Instance-adaptive Zero-shot Chain-of-Thought Prompting

Summary

This paper explores the inside interactions among three components(i.e., question, prompt, rationale) in zero-shot CoT reasoning through the saliency score analysis, and discover the distinct characteristics of good and bad reasoning in terms of information flow. Based on above findings, the authors propose the IAP - an instance-level adaptive prompting strategy designed to enhance CoT reasoning by selecting an appropriate prompt that can guide LLMs to reason correctly from a given set of prompts for each question. To demonstrate the effectiveness of the proposed method, the authors conducted extensive experiments on the LLaMA-3 8B, LLaMA-2 13B, and Qwen14B models across Math, Logic, and Commonsense tasks.

Strengths

* **The motivation of this paper is quite intuitive.** Through the example in Figure 1, the authors illustrate that an instance-wise zero-shot CoT prompt is more plausible for better reasoning and may achieve a cap-breaking performance compared to the task-level optimal prompt, which is easy to understand. * **Using Neuron saliency score to analyze the information flow of the question, prompt, and rationale is innovative and promising.** The authors conduct both qualitative and quantitative analyses on good-bad reasoning instances, discovering different patterns of information flow in good-bad reasoning. They further delve into a deeper analysis from the perspectives of Layer and Head, yielding valuable conclusions. * **Extensive experiments are conducted to validate the effectiveness of proposed method.** The authors test the proposed IAP and comparison methods on LLaMA-3 8B, LLaMA-2 13B, and Qwen14B across Math, Logic, and Commonsense reasoning tasks.

Weaknesses

* **The proposed method is not clearly explained, some details are missing.** For instance, in Sequential Substitution (IAP-ss), it is not specified what the corresponding threshold is or how it was obtained. Similarly, in Majority Vote (IAP-mv), there is no explanation provided for how many top maximum scores are preserved. The lack of details in the proposed methods can be quite confusing. * **The experimental setup is unfair.** For the main "effective" method proposed, IAP-mv, which obtains the final results through Majority Vote (similar to self consistency), it is unreasonable to compare it with baselines that only perform inference once. The improvement brought about by this method could very well be the result of ensemble inference from multiple reasoning paths [1][2][3], rather than the strategy implemented by the author's previous good-bad reasoning findings. * **Some experimental conclusions are missing or somewhat unreasonable.** For instance, in the experiment of Consistency & Complementary, the authors only provide the experimental setup and show the corresponding results in Table 2 without giving any conclusions or analysis. Furthermore, in the analysis experiment for Efficacy, the author demonstrated through Figure 6 that while IAP-ss also incurs additional overhead, its accuracy compared to 0-shot also declines to some extent. Therefore, the conclusion stated by the authors that "the two IAP strategies can be employed as trade-offs in different demand prioritization applications" is not in line with the facts, as IAP-ss does not provide any gain compared to the baseline 0-shot. [1] Making Large Language Models Better Reasoners with Step-Aware Verifier [2] Diversity of Thought Improves Reasoning Abilities of LLMs [3] Answering Questions by Meta-Reasoning over Multiple Chains of Thought

Questions

See weeknesses. Although this paper performs quite well in the section of Information Flow Analysis on Zero-shot CoT, there are numerous flaws in its experiments that fail to convince me. If the authors can address my concerns, I would consider raising my score.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

I think the authors have addressed their limitations.

Summary

The paper introduces an instance-adaptive prompting algorithm for zero-shot Chain-of-Thought (CoT) reasoning in large language models (LLMs). Traditional task-level prompts are insufficient for all instances, so the authors propose a strategy that differentiates good and bad prompts based on information flow from the question to the prompt and rationale. Using neuron saliency score analysis, the study reveals that successful reasoning requires prompts to aggregate semantic information from the question. The proposed instance-adaptive prompting strategy (IAP) demonstrates consistent improvements across multiple reasoning tasks with LLaMA-2, LLaMA-3, and Qwen models.

Strengths

- **Comprehensive Analysis**: Uses neuron saliency scores to understand information flow during reasoning. - **Innovative Approach**: Tailors prompts to individual instances, improving upon uniform task-level prompts.

Weaknesses

- **Limited Scope of Neuron Saliency Score Analysis**: The neuron saliency score analysis in Section 2 is conducted on only one LLM (not state in paper, personal guess) and one dataset, GSM8K. More evidence and diverse datasets are needed to support this analysis comprehensively. Additionally, Section 2.3, "Head Analysis," lacks a comparative study between good and bad reasoning instances. - **Insufficient Experimental Details**: The experimental details provided are not sufficiently clear. Based on lines 263-264, my understanding is that IAP uses 9 different zero-shot CoT prompts to compute the S score. For IAP-ss, the process stops and uses the current result upon encountering the first prompt that meets the threshold. For IAP-mv, results from all 9 prompts are saved, and the top K are used for voting. This raises several questions: 1. How does the performance of IAP-mv compare to directly using a majority vote across the 9 prompt results? 2. What is the distribution of results for each method? For instance, which prompts tend to have higher S scores? This analysis could reveal conclusions such as "Let’s think step by step" being more suitable for math questions, while "Don’t think. Just feel." may be better for MMLU, aligning more closely with the instance-wise topic. - **Figure 1 Inconsistency**: Figure 1 does not align with the rest of the paper. The figure and its description suggest addressing an overly complex problem, which is not the case in practice. - **More Model Variants**: There is a need for experiments on larger model sizes to validate the findings. - **Terminology Inconsistency**: The terminology used is inconsistent. In Figure 3, the terms "Good reasoning" and "Bad reasoning" are used, while in Figure 4, the terms change to "Good prompt" and "Bad prompt."

Questions

N/A

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Summary

This paper analyzed the mechanism of the large language models (LLMs) zero-shot Chain-of-Thought (CoT) reasoning, in which the authors found a pattern to discriminate a good reasoning path and a bad one with the saliency scores. Based on the findings, this paper proposed a set of instance-adaptive prompting approaches for zero-shot CoT reasoning. Experimental results on various tasks with distinct LLMs demonstrated the effectiveness of the proposed methods, validating the correctness of the findings.

Strengths

1. This paper is well-written and easy to follow. 2. This paper emphasized the instance-level CoT prompting rather than the task-level, providing a novel and fine-grained research object for LLM reasoning, which had not been explored in earlier works. 3. This work employed an interesting information flow analysis to delve into the mechanism of zero-shot CoT during the LLM inference, encouraging relevant CoT research. 4. The authors presented two simple yet effective instance-adaptive zero-shot CoT prompting approaches based on elaborative analysis, and empirical results verified their observations.

Weaknesses

1. Experiments only covered 7B and 13/14B models, involving larger model in the experiments would be more marvelous. 2. In Section 4.3, the authors didn’t interpret the results of consistency and complementary results in detail.

Questions

1. Why did the authors choose the saliency score instead of the widely used attribution score as the information flow analysis method? 2. What did some results with underlines in Table 1 mean? 3. Should the colon after the “remains” in the line 44 be removed? 4. Should the “and” in the “, and Tracking Shuffled Objects …” in line 258 be removed? 5. In Table 3 in the Appendix, why didn’t the authors highlight the best results and explain the results more? All my concerns have been addressed by the authors' rebuttal.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have adequately clarified the limitations of this work.

Summary

The authors argue that a single, task-level prompt is insufficient for addressing the diverse needs of different instances within a dataset. To overcome this limitation, they propose an instance-adaptive prompting (IAP) algorithm that differentiates between effective and ineffective prompts for individual instances. The authors also provide a detailed examination of the information flow at different layers and heads of the LLMs, offering insights into the internal mechanisms that contribute to reasoning quality. The proposed IAP strategy is shown to be effective across multiple models and tasks, highlighting its potential for advancing zero-shot reasoning in LLMs.

Strengths

1. The IAP algorithm is a creative solution that addresses the limitations of previous prompt strategy. The originality lies in the adaptive differentiation of prompts for individual instances and the use of information flow analysis to understand and enhance reasoning mechanisms within LLMs. 2. The quality of the experimental design is high, with comprehensive testing across different models and reasoning tasks.

Weaknesses

1. As the number of available prompts grows, the strategy for selecting the best prompt could become increasingly complex. 2. The titles of some figures are verbose (e.g. Figure 2). Keeping the title concise and giving a detailed explanation of the figures in the main text will make the paper better.

Questions

1. How were the saliency score thresholds determined, how do these thresholds affect the results? 2. In some paragraphs, full stops are missing, for example, at the end of the "Experiments" section and at the end of the first paragraph in the "Preliminary analysis" section.

Rating

7

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

it appears that the authors have made an effort to address the limitations of their work.

Response to Rebuttal by Authors

Thank you for your responses. After reading your rebuttal, I still have the following two major concerns: (1) IAP-ss needs to obtain the corresponding threshold on the training set. For the realistic inference scenarios, there is often no training set for parameter tuning, making the practical application scenarios of IAP-ss very limited. (2) The experiment setup for comparing self-consistency. The rebuttal does not provide a specific implementation explanation for self-consistency, which makes me confused. I hope the authors can give further explanations.

Response to the authors' rebuttal

Thank you for the authors' responses. I have read the rebuttal and all my concerns have been addressed. This paper analyzed the inner machanism of zero-shot CoT and proposed a novel instance-adaptive prompting strategy. I also reviewed the comments from the other reviewers, and believe this is quite a solid work. I will raise the score from 7 to 8.

Response to the authors

Thanks to the authors for resolving my concern. I will change me score from 6 to 7.