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

Introduction

Large language models (LLMs) have demonstrated capabilities at tackling copious reasoning tasks through Chain-of-Thought (CoT) . Compared to the few-shot setting for CoT generally, zero-shot CoT prompting can achieve approximate performance with merely one natural language prompt rather than complicated demonstrations, which has been proven as a simple and efficient paradigm . Numerous efforts have been thrown into searching for better prompts that can benefit zero-shot CoT reasoning. Plan-and-Solve employs a human-crafted prompt to break down the question and automatically generates reasoning steps. OPPR takes the LLM as an optimizer to update a zero-shot CoT prompt iteratively and produce corresponding optimized prompts for a given task. Self-discover selects relevant atomic reasoning modules (e.g. breaking down problems, critical thinking) for a given task, then adapts and customizes those modules to fit the task.

All prior methods focused on constructing prompts from the task perspective, aiming to find the optimal task-level prompt. Seeking the optimal prompt for a given task may achieve compelling performance, beating other prompts on the dataset scale. However, from the perspective of instance, the task-level optimal prompt within a dataset may have adverse effects on certain instances, whereby the model, capable of correctly answering them under other sub-optimal task-level prompts . Figure 1 illustrates an instance from GSM8K dataset , this is a simple question that can be straightforwardly answered correctly under "Don’t think. Just feel.", which is generally regarded as a less favorable prompt, but "Let’s think step by step" guides the LLM to bad reasoning in some steps. Therefore, 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.

Nevertheless, the severe challenge of choosing one of the suitable prompts for each instance remains: the difficulty of understanding why some reasoning processes succeed while others fail. To meet such a challenge, we intend to detect the mechanism of zero-shot CoT which is an unclear mystery . Neuron saliency score analysis is an important approach for observing the information flow during the model inference , by which we can observe a click of the dynamic reasoning process in certain steps. After comprehensive investigation across several LLMs and tasks, we find that a successful reasoning procedure tends to satisfy the following conditions: the semantic information of the question should be aggregated to the prompt first, and the reasoning steps gather information from both the original question and the synthesized question-prompt semantic information. Otherwise, it is more likely to be a failure reasoning. Such a saliency score phenomenon is in line with human intuition, as the question is the beginning of reasoning, one needs to understand it first, then solve it following the rules within the prompt while always concerning the question itself.

Inspired by the above findings, we further propose an instance-adaptive prompting strategy (IAP) for zero-shot CoT reasoning. Given a list of prompts in distinct styles, we try to recognize good ones that elicit LLMs to reason toward the correct answer while avoiding bad CoT reasoning, referring to the analytical results. We conduct comprehensive experiments with IAP and existing methods with multiple LLMs on various tasks. Experimental results show that the IAP can consistently improve the overall performance of LLMs such as LLaMA-2 13B, LLaMA-3 8B, and Qwen 14B on kinds of reasoning tasks including math, logic, and commonsense reasoning. Specifically, the IAP strategy achieves a 2%-4% accuracy enhancement across tasks and models compared to the optimal task-level prompt. Our contributions can be summarized as follows:

We look into the inside interactions among three components (i.e., question, prompt, rationale) in zero-shot CoT reasoning through the saliency score analysis and discover that good reasoning rationale tends to aggregate information from both the question and the prompt, in which the prompt first gathers information from the question. In contrast, bad reasoning probably ignores one of them.

We propose the IAP – an instance-level adaptive prompting strategy based on our findings to achieve better CoT reasoning by selecting a proper prompt that can elicit LLMs to reason from some given prompts for each question correctly.

Extensive experiments illustrate the superior performance of our instance-level adaptive prompting zero-shot CoT strategy, demonstrating the effectiveness of our findings for differentiating the reasoning processes with saliency scores.

Information Flow Analysis on Zero-shot CoT

It is critical to determine the key factors for good zero-shot CoT reasoning, therefore we dive into the LLMs inference process in disparate parts. There are three main components in zero-shot CoT: question $q$, prompt $p$, and rationale $r$, and we need to choose a proper tool to analyze the semantic information interactions among these components. The saliency score is a common practice for analyzing the information flow in In-Context Learning , and we intend to adapt it to CoT reasoning to observe the information flow in the zero-shot setting. The saliency matrix is computed by multiplying an attention matrix and its gradient for the target output element-wise as follows:

We define the reasoning that produces the right answer to a given question as good reasoning; otherwise as bad reasoning. Given various information interactions happen among reasoning steps during the model inference, choosing which step (i.e., output token) to explore is critical. Despite the most popular practice being the last step for the In-Context Learning , there are distinct circumstances in the CoT reasoning , our investigation shows that not all the final answers appear at the reasoning last step. To eliminate the effect of distinct LLM generation styles as much as possible, we adopt uniformly the answer generation step for all tasks as our observation time, concretely, we implement that with several regular expressions to recognize the answer step during model inference. More details are in Appendix A.2.

The good and bad reasoning patterns under "Don’t think. Just feel." are in line with the ones under "Let’s think step by step.", which is shown in Figure 2(c) and 2(d). Figure 2(c) illustrates that even such a prompt may guide LLMs to output the answer in very few steps after the question,

the prompt tokens still capture information from the question plainly, and the limited rationale tokens proactively take advantage of information from both the question and prompt. The phenomenon in a few cases cannot illustrate any universal pattern, hence, we randomly sample 100 instances including an even number of good and bad ones to test the suitability in a larger scope. Figure 3 elaborates that good reasonings have higher mean values on the question-to-prompt, question-to-rationale, and prompt-to-rationale than those bad, justifying the above phenomenon in a broader context. We can conclude that: For prompts that enable LLMs to reason correctly, there are significant saliency scores in the question-to-prompt and pronounced saliency scores from the question and prompt to the rationale; In contrast, for the prompts that do not lead LLMs to reason correctly, the saliency scores from the question to the prompt are usually not significant, or the flow from the question and the prompt to the rationale is not substantial. These findings align with the human cognitive process: given a question, one needs to comprehend it first, and then address it by applying the guidelines provided in the prompt while always concerning the question itself.

With the saliency scores phenomenon during zero-shot CoT reasoning, we believe the strength of saliency scores among them may affect LLMs’ reasoning quality. Hence, we obtain the saliency scores among the question, prompt, and CoT rationale:

2 Layer analysis

As observed in Figure 4(a), there is a pronounced peak in shallow layers of the LLM, demonstrating a substantial transfer of semantic content from the question to prompt in the good reasoning. This trend suggests that when the model formulates a robust prompt, it effectively aggregates the critical aspects of the original question at the outset, setting a strong foundation for later steps. Figure 4(b) maintains lower, yet consistent, saliency scores through the majority layers for both good and bad prompts when transferring information from the question to the rationale. This implies that while the question’s semantics are integral to crafting the rationale, the direct influence is far less than the initial aggregation seen in question to prompt saliency scores. Figure 4(c) depicts the information flow from the prompt to the rationale, we observe a minor but stable ascending trend for good prompts. This gradual integration underscores the importance of the prompt in orchestrating the connection between the given question and rationale, particularly in the later stages process within the LLM.

Through a layer-wise analysis, we notice that the question’s information first aggregates to the prompt in shallow layers, which suggests that an appropriate prompt acts as a catalyst, enhancing the model’s ability to integrate and leverage the question’s meaning. Subsequently, the reasoning gathers and refines information from both the original question and the synthesized question-prompt semantics, culminating in a coherent and contextually informed rationale. These information aggregation phenomena signify that shallow layers of the model are capable of encoding the semantic information of the question and prompt. The insights drawn from these findings evoke the potential for interpretability and reasoning capabilities of LLMs, indicating that the judicious formation of prompts can orchestrate the saliency scores in ways that affect rationales’ quality.

3 Head analysis

Multi-head attention is the fundamental component in the Transformer decoder to learn the same sequence from multi-view, like different positions of Transformer blocks, scattered heads are sensitive to their locations. Figure 5 provides an in-depth examination of the instance-level saliency scores within the attention mechanism of the LLM. Representation of saliency scores as heatmap visualizations offers a detailed perspective on how semantics flow the question, prompt, and rationale propagates through individual attention heads across various layers.

Figure 5(a) highlights the saliency scores from the question to the prompt, attention heads at the front of the middle and end positions effectively concentrate question semantics and aid their embedding into the prompt context. Notably, this pattern corroborates our understanding that certain heads are specialized in aggregating the shallow layers, which is essential for formulating coherent prompts. Figure 5(b) shows the transition from the question to the rationale, reflecting the model’s nuanced strategy of parsing the question to spawn a rationale and this aligns with the layer-level analysis, asserting the importance of inheriting question semantics, albeit less evidently than the question-to-prompt transition. Figure 5(c) delineates the flow from the prompt to the rationale, scattered saliency scores across heads denote that while all heads partake in the progression towards a rationale, middle and behind heads are pivotal in harmonizing the prompt with the rationale context. This staged intertwining of prompt and rationale semantics accentuates the sophisticated nature of information assimilation in the latter reasoning.

Combining the above 3 types of head-wise analysis, we note that the distribution and intensity of attention across heads are not homogeneous but are rather intricately patterned to orchestrate a hierarchical and systematic progression of semantics. The saliency scores from the question to the prompt and rationale are proven to be the core in the early phase, setting a solid foundation for rational derivation. The subsequent interactions that spawn the rationale further underscore the nuanced employment of attention heads in synthesizing compounds of the initial question with emergent prompt semantics. The intelligence encapsulated in this fine-grained attention tracing elucidates the role of discrete heads in sculpting the LLM’s reasoning.

Adaptive Instance-level Zero-shot CoT

To discover the zero-shot CoT reasoning capabilities of LLMs, we meticulously decompose the saliency score inherent in layers and heads to discern patterns of good and bad CoT reasonings in a fine-grained manner. In Section 2.2 and 2.3, we find that good reasonings always have higher saliency scores than bad ones in the question-to-prompt, question-to-rationale, and prompt-to-rationale, we further discover the front heads of the middle and end positions in shallow layers contribute to the saliency scores in both the question-to-prompt and question-to-rationale.

We first compute the saliency scores of question-to-prompt, question-to-rationale, as well as prompt-to-rationale. Then, for each question and a certain prompt, we compute the synthesized saliency score as follows:

Based on the above findings, we believe that a prompt with saliency scores surpassing the corresponding threshold is considered a good prompt for a given question, consequently mitigating the need to explore further prompts. This process terminates upon either identifying an optimal prompt or traversing all candidates.

Majority Vote (IAP-mv)

Alteratively, the IAP-mv necessitates the computation in Eq. 4 across all candidate prompts, then preserves the top maximum scores, predominant answer among these top scores is selected as the final answer. This synergistic combination ensures that the chosen prompt not only aligns with the LLM’s inherent reasoning pattern but also complies with the collective intelligence inferred from an assorted selection of potential prompts.

Both methods have pros and cons: IAP-ss possesses the efficiency of a heuristic-based sequential evaluation, which needs less computational resource; while IAP-mv owns the robustness supported by the consensus-based vote. Correspondingly, IAP-ss can be constrained in its performance potency since a few irregular instances may depart from our findings; though IAP-mv may achieve better performance, it demands the comprehensive evaluation of all candidate prompts. In summary, the IAP contributes a novel perspective on the paradigm of instance-level prompting strategies that drive the frontier of zero-shot CoT reasoning with LLMs.

Experiments

Models. We test IAP and comparison methods on LLaMA-3 8B , LLaMA-2 13B , and Qwen 14B since they are popular Transformer decoder-only LLMs, which is convenient for exploiting and analyzing inside architectures. We set the generation mode to greedy-decoding to minimize irrelevant confounders during the model inference to ensure the answers to fixed questions under the same model and prompt, and all the experiments are run on an 8x NVIDIA A100 GPU server.

Baselines. OPPR takes the LLM as an optimizer to update a zeros-hot CoT prompt iteratively and produce corresponding optimized prompts for copious tasks. Self-Discover selects relevant atomic reasoning modules (e.g., decomposing problems, critical thinking) for a given task, then adapts and customizes those modules to fit the task. These two frameworks aim to search for an appropriate prompt, which is essentially similar to our purpose, so we choose them as comparisons with the IAP to observe the performance difference between instance-level and task-level zero-shot CoT prompting.

Tasks & Metrics. GSM8K is a challenging dataset for assessing the capability of language models in multi-step math reasoning. SVAMP is presented for one-step math reasoning, which is easier than GSM8K. CommonsenseQA is designed to evaluate a model’s capacity for commonsense reasoning with questions that demand commonsense knowledge. The MMLU can assess a model’s multi-task learning abilities across natural language inference, commonsense reasoning, question answering, etc. Causal Judgement and Tracking Shuffled Objects are two sub-tasks in BBH , the former specifically tests a model’s ability to reason about the dynamics and interactions of objects in a given scenario and the latter presents scenarios that require identifying the underlying causes and effects of specific events or phenomena. We select the GSM8K and SVAMP for math reasoning, Causal Judgement, and Tracking Shuffled Objects-5-Objects for logic reasoning, CSQA, and MMLU for Commonsense reasoning. For all tasks, we adopt Accuracy as the only evaluation metric.

Zero-shot CoT Prompts. In the following part, we use #1 to represent "Let’s think step by step.", #2 denotes "First,", #3 is "The answer is after the proof.", #4 is "Before we dive into the answer,", #5 is "Let’s solve this problem by splitting it into steps.", #6 is "Let’s think about this logically.", #7 is "It’s a beautiful day.", #8 is "Don’t think. Just feel.", and #9 is "By the fact that the earth is round," and we implement the IAP by enabling these 9 prompts as the candidates.

2 Results

Prompts steer these LLMs to achieve different results in multiple tasks, and no single prompt can get an overwhelming performance on all datasets, which makes our research on the mechanism of zero-shot CoT valuable. Table 1 shows the zero-shot CoT reasoning results with LLaMA-3 8B and Qwen 14B and various prompts on 3 reasoning tasks, we put the results of LLaMA-2 13B in the Appendix since LLaMA models share a quite similar architecture.

Math reasoning. Compared with these prompts, IAP-mv improves the LLaMA-3 and Qwen’s accuracy on GSM8K from 64.52%, 60.50% to 66.34%, 62.81% respectively. On SVAMP, IAP-mv obtains a 1.33% improvement on both two models compared to the task-level-optimal prompt. It is worth noting that OPPR and Self-discover are unstable with different LLMs and datasets, indicating the unstable characteristics of task-level prompting. Results on these two math reasoning datasets demonstrate the IAP can benefit the math reasoning task.

Logic reasoning. For Causal Judgement, IAP-ss and IAP-mv enhance the accuracy of the task-level optimal prompt and IAP-mv outperforms OPPR, which is optimized by numerous iterations. For Tracking shuffle Objects, IAP-mv performs well with Qwen while achieving a sub-optimal accuracy with LLaMA-3, IAP-mv still obtains strong results, improving 2.4% and 2.3% with LLaMA-3 and Qwen, respectively.

Commonsense reasoning. On CSQA, the IAP improves the accuracy of the former best for 3.44% with LLaMA-3, and 2.45% with Qwen. On MMLU, LLaMA-3 and Qwen obtain improvement to a large margin, either. We note that improving IAP-mv and IAP-ss on commonsense reasoning is quite salient, demonstrating the effectiveness of the saliency score-based prompting strategies.

3 Ablation studies

The success of zero-shot prompting for CoT reasoning lies in the semantic information within those prompts, when the LLM receives a prompt, it would generate rationales by obeying the meaning of the prompt as much as possible. According to semantics, categorizes these zero-shot CoT prompts into 3 types: instructive, misleading, and irrelevant, and we further define that prompts in the same category are consistent, or otherwise they are complementary. Our 9 prompts can be divided into 3 consistency groups, but the irrelevant group contains only one prompt, thus we evaluate the complementary on the other two. For the complementary groups, we build them two-by-two. Table 2 depicts the performance of each group. We employ IAP-mv since it manifests a stronger capability in taking advantage of multiple prompts.

Efficacy

As we mentioned in Section 3, the IAP-mv trades efficiency for performance, and IAP-ss emphasizes efficiency. We introduce the reasoning time (seconds) for each iteration complete as the metric to measure the efficiency and conduct time-consuming experiments under the same setting to show the cost of IAP-mv, IAP-ss, and Self-Discover on the Tracking Shuffle Objects dataset, results are shown in Figure 6. All these strategies increase the computation cost to a certain degree, while IAP-ss may bring accuracy decreases than the task-level optimal prompt, it beats Self-discover. Though IAP-mv is the most time-consuming, it can improve performance, therefore, the two IAP strategies can be employed as trade-offs in different demand prioritization applications.

Related Work

CoT reasoning advances the reasoning abilities of LLMs by demonstrating a series of logical steps preceding the input demonstration. Building on the groundwork laid by CoT, Self-consistency innovates through a margin decoding strategy that emphasizes the majority paths to derive the final answer, presenting a significant leap in CoT reasoning. Similarly, the Least-to-most strategy decomposes a complex question into manageable subquestions, addressing them progressively to achieve a comprehensive solution. Furthermore, the Plan-and-Solve automates the generation of reasoning steps through a meticulously crafted prompt, streamlining the breakdown of questions into digestible parts that can be tackled sequentially.

Promisingly, the AutoHint framework augments the original prompt with enriched instructions extracted from contextual demonstrations. Similarly, the COSP capitalizes on answer pools derived from training sets to compute outcome entropy, inspired by the notion of self-consistency, thereby refining the selection process for QA pairs used during test set demonstrations. In specialized prompting, MathPrompter specifically caters to mathematics problems, employing handcrafted prompts to generate diverse algebraic expressions or Python functions. In contrast, Progressive-Hint Prompting facilitates dynamic interactions between users and LLMs, guiding the reasoning with hints to generate from previous answers. Moreover, InstructZero leverages an open-source LLM to enhance soft prompts relevant to Bayesian tasks, iteratively optimizing prompts to navigate through complex reasoning landscapes.

Advanced prompting approaches such as SelfzCoT and Meta-prompting showcase the evolutionary trajectory of prompting, which generates semantic and code prompts through a root prompt to obtain precise answers, while Meta-prompting deconstructs complex tasks into simpler sub-tasks, each addressed by specialized models to foster inter-model communication and apply intricate reasoning. Lastly, methodologies like OPRO and the innovative concept of evolutionary prompting aim to recursively optimize CoT prompts and generate varied prompts through mutations and crossovers, respectively. Self-discover selects relevant atomic reasoning modules (e.g., breaking down problems, critical thinking) for a given task, then adapts and customizes those modules to fit the task. Implement the customized reasoning structure when solving task instances. These workarounds significantly contribute to the development of zero-shot CoT prompts that guide LLMs toward more accurate problem framing, intermediate reasoning, and final answers.

Conclusion

In this paper, we aim to delve into the mechanism of LLMs in zero-shot CoT reasoning from the perspective of information flow to understand what happened during this process, and we find stronger saliency scores within question-to-prompt and question-to-rationale can lead to better LLM reasoning. To investigate these phenomena in a nuanced manner, we go deep into the Transformer layers and attention heads in the LLM and find the front of the middle and final heads in shallow layers carry more information during information flows. Inspired by that, we present an instance-adaptive zero-shot prompting strategy for better CoT reasoning. To demonstrate our findings, we conduct comprehensive experiments on several LLMs and tasks, and the results show our proposed strategies can improve the performance of LLMs on all candidate prompts, highlighting our interpretation of zero-shot CoT in the view of information flow.

Limitations

In this work, we select the answer step as the key step to investigate the saliency scores, even in most instances it can be located well, and some irregular answers can not be identified precisely, such a factor may affect the accuracy of our analysis. Despite our research in this paper providing insight into understanding the underlying workflow of zero-shot CoT reasoning, it cannot be the only interpretation, and we believe there must be better means to explain that.

References

Appendix A Appendix

Table 3 and Table 4 are supplementary for Table 1 and Table 1, the results here basically are coincidence with the Experiment section in the main body. However, the IAP can not obtain consistent enhancement, which may caused by the irregular output formats of LLaMA-2.

A.2 Answer Step Recognition

We prepare 3 types of answer formats to recognize the answer step while LLMs reasoning, concretely, employs the regular expression to judge whether the model has just output the answer to the given question. Once we detect some pre-defined patterns, we break the LLM’s generation for loop and compute the saliency scores at this time step. We put the recognition formats in Table 5.