To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning

Original Paper: https://arxiv.org/abs/2409.12183

By: Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, Greg Durrett

Abstract:

Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra ``thinking'' really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications.

Summary Notes

Figure: Left: meta-analysis of CoT literature; each point is a reported delta of CoT over direct answering for some (LLM, task) pair. Right: average performance of using zero-shot CoT v.s. direct answer prompts across five general reasoning categories, covering 20 datasets with 14 LLMs evaluated on each. In both sets of results, math and other kinds of symbolic reasoning are the domains that consistently see substantial improvements from CoT (red dotted line indicates the mean improvement from CoT across experiments).

Introduction

In the realm of large language models (LLMs), the Chain-of-Thought (CoT) technique has emerged as a promising method for enhancing reasoning capabilities. CoT has been lauded for its ability to provide human-readable explanations and improve language models' performance on complex tasks. But is this technique equally effective across all problem domains? A comprehensive study has delved into this question, analyzing over 100 papers and conducting experiments on 20 datasets with 14 different models. This blog post explores the findings of this research, highlighting where CoT excels, where it falls short, and its implications for future applications.

The Role of CoT in Symbolic Reasoning

Chain-of-Thought prompting involves breaking down a problem into intermediate steps, allowing models to execute a sequence of logical operations. This approach is particularly helpful in tasks that require mathematical, logical, or algorithmic reasoning. The study reveals that CoT significantly improves performance in these domains, with tasks involving symbolic reasoning showing average improvements of up to 56.9% with CoT compared to 45.5% without it.

Methodologies and Key Findings

The researchers employed a meta-analysis of existing literature and their own experiments across various models and datasets. They distinguished between zero-shot and few-shot CoT prompting, comparing these against direct answer strategies. The study found that:

1. Mathematical and Logical Tasks: CoT provides substantial benefits in tasks that can be broken down into formal, sequential steps, such as solving equations or logical puzzles. For instance, in the GSM8K dataset, which involves grade-school math problems, CoT improved accuracy by up to 66.9%.
2. Symbolic Execution: The primary advantage of CoT is its ability to execute symbolic operations effectively. However, when compared to external symbolic solvers, CoT's performance is still overshadowed, indicating that while CoT can plan and execute solutions, it does not yet match the precision of dedicated symbolic solvers.
3. Non-Symbolic Tasks: In contrast, CoT offers little to no improvement for tasks that do not inherently involve symbolic reasoning, such as commonsense reasoning or context-aware QA. The study suggests that for these tasks, CoT may not be necessary, and other prompting strategies could achieve similar results at a lower computational cost.

Implications and Future Applications

The findings suggest that CoT should be applied selectively, primarily in domains where symbolic reasoning is paramount. For tasks outside this scope, the study advocates for exploring new paradigms that leverage intermediate computations more effectively. This might include integrating search algorithms, interacting agents, or models fine-tuned with more sophisticated reasoning techniques.

Conclusion

The research underscores the importance of understanding the strengths and limitations of CoT in enhancing the reasoning capabilities of LLMs. As we continue to push the boundaries of AI, it's clear that while CoT is a powerful tool for specific tasks, it is not a one-size-fits-all solution. Future advancements in LLMs will likely require a diverse array of methods to address the varied challenges across different domains. As engineers and researchers, focusing on the nuanced applications of techniques like CoT will be crucial in developing more robust and versatile AI systems.

Discussion and Related Work

This study contributes to the ongoing discourse on the effectiveness of CoT in reasoning and planning tasks. While early work highlighted CoT's potential, this research provides a more granular understanding of where CoT truly shines. It aligns with the notion that while deliberation should theoretically enhance performance across various tasks, its practical benefits are most pronounced in domains that require explicit symbolic manipulation.

Limitations and Areas for Future Research

While offering significant insights, the study acknowledges potential limitations, such as data contamination and the focus on English-language models. Further research could explore CoT's applicability in different languages and its integration with other cutting-edge AI techniques to enhance its efficacy in non-symbolic reasoning tasks.

Acknowledgments

The study was supported by contributions from numerous researchers across institutions, highlighting the collaborative effort in advancing AI research. The findings are a testament to the intricate balance between technical innovation and practical application in the ever-evolving field of artificial intelligence.