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.

Slim-SC: Thought Pruning for Efficient Scaling with Self-Consistency

Created by
  • Haebom

Author

Colin Hong, Xu Guo, Anand Chaanan Singh, Esha Choukse, Dmitrii Ustiugov

Outline

This paper theoretically and experimentally analyzes the inefficiency of Self-Consistency (SC), a Test Time Extension (TTS) technique, and proposes Slim-SC, a novel method to improve it. SC generates multiple inference processes in parallel and selects a final answer through majority voting, but suffers from high computational costs. Slim-SC utilizes a stepwise pruning strategy that removes redundant chains by exploiting chain similarity during the inference phase, reducing inference latency and KVC usage by up to 45% and 26%, respectively, while maintaining or improving accuracy. This provides a simple yet efficient TTS alternative to SC.

Takeaways, Limitations

•
Takeaways:
◦
We present theoretical and experimental analyses of the inefficiencies of SCs to suggest potential improvements.
◦
We demonstrate that Slim-SC can significantly reduce the computational cost of SC while maintaining or improving accuracy.
◦
Provides a simple and efficient alternative to SC.
◦
Contributes to research on TTS techniques to improve the inference performance of LLM.
•
Limitations:
◦
The performance improvements of Slim-SC may be limited to specific datasets and LLM architectures.
◦
Further research may be needed on how to measure similarity at the thinking stage.
◦
Experiments with more diverse LLM architectures and datasets are needed.
👍
Made with Slashpage