Towards Better Chain-of-Thought: A Reflection on Effectiveness and Faithfulness

Jiachun Li, Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao, Jiang, Xiaojian, Liu, Kang, Zhao, Jun

Abstract

Chain-of-thought (CoT) prompting demonstrates varying performance under different reasoning tasks. Previous work attempts to evaluate it but falls short in providing an in-depth analysis of patterns that influence the CoT. In this paper, we study the CoT performance from the perspective of effectiveness and faithfulness. For the former, we identify key factors that influence CoT effectiveness on performance improvement, including problem difficulty, information gain, and information flow. For the latter, we interpret the unfaithful CoT issue by conducting a joint analysis of the information interaction among the question, CoT, and answer. The result demonstrates that, when the LLM predicts answers, it can recall correct information missing in the CoT from the question, leading to the problem. Finally, we propose a novel algorithm to mitigate this issue, in which we recall extra information from the question to enhance the CoT generation and evaluate CoTs based on their information gain. Extensive experiments demonstrate that our approach enhances both the faithfulness and effectiveness of CoT.

Introduction

Chain-of-Thought (CoT) prompting has emerged as a powerful technique to enhance the reasoning capabilities of Large Language Models (LLMs). By encouraging models to generate intermediate reasoning steps, CoT often leads to superior performance on complex tasks. However, its effectiveness is inconsistent across different domains, and concerns regarding the "faithfulness" of these reasoning chains—whether the model actually relies on its generated logic—persist. This paper investigates these two dimensions: effectiveness and faithfulness.

Analyzing CoT Effectiveness

We identify three primary factors that dictate whether CoT will improve model performance: problem difficulty, information gain (IG), and information flow.

The Role of Problem Difficulty

CoT is most beneficial for "hard" problems where direct answering is likely to fail. On simpler tasks, the overhead of generating a reasoning chain often introduces noise, potentially degrading accuracy.

CoT effectiveness vs. problem difficulty in GSM8k and WinoGrande.

Information Gain (IG) and Domain Variance

We define Information Gain (IG) as a measure of the "new" information a CoT provides relative to the original question. Our analysis reveals a stark contrast between mathematical and commonsense reasoning. In mathematical tasks, IG is typically low because the reasoning must strictly follow from the prompt's constraints. Conversely, in commonsense tasks, IG is high as models often draw from internal knowledge, making the CoT redundant or even distracting.

Information Gain (IG) across various reasoning datasets.

The Faithfulness Problem: The "Backdoor" Effect

A critical issue in CoT is unfaithfulness, where a model provides a correct answer despite a flawed reasoning chain. We discovered that this occurs because models can "recall" correct information directly from the question context during the final answer generation, effectively bypassing their own incorrect intermediate steps. This "backdoor" mechanism suggests that the model's internal attention can override the explicit reasoning path.

The QUIRE Algorithm

To address these issues, we propose QUIRE (QUestion Information Recall and Enhancement). QUIRE operates in three stages:

  1. Recall: Using Integrated Gradient Attribution (IGA), we identify the specific parts of the question the model relied on for its initial answer.
  2. Enhancement: These recalled snippets are injected back into the prompt as hints to guide the generation of a more faithful CoT.
  3. IG-Weighted Voting: Instead of standard majority voting, we weight candidate answers based on the Information Gain of their respective reasoning chains, prioritizing chains that provide meaningful, non-redundant logic.

Results

Extensive evaluations using models like Llama3.1-8B and Gemma2-9B across nine datasets demonstrate that QUIRE significantly improves both accuracy and faithfulness. For instance, we observed an accuracy boost of up to 2.4% on ProofWriter and a 5.6% improvement in faithfulness (measured via Faithful BertScore) on ProntoQA.

Conclusion

Our work provides a deeper understanding of why CoT succeeds or fails. By identifying the "backdoor" effect and the influence of problem difficulty and information gain, we developed the QUIRE algorithm to foster more reliable and effective reasoning in LLMs. Future work will focus on extending these attribution-based methods to black-box models.

Summary updated

Abstract

Chain-of-thought (CoT) prompting demonstrates varying performance under different reasoning tasks. Previous work attempts to evaluate it but falls short in providing an in-depth analysis of patterns that influence the CoT. In this paper, we study the CoT performance from the perspective of effectiveness and faithfulness. For the former, we identify key factors that influence CoT effectiveness on performance improvement, including problem difficulty, information gain, and information flow. For the latter, we interpret the unfaithful CoT issue by conducting a joint analysis of the information interaction among the question, CoT, and answer. The result demonstrates that, when the LLM predicts answers, it can recall correct information missing in the CoT from the question, leading to the problem. Finally, we propose a novel algorithm to mitigate this issue, in which we recall extra information from the question to enhance the CoT generation and evaluate CoTs based on their information gain. Extensive experiments demonstrate that our approach enhances both the faithfulness and effectiveness of CoT.
BibTeX
@inproceedings{li2025towards,
  author = {Jiachun Li and Pengfei Cao and Yubo Chen and Kang Liu and Jun Zhao and Jiang, Xiaojian and Liu, Kang and Zhao, Jun},
  title = {{Towards Better Chain-of-Thought: A Reflection on Effectiveness and Faithfulness}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure