PAS: Plug-and-Play Prompt Augmentation System

1. The baseline methods in the experimental section only include BPO, which is insufficient to demonstrate PAS as the current state-of-the-art (SOTA) solely based on its performance exceeding BPO. How does PAS's generated prompts compare with iterative methods like APO, OPRO, and PromptAgent? I have concerns regarding this. Based on my experience, adding supplementary descriptions to prompts can indeed enhance the performance of large language models, but it is challenging to surpass methods like APO, OPRO and PromptAgent, which involve comprehensive prompt rewriting. I suggest that the authors add comparative experiments with these baselines and conduct comprehensive comparisons on more diverse datasets, such as mathematical problems (GSM8K), commonsense questions (CommonsenseQA), and language understanding tasks (SST-5, TREC), among others.

2. PAS generates complementary prompts using large language models and filters them through the model as well. However, previous research has shown that large language models may exhibit self-enhancement bias, meaning LLMs may have a tendency to favor their own generated answers. The authors' approach may need to address this potential issue, such as using another model to evaluate the quality. Additionally, past research has shown that the improvement achieved by prompts for large language models can sometimes be inexplicable. For example, OPRO found that on the GSM8K dataset, the prompt “Take a deep breath and work on this problem step-by-step.” performs nearly 9% better than “Let's think step by step.” Therefore, relying solely on LLM-based data selection from a semantic perspective may be insufficient. The authors should consider a more comprehensive evaluation approach, such as generating results using prompt + complementary prompt on the LMSYS-1M and WildChat datasets and directly evaluating with corresponding metrics.

3. In Appendix D, the author mentions that PAS is considered Controlled Generation Time. How is this specifically demonstrated? BPO also inputs prompts into the LLM and receives optimized prompts in return, so why isn’t BPO considered Controlled Generation Time? Additionally, the author claims that PAS supports Long Documents and RAG. How is this reflected, and why doesn’t BPO qualify? I feel that the author’s analysis does not clearly address these issues. For these two points, I recommend that the author add more compelling evidence, such as additional experiments or a more detailed comparative analysis.

4. Writing suggestions:

a) The reference for BPO in line 94 should be added earlier; the citation appearing only by line 106 might impact reading fluency.

b) In the related work section, the first paragraph of the “Automatic Prompt Engineering” section addresses prompt design techniques but has little relevance to the “Automatic” aspect of APE. I recommend reclassifying it into a different related work section (or removing it), such as “Prompt Design” or “Prompt Engineering,” while keeping the other parts under “Automatic Prompt Optimization.” I hope the authors will consider this suggestion.

c) The titles of sections 3.2 and 3.3 (“PROMPTS COMPLEMENTARY DATASET” and “PROMPT COMPLEMENTARY DATASET”) are too similar, making it challenging for readers to distinguish between them. Adjusting them for clarity would be helpful.

Based on the questions I raised in the "Questions" section, I believe that this paper is generally unclear in its presentation, I strongly suggest that the author provides a clearer explanation in the next version.

1. Poor Citation Practice: The paper uses several important benchmarks for its main evaluation results, yet did not cite their works (eg. Arena-Hard, AlpacaEval). The authors need to make sure to credit all the works they used from other researchers.

Suggestion: The authors should add the appropriate citations and update the reference list.

2. Clarity: The paper has clarity issues, including several important parts of the paper being confusing. Examples and comments below:

Line 262: We manually design 4-5 few shot examples for each category, where we call it few-shot golden data. Then we utilize the prompt dataset from section 3.1 to generate high-quality (prompt, complementary prompt) pairs utilizing few-shot learning.

It is unclear how or under what criteria the few shot examples are designed. They author did not link to these few-shot examples nor explain what necessary steps were taken to ensure these design choices are unbiased.

Section 4.4: Human Evaluation

The paper did not explain how one of the main evaluation results, human evaluation, was conducted. Not knowing how the experiment was conducted (eg. who were the participants in evaluation and what specific dataset used) raises questions about the validity of the experiment.

Suggestion: Include information such as the number and qualifications of evaluators, the evaluation criteria provided to them, the dataset used, and any measures taken to ensure unbiased assessments.

3. Lack of validations: The presented method suffers from lack of validations, especially in crucial components which leverage LLM for decision making, which are known to be unreliable. Examples and comments below:

Line 244: Here $Q_{score}(p_i)$ represents the quality score assigned by BaiChuan 13b model to prompt $pi_i$, and $\tau$ denotes the quality threshold. By employing quality selection, we aim to enhance the overall quality of the prompt data.

Line 251: This results in a classification model capable of categorizing prompts into common categories such as Q&A and coding.

The authors claim the curation pipeline selects high quality data and accurately classifies categories. However, the author did not provide any validation results. Presentation of quantitative metrics for data quality selection and classification accuracy would be beneficial.

3. Generalizability The authors claim their system can be generalized to any LLMs and obtain SOTA performance. However, this claim raises concerns. Examples and comments below:

Line 392: To address Q1, we used Qwen-2-7B-Instruct as the base model due to its outstanding performance.

The author chose to present Qwen-2-7B-Instruct’s results because they are more favorable. To demonstrate the method is indeed robust to model as the author claimed, results should be shown for more models.

Suggestion: The authors should include a comprehensive table or figure showing results across a wider range of models, including both smaller and larger models from different families. This would help demonstrate the claimed robustness and generalizability.

Line 386: To evaluate the effectiveness of our PAS model, we used three comprehensive benchmarks Arena-hard, Alpaca-Eval 2.0 and Alpaca-Eval 2.0 (LC) to thoroughly assess the model’s performance.

These benchmarks may not be representative of the real-world use case of LLM. The authors claim they are comprehensive, however, Arena-Hard contains difficult user queries from Chatbot Arena and AlpacaEval are mainly information retrieval tasks. For instance, Arena-Hard queries are selected for being well-defined, and do not necessarily require complementary augmentation. Unfiltered prompts in the wild, which are not well-defined, might suit better for evaluating the effectiveness of PAS. Hence, it is difficult to tell from these results whether PAS truly improves compared to the baseline.

Suggestion: The authors should include an additional evaluation using a dataset of unfiltered, real-world prompts. Discussion the limitations of the current evaluation benchmarks and how to address these limitations in future work would also be beneficial.

4. Numerical Significance

Line 439: Notably, PAS exhibits a marked improvement in performance metrics compared to BPO, exceeding the baseline by 3.41 points on average. This is particularly evident in models like GPT4-0613, where the average score improvement is as high as 5.89 points. Even in cases where the improvement is smaller, such as Llama3-70b-Instruct and GPT-4-turbo-2024-04-09, PAS still manages to outperform BPO for more than 1 point, indicating its robustness and consistency.

It is difficult to tell whether the improvement is truly significant, when the difference between PAS and baseline is within variance of the benchmark. The author should provide confidence intervals for the presented results. It also appears that Arena-Hard provides confidence interval for model evaluation, which the authors should report.

1. I agree that prompting is central to improving LLM performance, and this work is timely and interesting. However, my first concern is the positioning of the work relative to the rich literature on automated prompt engineering techniques. While the authors briefly describe other approaches, they do not clearly articulate how PAS is different from and better than these techniques.

2. My second concern is that the paper fails to clearly articulate the novelty of the work. While generating complementary prompts is interesting, there is nothing new in this approach. Additionally, once the data is generated, a simple supervised fine-tuning (SFT) is applied for complementary prompt generation.

3. My third concern is that the paper lacks a comprehensive comparison to various other approaches (baselines). The authors only compare their method against BPO, but it’s unclear why they chose this single baseline, especially when there are several recent approaches, such as OPRO, EVOPrompt, PromptBreeder, and PromptWizard.

Limited Details on Dataset Creation: While the paper describes the process of creating the prompt-complementary prompt dataset, more details are needed. Specifically, the types of prompts included in the initial dataset (LMSYS-1M and WildChat), the criteria for quality selection and classification categories should be further elaborated. A more detailed analysis of the dataset distribution (Figure 12) with examples would be beneficial.

Clarity on PA Model Architecture: The paper lacks specifics about the architecture and training details of the PA model. Information on the LLM used for training the PA model, the fine-tuning process (SFT), and hyperparameter choices is essential for reproducibility and understanding the system's inner workings.

Comparison with other Methods: The paper primarily focuses on comparison with BPO. Including comparison with a wider range of APE methods (OPRO, APO, APE) would strengthen the paper's claims of achieving SOTA performance. Also there are other prompt augmentation methods which could/should be studied.

Discussion on Limitations: The paper could benefit from a more explicit discussion of the limitations of PAS. For instance, potential failure cases where the PA model fails to generate relevant complementary content, or the impact of the quality of the base prompt on the effectiveness of PAS, could be addressed.