Reasoning Era: A Survey of Long Chain-of-Thoughts for Reasoning Large Language Models

Created by
  • Haebom

Author

Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, Wanxiang Che

Outline

This paper provides a comprehensive investigation of the Long CoT, which plays a crucial role in improving the reasoning ability of large-scale language models (LLMs). We clarify its differences from the traditional Short CoT, and analyze the core features of Long CoT, such as deep reasoning, extensive exploration, and actionable reflection. In addition, we investigate phenomena such as overthinking and inference time expansion, and suggest future research directions such as multimodal inference integration, efficiency improvement, and enhanced knowledge framework.

Takeaways, Limitations

•
Takeaways:
◦
We clearly define the differences between Long CoT and Short CoT and propose a new classification system, thereby laying the foundation for LLM inference research.
◦
We analyzed the core features of Long CoT (deep inference, broad exploration, and actionable reflection) to identify the cause of the performance improvement.
◦
It has provided insights into phenomena such as excessive thinking and extended reasoning time.
◦
It can contribute to the advancement of the field of LLM Inference by suggesting future research directions.
•
Limitations:
◦
Although this is the first comprehensive investigation of Long CoT, there may still be a lack of consensus on the precise definition or measurement method of Long CoT.
◦
The future research directions suggested in the paper may be hypotheses that have not yet been verified.
◦
Further research is needed on generalizability across different LLM architectures and datasets.
👍
Made with Slashpage