GRAD-SUM: Leveraging Gradient Summarization for Optimal Prompt Engineering
Abstract
Prompt engineering for large language models (LLMs) is often a manual time-intensive process that involves generating, evaluating, and refining prompts iteratively to ensure high-quality outputs. While there has been work on automating prompt engineering, the solutions generally are either tuned to specific tasks with given answers or are quite costly. We introduce GRAD-SUM, a scalable and flexible method for automatic prompt engineering that builds on gradient-based optimization techniques. Our approach incorporates user-defined task descriptions and evaluation criteria, and features a novel gradient summarization module to generalize feedback effectively. Our results demonstrate that GRAD-SUM consistently outperforms existing methods across various benchmarks, highlighting its versatility and effectiveness in automatic prompt optimization.
GRAD-SUM: Leveraging Gradient Summarization for Optimal Prompt Engineering
Derek Austin††thanks: Primary author. Galileo Technologies Elliott Chartock Galileo Technologies
| Prompt State | Prompt |
|---|---|
| Initial Prompt | Answer the question: {question} |
| Optimized Prompt |
You are tasked with solving a grade school math question. Follow these detailed steps to ensure a clear and accurate solution:
1. Identify the Mathematical Operation: Begin by determining the type of math operation required (e.g., addition, subtraction, multiplication, division). Clearly state this in your response and explain why this operation is needed for the given question. 2. Outline the Thought Process: Provide a logical and structured approach to solving the problem. Explain your reasoning in detail, as if you are teaching the concept to a student. Ensure each step is connected and follows logically from the previous one, making it easy to follow. 3. Step-by-Step Calculations: Break down the problem into smaller, manageable steps. Show each calculation in detail to demonstrate how you arrive at the solution. Include all intermediate steps and results to offer a comprehensive understanding. 4. Verify Accuracy: After arriving at the solution, double-check your calculations and the final numerical answer to ensure its accuracy. Explain how you verified the correctness of your answer, such as by using inverse operations or checking with a different method. 5. Final Answer: Present the final answer in a clear and precise manner, ensuring it is easy to understand. State explicitly that the answer has been verified for accuracy. Here is the question: {question} |
1 Introduction
The introduction of machine learning and artificial intelligence has automated many tasks, but creating and refining prompts for LLM applications remains largely a manual process. Practitioners often struggle to find the right prompt to elicit accurate and high-quality outputs. Common practice is to judge output quality by implementing fully automated evaluation feedback loops, where LLMs assess outputs based on user-provided evaluation criteria, these are often called LLM-as-a-judge methods. Seeing as evaluation is an automatic process in LLM-as-a-judge scenarios, the main bottleneck in automating prompt engineering is the actual refinement of the prompt. Although there have been attempts to automate this process (Yang et al., 2023) (Pryzant et al., 2023) (Zhou et al., 2023), we found that these methods either could not be adapted seamlessly to all tasks or were too costly to implement at scale.
Automatic prompt optimization methods broadly fall into two categories: Monte Carlo search methods Yang et al. (2023) and feedback-based (or gradient-based) methods (Pryzant et al., 2023)Fernando et al. (2024)(Prasad et al., 2023). Monte Carlo search methods start with a base prompt and iterate over hundreds if not thousands of possible prompts in order to find the best performer, an inefficient process. These methods use minimal feedback on why their responses might be incorrect and instead search somewhat blindly through prompt space to identify the best possible prompt, making small changes along the way. Feedback-based methods, on the other hand, use ratings and explanations of how their answer could be improved from previous iterations to generate new prompts. They aim to emulate a traditional optimization process by iteratively refining prompts based on detailed feedback, closely mimicking the approach of a human prompt engineer. As shown in (Yang et al., 2023), Monte Carlo methods take many iterations to converge, and thus are likely to be too costly for most practitioners. Therefore, in approaching the problem of automatic prompt engineering we choose to focus on feedback-based methods.
One of the seminal works that inspired our feedback-based approach was presented in "Automatic Prompt Optimization with ‘Gradient Descent’ and Beam Search" (Pryzant et al., 2023). ‘Gradients’ in this context, refers to natural language descriptions of errors in LLM responses and suggestions for improving the system’s response in the future. The method presented in the aforementioned paper uses each of these gradients individually to generate a new prompt. Their approach addressed classification tasks such as hate speech detection, lie detection, jailbreak detection, and sarcasm detection, using a restrictive matching evaluator where predictions must exactly match the expected answers. As most LLM use cases do not have a strict expected output we improve upon their work by introducing LLMs-as-a-judge evaluations - which enable accurate judgment of any response for all possible tasks.
Our approach, GRAD-SUM, introduces a task description input that allows models to incorporate high-level, domain-specific information provided by the user. Additionally, we introduce natural language evaluation criteria also provided by the user, enabling our method to use LLM-as-a-judge metrics, extending (Pryzant et al., 2023) method to all possible use cases. We also improve upon (Pryzant et al., 2023) method, by introducing a new gradient summarization module after discovering that editing the prompt based on feedback from a single output, as done in (Pryzant et al., 2023) often resulted in highly specific candidates prompts that failed to generalize to the broader dataset population. To address this, our gradient summarization module takes all computed gradients and produces a generalized summary of the feedback, akin to averaging gradients over many samples, a common practice in machine learning. Overall, our approach combines a novel gradient summarization module, user-provided task descriptions and evaluation criteria to create a more flexible, scalable and performant solution for automated prompt engineering than any existing today.
2 Related Work
Customizing an LLM to a user’s specific task can fit into two broad categories: parametric and non-parametric approaches. Parametric approach’s usually consist of parameter efficient fine-tuning (Hu et al., 2022) (Li and Liang, 2021) where only a portion of parameters are updated based on backpropagating gradients. Prefix tuning introduces new learnable tokens that one can think of as a prompt in continuous space, that are learned for specific tasks (Li and Liang, 2021). However, as these tokens defined in continuous space they are not easily interpretable. Other approaches like (Shin et al., 2020) backpropagate gradients and choose the best discrete prompts. However, even this technique suffers from interpretability as most prompts are not grammatically correct or coherent sentences.
Most practitioners currently rely on black box API based LLM’s for language generation where there is no potential for backpropagation of any kind (Brown et al., 2020). Therefore, prompt optimization becomes a discrete optimization problem over an extraordinarily large search space. Some methods attempt to tackle this problem through the use of complementary models. (Deng et al., 2022) attempt to use a policy network that receives rewards from an evaluator function in order to update a discrete prompt. (Chen et al., 2023) uses an open source model to convert a soft-prompt defined in embedding space into a discrete prompt that was then fed to the LLM API. Others use evolutionary algorithms in order to modify, delete and add a list of candidate prompts that are evaluated each iteration (Fernando et al., 2024) (Prasad et al., 2023). Simple Monte Carlo sampling around candidate prompts (Yang et al., 2023) can complement these methods or be implemented entirely on their own.
Most of these methods attempt to use some sort of signal from an evaluator LLM in order to asses the quality of the output. However, it has been shown that LLMs are not great at self-correction without use of auxiliary knowledge or tools (Stechly et al., 2023) (Ma et al., 2024) (Huang et al., 2024) (Gou et al., 2024). Therefore the introduction of specific evaluation criteria, as in our method, the use of tools, as in (Gou et al., 2024), or modifying LLM critiques to better find the root of the errors (Ma et al., 2024) becomes of paramount importance to finding an optimized prompt.
3 Method
Our method consists of 5 distinct modules as seen in Figure 1: generation, evaluation, gradient generation, gradient summarization, and prompt editing. Throughout our optimization process, we utilize beam search with a beam size of three candidate prompts, feeding each potential prompt to the modules introduced below. Figure 1 showcases one iteration for one potential prompt.
3.1 Generation
The generation module necessitates a dataset, a prompt, and an LLM with which to generate outputs. In usual workflows the prompt will consist of formatting ‘slots’ where the dataset will provide text that will be used to fill in the slots as shown in Figure 1. For instance, a prompt for a retrieval augmented generation (RAG) workflow could be:
“Given the following context: {context}. Answer the question {question}."
The dataset should then consist of the columns ‘context’, ‘question’ and optionally target outputs which can be used in the evaluate module. We then feed these formatted prompts to the generation function and receive LLM generated responses.
3.2 Evaluation
Our evaluation module then takes in the generations from the generation function, user-defined evaluation criteria, and optionally an expected answer. The module in turn returns a rating as to how well the response performs as judged by the evaluation criteria as well as an explanation for the rating. We ask for the explanation before the evaluator module returns a rating as chain of thought has been shown to dramatically improve LLM performance, especially on reasoning tasks (Wei et al., 2022). For use cases with an expected answer, evaluation criteria can be as simple as:
“Does the LLM generated response semantically match the expected answer? If a reasonable human would judge it to, give the response a 1, otherwise a 0."
We note that we only ask for a binary indicator as empirically we found LLM’s are far better and more consistent at binary indicators than a sliding numeric scale.
3.3 Gradients
The gradient module receives the current prompt, the evaluation criteria, a description of the task at hand, and a maximum of 5 generation responses that received a rating of 0 from the evaluation module. Note that any generation receiving a rating of 1 in the evaluate step will be excluded from this process as it cannot be improved, thus it is possible the gradient module will receive less than 5 inputs. An LLM is then asked to evaluate the input and output for the specific prompt and provide actionable methods for improving the prompts to address deficiencies noted in the rating explanation. The gradient module leverages all available information—the task description, input to the generation function, the LLM response from the generate function, evaluation criteria, and explanations for poor ratings—to identify areas for improvement. Below is a sample of what a gradient will look like:
“The model should draw upon the entire context and state its reasoning explicitly."
This step aims to emulate human evaluation of generations, determining how to adjust the prompt to achieve the desired outcomes.
| Dataset | Generating Model | Optimization Method | Initial Validation Rating | Final Validation Rating |
| GSM8K | GPT 3.5 | DSPY | 0.635 | 0.755 |
| GSM8K | GPT 3.5 | GRAD-SUM | 0.635 | 0.82 |
| Orca Math | GPT 3.5 | DSPY | 0.395 | 0.455 |
| Orca Math | GPT 3.5 | GRAD-SUM | 0.395 | 0.575 |
| Neural Bridge RAG | GPT 3.5 | DSPY | 0.605 | 0.885 |
| Neural Bridge RAG | GPT 3.5 | GRAD-SUM | 0.605 | 0.915 |
| HellaSwag | GPT 3.5 | DSPY | 0.575 | 0.48 |
| HellaSwag | GPT 3.5 | GRAD-SUM | 0.575 | 0.795 |
| HotPot QA | GPT 3.5 | DSPY | 0.575 | 0.626 |
| HotPot QA | GPT 3.5 | GRAD-SUM | 0.575 | 0.725 |
| MMLU | GPT 3.5 | DSPY | 0.45 | 0.56 |
| MMLU | GPT 3.5 | GRAD-SUM | 0.45 | 0.625 |
| MT & Vicuna Bench | GPT 3.5 | DSPY | 0.831 | 0.823 |
| MT & Vicuna Bench | GPT 3.5 | GRAD-SUM | 0.831 | 0.95 |
3.4 Gradient Summarization
While methods like (Pryzant et al., 2023) use the gradients individually to then generate new prompts (ie. 5 gradients would lead to generating 5 new prompts), we find that this leads to prompts that are far too specific to certain questions. Additionally, the task of evaluating these candidate prompts consumes API calls that can lead to costly compute bills. Therefore, we found the most effective way of generating new general-purpose prompts was to summarize all gradient feedback into one general paragraph that could apply to the dataset as whole. We feed the task description and the gradients computed in the previous step to the gradient summarization module and use an LLM to generate a one-paragraph summary of the critiques taking into consideration the task at hand. This step can be thought of as analogous to averaging gradients over a mini-batch to stabilize training.
3.5 Prompt Editor
Our prompt editor module then takes in the current prompt, the summarized gradient, and the task description and outputs a candidate prompt for each prompt within our beam (doubling our beam size momentarily). The new prompts should likely address the critiques provided in the previous iteration. Each candidate prompt is checked to ensure every slot in the current prompt is present in the new prompt in order to avoid information loss of any kind. We then perform an evaluation of each new candidate prompt. Specifically, we take a randomly sampled subset (five rows) of the dataset and feed the responses to our generate and evaluate modules, choosing the two candidate prompts with the highest average rating. As our beam is now 5 prompts wide we reduce to a beam size of 3 through selecting the top 3 performers as given by the upper confidence bound (Kuleshov and Precup, 2014). This reduction step thus always retains the best-performing prompt from previous iterations, reducing the variance from evaluating candidate prompts on a small sample size.
4 Experiments & Results
In order to benchmark our method we utilize common datasets: GSM8k (Cobbe et al., 2021), Orca Math (Mitra et al., 2024), Neural Bridge RAG (AI, ), Hella Swag (Zellers et al., 2019), HotPot QA (Yang et al., 2018), MMLU (Hendrycks et al., 2021) and a mixture of the first turn from MT Bench and Vicuna Bench (Zheng et al., 2023). We do not use any expected answers in our MT Bench and Vicuna Bench results, highlighting our use of user-provided evaluation criteria extends our technique to datasets without expected answers. For each dataset if there is an available training dataset we extract 30 random samples for our training set and 200 random samples from the validation or test set in order to simulate typical industry evaluation flows. If there is no available test or validation set we sample 200 from the train set and hold them out during training. 10 iterations are run over the train set with 10 rows being evaluated each iteration for each prompt (thus 30 calls for 3 candidate prompts) and the best prompt (the best performing prompt left in our beam), is extracted and used in our final validation. All datasets used can found on our GitHub.111https://github.com/rungalileo/prompt_optimization_datasets Seeing as (Pryzant et al., 2023) is not directly comparable to our approach on these benchmarks due to their classification / exact-matching framework222Figure 2 compares to the method introduced in (Pryzant et al., 2023) after extending their work to use LLM-as-a-judge metrics, showcasing the gradient summarization module improves upon their approach., we instead compare to the popular prompt optimizer, DSPY (Khattab et al., 2023). As DSPY is an abstracted prompting tool that does not allow users to directly control the initial prompt, we extracted DSPY’s initial prompt and used it as the starting point for our method to ensure a fair comparison. The same evaluation criteria is used with DSPY and we create a custom LLM-as-a-judge metric within their system in order to make our methods as comparable as possible. We utilize the DSPY COPRO optimizer and allow for 10 passes over the dataset to ensure a fair comparison. For our method we conduct evaluation, gradient generation and summarization with GPT 3.5 (‘gpt-3.5-turbo-0125’). Prompt editing is performed with GPT4o (‘gpt-4o-2024-05-13’) in order to ensure high-quality new prompts. The initial prompts, the final prompts, the task descriptions and evaluation criteria used for the respective tasks are provided in the appendix.
As you can see in Table 2 our method improves and outperforms DSPY on all use cases. We average an improvement of 14% over our initial prompts, showcasing the robustness and strength of our optimizer.
Gradient Summarization
We also conduct an ablation where we remove the summarization module in order to assess it’s efficacy. Removing the summarization module is analogous to comparing our method with (Pryzant et al., 2023) after extending their method to use LLM-as-a-judge evaluation. We note an average validation increase of 5% with the introduction of the summarization module. Upon examining the HotPot QA final prompt, it became obvious the prompt became far too specific to single gradients found throughout the training loop. Below is an extract from the final HotPot QA prompt that clearly attaches to information relevant to individual examples:
“Reasoning: Let’s think step by step to logically deduce the answer. Focus on the relevant information in the context to arrive at the correct answer. Ensure you consider all details, specific terms, and names mentioned in the context, such as the regiment within the Guards Division, the term ’shorts,’ the founding year of Image Comics, middle names, the specific crime Brian Nichols was on trial for, and the full names of individuals. Provide a comprehensive response covering all individuals mentioned in the context where applicable."
Thus, we find that the gradient summarization module is essential for maintaining high-quality prompts that can generalize to a validation dataset.
5 Conclusion
We introduce a scalable, flexible prompt optimization framework that improves LLM performance by an average of 14% across many popular datasets. In all scenarios our framework is also able to find a better prompt than our initial prompt. Finally, our framework can be used across many of scenarios including those without expected answers, illustrating the flexibility of our framework. By automating the prompt engineering process and demonstrating consistent performance gains across multiple datasets and models, our work enables efficient, flexible and scalable utilization of LLMs in real-world applications, reducing the time and effort required to achieve high-quality outputs.
6 Limitations
Our system currently supports only LLM-as-a-judge metrics. While these metrics are effective in many scenarios, they may not be suitable for all tasks. Expanding the system’s capability to support any user-defined metric, including numerical and domain-specific metrics, is left for future work. The use of LLMs for evaluation also introduces potential biases inherent in the models themselves. This has the potential to affect the accuracy of the evaluation process. Finally, there is still a dependence on the user to define the task descriptions and evaluation criteria, which have a large impact on final prompt quality. Streamlining these aspects to require minimal user input while maintaining effectiveness remains a challenge.
References
- (1) Neural Bridge AI. Neural-bridge/rag-dataset-12000 · datasets at hugging face.
- Brown et al. (2020) Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
- Chen et al. (2023) Lichang Chen, Jiuhai Chen, Tom Goldstein, Heng Huang, and Tianyi Zhou. 2023. Instructzero: Efficient instruction optimization for black-box large language models.
- Cobbe et al. (2021) Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
- Deng et al. (2022) Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric Xing, and Zhiting Hu. 2022. Rlprompt: Optimizing discrete text prompts with reinforcement learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3369–3391.
- Fernando et al. (2024) Chrisantha Fernando, Dylan Sunil Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2024. Promptbreeder: Self-referential self-improvement via prompt evolution.
- Gou et al. (2024) Zhibin Gou, Zhihong Shao, Yeyun Gong, yelong shen, Yujiu Yang, Nan Duan, and Weizhu Chen. 2024. CRITIC: Large language models can self-correct with tool-interactive critiquing. In The Twelfth International Conference on Learning Representations.
- Hendrycks et al. (2021) Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations.
- Hu et al. (2022) Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
- Huang et al. (2024) Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large language models cannot self-correct reasoning yet. In The Twelfth International Conference on Learning Representations.
- Khattab et al. (2023) Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan A, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2023. DSPy: Compiling declarative language model calls into self-improving pipelines. In R0-FoMo:Robustness of Few-shot and Zero-shot Learning in Large Foundation Models.
- Kuleshov and Precup (2014) Volodymyr Kuleshov and Doina Precup. 2014. Algorithms for multi-armed bandit problems. arXiv preprint arXiv:1402.6028.
- Li and Liang (2021) Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582–4597.
- Ma et al. (2024) Ruotian Ma, Xiaolei Wang, Xin Zhou, Jian Li, Nan Du, Tao Gui, Qi Zhang, and Xuanjing Huang. 2024. Are large language models good prompt optimizers? arXiv preprint arXiv:2402.02101.
- Mitra et al. (2024) Arindam Mitra, Hamed Khanpour, Corby Rosset, and Ahmed Awadallah. 2024. Orca-math: Unlocking the potential of slms in grade school math. arXiv preprint arXiv:2402.14830.
- Prasad et al. (2023) Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. 2023. Grips: Gradient-free, edit-based instruction search for prompting large language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3845–3864.
- Pryzant et al. (2023) Reid Pryzant, Dan Iter, Jerry Li, Yin Lee, Chenguang Zhu, and Michael Zeng. 2023. Automatic prompt optimization with “gradient descent” and beam search. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7957–7968.
- Shin et al. (2020) Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222–4235.
- Stechly et al. (2023) Kaya Stechly, Matthew Marquez, and Subbarao Kambhampati. 2023. GPT-4 doesn’t know it’s wrong: An analysis of iterative prompting for reasoning problems. In NeurIPS 2023 Foundation Models for Decision Making Workshop.
- Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
- Yang et al. (2023) Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. 2023. Large language models as optimizers. arXiv preprint arXiv:2309.03409.
- Yang et al. (2018) Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369–2380.
- Zellers et al. (2019) Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800.
- Zheng et al. (2023) Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Preprint, arXiv:2306.05685.
- Zhou et al. (2023) Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations.
Appendix A Prompt Optimization Results
For each dataset we list out the initial prompt, the task description, the evaluation criteria and the final prompt obtained. We provide the final prompt for all methods: GPT 3.5 with DSPY, and GPT 3.5 with our method.
A.1 GSM8k
Task Description:
The task is answering grade school math questions.
Evaluation Criteria:
Does the output align with the expected answer? The questions are math questions. Check if the answer matches the expected answer. Give it a 1 if a math teacher would consider the answer correct. Give it a 0 if the answer is incorrect. Do not worry about intermediate calculations, only the final answer.
Initial prompt for our method & DSPY:
The task is to answer math questions.
—
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question provided
—
Question: {question}
Reasoning: Let’s think step by step in order to Answer:
${answer}
Answer:
Best Prompt GPT 3.5 (DSPY):
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Step-by-Step Explanation:» the answer to the question provided
—
Question: {question}
Reasoning: Letś think step by step in order to find the solution. We will break down the equation and solve for the unknown variable. Step-by-Step Explanation: [Insert detailed steps and calculations here]
Step-by-Step Explanation:»
Best Prompt GPT 3.5 (Our method):
The task is to answer grade school math questions accurately and clearly. Follow these detailed steps to ensure clarity and correctness in your response.
Question:
{question}
Reasoning: Let’s think through the problem step by step to produce the correct answer. Follow these instructions carefully:
1. Identify Key Information:
- Read the
{question}
carefully.
- Identify and list the key pieces of information given.
- Clearly state what the question is asking for.
2. Choose the Appropriate Method:
- Determine which mathematical methods or formulas are needed to solve the problem.
- Explain why this method or formula is appropriate for the given question.
3. Apply the Method:
- Carefully apply the chosen method or formula to the problem.
- Show each step of your calculation clearly and in order.
- Explain your reasoning at each stage to ensure clarity.
4. Double-Check:
- After arriving at a solution, double-check your calculations to confirm their accuracy.
- Review your explanation to ensure it is logical and easy to understand.
- Revisit the question to make sure your solution addresses exactly what was asked.
Answer:
A.2 Orca Math
Task Description:
The task involves solving math word problems.
Evaluation Criteria:
Does the output align with the expected answer? The questions are math questions. Check if the answer matches the expected answer. Give it a 1 if a math teacher would consider the answer correct. Give it a 0 if the answer is incorrect. Do not worry about intermediate calculations, only the final answer.
Initial prompt for our method & DSPY:
The task is to answer math questions.
—
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question provided
—
Question: {question}
Reasoning: Let’s think step by step in order to Answer:
${answer}
Answer:
Best Prompt GPT 3.5 (DSPY):
Take advantage of the large size of the language model and generate detailed explanations for the math questions instead of just the answers.»
—
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Explanation: the answer to the question provided
—
Question: {question}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (Our method):
The task involves solving math word problems accurately by following a clear and logical reasoning process. Please adhere to the following format to ensure a coherent and cogent response:
—
Question:
{question}
Reasoning: Let’s think step by step in order to produce the answer. We will break down the problem into smaller parts, perform the necessary calculations, and show all work and thought processes clearly. Ensure each step logically follows from the previous one and aligns with the evaluation criteria of correctness, completeness, and clarity.
Answer: Provide the final answer to the question based on the reasoning process above.
—
Question: {question}
Reasoning: Let’s think step by step in order to produce the answer. We…
Answer:
A.3 Neural Bridge RAG
Task Description:
The task is a question answering task given specific context that should have the answer.
Evaluation Criteria:
Does the llm output match the expected answer? If the model says it does not have enough context to answer the question give it a 0. Otherwise judge whether a human would grade the output as matching the expected answer. Adding context around the answer is fine as long as the answer is correct according to the expected answer within the brackets <EXPECTED ANSWER>. If it does match give it a 1. If it does not give it a 0.
Initial prompt for our method & DSPY:
The task is a question answering task given context that should have the answer.
—
Follow the following format.
Question:
${question}
Context:
${context}‘
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question given the context provided
—
Question: {question}
Context: {context}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (DSPY):
Guide the model to closely examine the question through context, utilize advanced inference techniques to deduce implicit information effectively, and generate a comprehensive and sophisticated response geared towards enhancing overall understanding and insights.»
—
Follow the following format.
Question:
${question}
Context:
${context}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Sophisticated Answer: the answer to the question given the context provided
—
Question: {question}
Context: {context}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (Our method):
You are tasked with answering a question based on a given context. Follow the detailed steps below to ensure your response is specific, comprehensive, and derived solely from the context provided.
1. Carefully read and fully understand the question. Clarify exactly what information is being sought by the question.
2. Thoroughly analyze the given context. Identify and extract all relevant pieces of information that directly pertain to the question.
3. Think logically and proceed step-by-step to connect the extracted details from the context to the question. Clearly explain your reasoning process, ensuring it is detailed, coherent, and directly related to the context.
4. Provide the final answer based on your detailed reasoning process, ensuring it directly addresses the question.
Use the following format for your response:
Question:
{question}
Context:
{context}
Reasoning: Let’s think step by step to understand and find the answer. First, identify specific details from the context that are relevant to the question. Then, logically connect these details to derive the answer. Ensure your reasoning is clear, detailed, and coherent.
Answer: The answer to the question based on the context provided
A.4 HellaSwag
Task Description:
The task is to complete a sentence with the most logical of 4 possible options.
Evaluation Criteria:
Does the output match the the expected answer? If it does give it a 1. If it does not give it a 0. It does not have to match exactly but it should be close enough that a reasonable human would consider the output to match the expected answer. Make sure that the chosen completion is the correct completion.
Initial prompt for our method & DSPY:
The task is to complete a sentence with the most logical of 4 possible options.
—
Follow the following format.
Context:
Context for the question
Ending 1:
First possible ending for the question
Ending 2:
Second possible ending for the question
Ending 3:
Third possible ending for the question
Ending 4:
Fourth possible ending for the question
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question given the possible endings provided
—
Context: {context}
Ending 1: {ending1}
Ending 2: {ending2}
Ending 3: {ending3}
Ending 4: {ending4}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (DSPY): The task is to complete a sentence with the most logical of 4 possible options.
—
Follow the following format.
Context:
Context for the question
Ending 1:
First possible ending for the question
Ending 2:
Second possible ending for the question
Ending 3:
Third possible ending for the question
Ending 4:
Fourth possible ending for the question
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question given the possible endings provided
—
Context: {context}
Ending 1: {ending1}
Ending 2: {ending2}
Ending 3: {ending3}
Ending 4: {ending4}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (Our method):
Your task is to complete a sentence with the most logical of 4 possible options. Follow these detailed instructions to ensure the best possible output.
### Task Instructions
1. Understand the Context: Carefully read the provided context to grasp the situation, the main idea, and any underlying nuances. Pay close attention to the details that are crucial for understanding the sentence.
2. Analyze Each Option: Evaluate each possible ending individually. Consider its relevance, logical consistency, and coherence with the context. Think about how each option fits with the main idea and details provided in the context.
3. Compare Against the Context: Cross-check each ending with the context to ensure it fits seamlessly and supports the overall meaning. Eliminate any options that do not make sense or disrupt the flow of the sentence.
4. Reasoning Process: Explain your thought process step-by-step. Consider the context and evaluate each ending thoroughly. Justify why each ending is or isn’t suitable, providing clear reasons for your choices.
5. Select the Most Logical Ending: Choose the ending that best completes the sentence in a coherent and meaningful way. Ensure that the chosen ending aligns perfectly with the context and enhances the overall understanding of the sentence.
### Format
Context:
{context}
Ending 1:
{ending1}
Ending 2:
{ending2}
Ending 3:
{ending3}
Ending 4:
{ending4}
Reasoning: Let’s think step-by-step to produce the answer. We need to consider the context and evaluate each ending one by one to determine which is the most logical. We should look for consistency, relevance, and coherence in the story.
Answer: The answer to the question given the possible endings provided.
—
Context: {context}
Ending 1: {ending1}
Ending 2: {ending2}
Ending 3: {ending3}
Ending 4: {ending4}
Reasoning: Let’s think step-by-step in order to…
A.5 Hotpot QA
Task Description:
The task is to reason over context given a question that will require logical deduction.
Evaluation Criteria:
Does the llm output answer the question correctly while only the context provided? Ensure that the model adheres to the context while providing a correct answer. If it does give it a 1. If it does not give it a 0. If the answer is that the context is not provided give the answer a 0.
Initial prompt for our method & DSPY:
The task is a question answering task given context that should have the answer.
—
Follow the following format.
Question:
${question}
Context:
${context}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question given the context provided
—
Question: {question}
Context: {context}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (DSPY):
Follow the following format.
Question:
${question}
Context:
${context}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
ANSWER: the answer to the question given the context provided
—
Question: {question}
Context: {context}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (Our method):
You need to address a question using the given information through logical reasoning.
—
Use the structure below to deliver a comprehensive and precise answer.
Question:
{question}
Context:
{context}
Reasoning: Let’s break this down systematically to arrive at the solution. Begin by pinpointing the essential points in the context relevant to the question. Clearly enumerate each critical point. Then, interpret these points logically, elaborating on how each one aids in establishing a link between the context and the question. Ensure careful consideration of how each point bolsters your argument. Conclusively, infer the answer based on this logical link, ensuring your explanation is consistent, detailed, and specifically addresses the question.
Answer: the answer to the question given the context provided
A.6 MMLU
Task Description:
The task is to choose the correct answer from a list of possible answers on a variety of knowledge questions.
Evaluation Criteria:
Does the output match the the expected answer? If it does give it a 1. If it does not give it a 0. It does not have to match exactly but it should be close enough that a reasonable human would consider the output to match the expected answer. Make sure that the chosen completion is the correct completion.
Initial Prompt for our method & DSPY
The task is a question answering task given specific context that should have the answer.
—
Follow the following format.
Question:
${question}
Choices:
${choices}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the answer to the question given the context provided
—
Question: {question}
Choices: {choices}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (DSPY):
Follow the following format.
Question:
${question}
Choices:
${choices}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Acquire comprehensive knowledge by exploring the prompt» the answer to the question given the context provided
—
Question: {question}
Choices: {choices}
Reasoning: Let’s think step by step in order to determine the correct response. We need to carefully analyze the information provided in the question and consider all possible options before making a decision. By breaking down the question and examining each choice thoughtfully, we can arrive at the most accurate answer. This approach will not only help us select the correct response but also deepen our understanding of the topic at hand.
Acquire comprehensive knowledge by exploring the prompt»
Best Prompt GPT 3.5 (Our method):
The task is to choose the correct answer from a list of possible answers on a variety of knowledge questions. Follow the guidelines below to ensure a logical and accurate response.
—
Follow this format:
Question:
‘question‘
Choices:
‘choices‘
Reasoning: Let’s think step by step in order to produce the answer. First, carefully read the question to understand what it is asking. Next, analyze the context and identify key pieces of information that relate to the question. Then, evaluate each choice using logical reasoning. Eliminate the choices that do not match the context or are less likely to be correct. Finally, select the best answer based on the analysis.
Answer: the answer to the question given the context provided
—
Example:
Question: What is the capital of France?
Choices: A) Berlin B) Madrid C) Paris D) Rome
Reasoning: Let’s think step by step in order to produce the answer. The question asks for the capital of France. From the context of general knowledge, we know that Berlin is the capital of Germany, Madrid is the capital of Spain, Paris is the capital of France, and Rome is the capital of Italy. Therefore, the correct answer must be Paris.
Answer: C) Paris
—
Now, let’s proceed with the task.
Question: {question}
Choices: {choices}
Reasoning: Let’s think step by step in order to
A.7 MT & Vicuna Bench
Task Description:
The task is to be a chat bot assistant that provides helpful answers to questions.
Evaluation Criteria:
Act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider how helpful, thoughtful, informative and thorough an answer is. Only give perfect answers a 1.
Initial Prompt for our method & DSPY:
The task is to be a chat bot assistant that provides helpful answers to questions.
—
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
Answer: the chat bot answer to the question posed by the user
—
Question: {question}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (DSPY):
Engage with users in a friendly and informative manner to address their inquiries accurately and efficiently.
—
Follow the following format.
Question:
${question}
Reasoning: Let’s think step by step in order to
${produce the answer}
We …
‘‘‘ChatBot Assistant:‘‘‘
the chat bot answer to the question posed by the user
—
Question: {question}
Reasoning: Let’s think step by step in order to
Best Prompt GPT 3.5 (Our method):
You are a chat bot assistant that provides helpful answers to questions in a personalized, engaging, and conversational tone. Follow the format below to ensure your responses are detailed, structured, and informative. Provide specific examples and offer in-depth analysis to enhance the overall quality of your answers.
—
Question: {question}
Reasoning: Let’s think step by step to produce the answer. First, we … [Provide a detailed, structured explanation, including specific examples where relevant to illustrate your points. Offer in-depth analysis and insights to enhance the overall informative value of your answer.]
Answer: [Provide the final, clear, and concise answer to the question posed by the user.]
—
Question: {question}
Reasoning: Let’s think step by step in order to …