# Token-Budget-Aware LLM Reasoning

**Authors:** Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, Zhenyu Chen  
**arXiv:** [2412.18547](https://arxiv.org/abs/2412.18547)  
**Date:** 2024-12-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.18547v4/

## Key points

- Authors: Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, Zhenyu Chen
- Enhances performance across tasks
- Effectiveness of Chain-of-Thought (CoT) reasoning
- Drawback: Increased token usage and costs in reasoning process
- Incorporate a reasonable token budget within the prompt
- Pivotal choice of token budget for compression strategy effectiveness
- Dynamically estimates token budgets based on problem complexity
- Guides reasoning process effectively with estimated budgets
- Successful reduction of token costs in CoT reasoning with marginal impact on performance
- Practical solution for optimizing LLM reasoning tasks
- Valuable insights for optimizing reasoning processes in language models
- For more details and access to their code repository, visit https://github.com/GeniusHTX/TALE.

## Summary

In their paper titled "Token-Budget-Aware LLM Reasoning," authors Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen address the critical role of reasoning in enhancing the performance of large language models (LLMs) across various tasks. They highlight the effectiveness of methods like Chain-of-Thought (CoT) reasoning in breaking down complex problems into manageable intermediate steps. However, they also identify a significant drawback in the form of increased token usage and associated costs. The researchers observe that the current reasoning process employed by LLMs is often unnecessarily lengthy and propose a solution to compress it by incorporating a reasonable token budget within the prompt. They emphasize that the choice of token budget is pivotal in determining the actual effectiveness of this compression strategy. To address this challenge, they introduce a novel token-budget-aware LLM reasoning framework. This framework dynamically estimates token budgets based on the complexity of each problem and utilizes these estimates to guide the reasoning process effectively. Through experiments conducted as part of their study, the authors demonstrate that their proposed method successfully reduces token costs in CoT reasoning while only marginally impacting performance. This approach offers a practical solution for striking a balance between efficiency and accuracy in LLM reasoning tasks. The research provides valuable insights into optimizing reasoning processes within language models and presents a promising avenue for future developments in this field. For more details and access to their code repository, interested readers can refer to https://github.com/GeniusHTX/TALE.