Ask Me Anything Prompting (AMA) is a novel strategy for enhancing the capabilities of large language models (LLMs). This approach, which methodologically collects multiple prompts and aggregates their responses, addresses the brittleness of single-prompt strategies and moves beyond the need for meticulously crafted prompts. It has proven to significantly improve task performance across various model types and sizes, enabling smaller, open-source LLMs to reach or surpass the performance levels of larger models like GPT-4.

Ask Me Anything: A simple strategy for prompting language models
Large language models (LLMs) transfer well to new tasks out-of-the-box simply given a natural language prompt that demonstrates how to perform the task and no additional training. Prompting is a brittle process wherein small modifications to the prompt can cause large variations in the model predictions, and therefore significant effort is dedicated towards designing a painstakingly “perfect prompt” for a task. To mitigate the high degree of effort involved in prompt-design, we instead ask whether producing multiple effective, yet imperfect, prompts and aggregating them can lead to a high quality prompting strategy. Our observations motivate our proposed prompting method, ASK ME ANYTHING (AMA). We first develop an understanding of the effective prompt formats, finding that question-answering (QA) prompts, which encourage open-ended generation (“Who went to the park?”) tend to outperform those that restrict the model outputs (“John went to the park. Output True or False.”). Our approach recursively uses the LLM itself to transform task inputs to the effective QA format. We apply the collected prompts to obtain several noisy votes for the input’s true label. We find that the prompts can have very different accuracies and complex dependencies and thus propose to use weak supervision, a procedure for combining the noisy predictions, to produce the final predictions for the inputs. We evaluate AMA across open-source model families (e.g., EleutherAI, BLOOM, OPT, and T0) and model sizes (125M-175B parameters), demonstrating an average performance lift of 10.2% over the few-shot baseline. This simple strategy enables the open-source GPT-J-6B model to match and exceed the performance of few-shot GPT3-175B on 15 of 20 popular benchmarks. Averaged across these tasks, the GPT-J-6B model outperforms few-shot GPT3-175B. We release our code here:

Let's break down the paper and see how we can apply this to ChatGPT or similar.


  • Introduction of AMA Prompting:
    • AMA Prompting revolutionizes the use of LLMs by collecting and aggregating multiple prompts, thereby enhancing model capabilities.
    • This approach effectively addresses the limitations of single-prompt strategies, reducing the necessity for perfect prompt design.
  • Effective Prompt Formats:
    • Research indicates that open-ended question-answering prompts are more effective than restrictive ones.
    • Utilizing this insight, AMA transforms task inputs into these more effective question-answering formats.
  • Scalable Prompt Collection:
    • AMA's scalable strategy reformulates task inputs into effective formats.
    • This involves leveraging the LLM itself to convert inputs into questions and generate corresponding answers.
  • Weak Supervision in Aggregation:
    • AMA applies weak supervision to combine the noisy predictions from various prompts.
    • This method accommodates the differing accuracies and interdependencies among the prompts.
  • Performance Across Model Families and Sizes:
    • AMA has shown consistent performance improvements, with an average lift of 10.2% over the few-shot baseline across diverse LLM families and sizes, including EleutherAI, BLOOM, OPT, and T0 models.
  • Comparison with Few-Shot GPT-3:
    • AMA has enabled smaller models like GPT-J-6B to outperform the few-shot GPT3-175B model in certain benchmarks.
  • Challenges and Limitations:
    • AMA encounters challenges in tasks requiring deep domain knowledge or dealing with temporally variable answers.
    • The strategy may be limited in tasks that cannot fully utilize the latent knowledge embedded within the model.
  • Reproducibility and Ethics:
    • The researchers have made AMA prompting code available to ensure reproducibility.
    • They recognize potential ethical risks in using AMA and advocate for responsible usage.
  • Acknowledgements and Funding:
    • The project was supported by DARPA, NIH, NSF, and other funding bodies.
    • Computational resources were provided by Together Computer, Numbers Station, and other contributors.

Ask Me Anything Prompting Strategy In-Depth

Overview of AMA Prompting

  1. Multi-Prompt Strategy: Unlike traditional methods that rely on a single prompt, AMA prompting uses multiple prompts for a single task. This diversity in prompts aims to capture a wider range of perspectives or interpretations of a task, leading to more robust and comprehensive responses.
  2. Aggregated Responses: The responses generated from these multiple prompts are then aggregated. This aggregation is crucial as it combines the insights from various prompts, thereby mitigating the risks of errors or biases that might be present in a single response.

Addressing Brittleness and Prompt Design

  1. Brittleness of Single-Prompt Strategies: Traditional single-prompt approaches are often brittle, meaning small changes in the prompt can lead to significant variations in the output. This brittleness can limit the practical usability of LLMs, as it requires precise and often complex prompt engineering to get the desired results.
  2. Reducing the Need for Perfect Prompt Design: Designing the perfect prompt can be a time-consuming and challenging process, often requiring iterative testing and refinement. AMA prompting alleviates this burden by using multiple imperfect prompts, each contributing to the final aggregated output. This approach inherently accepts and utilizes the imperfection in individual prompts.

Example of AMA Prompting

Let's consider a task where the model is asked to analyze the sentiment of a movie review. Instead of using a single prompt, AMA prompting would involve multiple prompts, each phrased differently to analyze sentiment. For example:

  • Prompt 1: "Read the following movie review. Is the sentiment expressed positive, negative, or neutral?"
  • Prompt 2: "Given this movie review, would you say the reviewer enjoyed the movie? Why?"
  • Prompt 3: "Summarize the tone of the movie review. Does it lean more towards positive, negative, or is it mixed?"

Each prompt approaches the task differently - directly asking for sentiment, inferring enjoyment, and requesting a summary of tone. The responses to these prompts are then aggregated to derive a more nuanced and accurate understanding of the review's sentiment.

AMA Prompting, by leveraging multiple prompts and their aggregated responses, offers a robust alternative to traditional single-prompt strategies. It not only addresses the brittleness associated with these traditional methods but also significantly reduces the pressure of crafting the perfect prompt, making LLMs more accessible and effective for a variety of tasks.

Ask Me Anything Prompting Example with ChatGPT

To illustrate how Ask Me Anything (AMA) Prompting can be integrated with ChatGPT, let's consider a detailed example. The process involves generating multiple prompts from a single user query, getting responses to these prompts, and then intelligently aggregating these responses into a cohesive answer.


Suppose a user asks ChatGPT: "Can you explain the causes and effects of climate change?"

Step-by-Step Implementation of AMA Prompting with ChatGPT:

Step 1: Generating Multiple Prompts

  • Action: Break down the user's question into multiple sub-questions or prompts, each targeting a specific aspect of the main question.
  • Example Prompts:
    1. "What are the primary natural causes of climate change?"
    2. "How do human activities contribute to climate change?"
    3. "What are the major environmental impacts of climate change?"
    4. "How does climate change affect human societies?"

Step 2: Generating Responses for Each Prompt

  • Action: ChatGPT generates answers for each of these prompts, treating them as individual queries.
  • Example Responses:
    1. Response to Prompt 1: Talks about natural factors like volcanic eruptions, solar radiation variations, etc.
    2. Response to Prompt 2: Discusses human contributions like fossil fuel combustion, deforestation, etc.
    3. Response to Prompt 3: Describes environmental impacts such as rising sea levels, increased extreme weather events, etc.
    4. Response to Prompt 4: Explains effects on human societies like migration due to sea-level rise, agricultural disruptions, etc.

Step 3: Aggregating Responses

  • Action: The responses are then aggregated to form a comprehensive answer. This involves evaluating the relevance and accuracy of each response and synthesizing them.
  • Method: Use principles of weak supervision to weigh the reliability of each response and merge them into a single, coherent narrative.

Step 4: Presenting a Unified Answer

  • Action: Present the synthesized answer to the user in a coherent and structured format.
  • Example Unified Answer: "Climate change is driven by a combination of natural factors such as volcanic activities and variations in solar radiation, and human activities, notably the burning of fossil fuels and deforestation. Its impacts are far-reaching, affecting the environment through rising sea levels and increased frequency of extreme weather events. Human societies are also profoundly affected, facing challenges such as displacement due to rising sea levels and disruptions in agricultural productivity."

Explanation of Benefits and Effectiveness:

  1. Comprehensive Understanding: This approach allows ChatGPT to cover the question's various dimensions, leading to a more thorough understanding.
  2. Reduced Bias and Error: Aggregating multiple responses mitigates the risk of bias or errors present in individual responses.
  3. Enhanced User Experience: The final answer provides a more detailed and nuanced explanation, likely leading to greater user satisfaction.

AMA Prompting, when applied to ChatGPT, enhances its ability to dissect complex queries into manageable parts and combine the insights gained into a robust, well-rounded answer. This method not only improves the depth and breadth of ChatGPT's responses but also enriches the user's experience through more informative and comprehensive answers.

Scenarios for Utilizing AMA Prompting

AMA Prompting, with its multiple prompt aggregation approach, can be highly effective in various scenarios. However, it's crucial to recognize situations where its application is most beneficial and where it might not be the optimal choice.

When to Use AMA Prompting:

  1. Complex or Multifaceted Questions: For inquiries that cover multiple aspects or require a nuanced understanding, AMA Prompting is ideal. It can dissect the question into smaller, more manageable parts, ensuring a comprehensive and detailed response.
  2. Situations Requiring Balanced Perspectives: In scenarios where a balanced view is essential, such as in discussions involving ethical considerations or multiple viewpoints, AMA Prompting can aggregate diverse perspectives to provide a well-rounded response.
  3. Learning and Educational Contexts: When used in educational settings, AMA Prompting can enhance understanding by breaking down complex topics into simpler sub-questions, making it easier for learners to grasp intricate subjects.
  4. Research and Analysis: In research scenarios where thoroughness is key, AMA can gather varied information on a topic, ensuring that the response is detailed and covers all necessary angles.

When Not to Use AMA Prompting:

  1. Simple or Direct Questions: For straightforward questions that require direct answers, AMA Prompting might overcomplicate the response, making it less efficient and potentially confusing.
  2. Time-Sensitive Situations: In scenarios where quick responses are critical, such as in real-time assistance or emergency situations, the time taken to generate and aggregate multiple prompts may not be practical.
  3. Highly Specialized or Niche Topics: If a question pertains to a very specialized field where expert knowledge is required, AMA Prompting might not always provide the depth and accuracy needed, unless the prompts are specifically tailored by an expert in that field.
  4. Limited Data Environments: In cases where there's limited information available on a topic, AMA Prompting may struggle to generate multiple relevant prompts, leading to responses that are not significantly different or insightful.
  5. Highly Personalized Responses: For questions that require personalized responses, such as in therapy or counseling, AMA's generalized approach might not be suitable. Personalized interactions often require empathy and a deep understanding of individual circumstances, which might not be effectively captured through multiple prompts.

The choice to use AMA Prompting should be driven by the nature of the inquiry and the context in which the response is needed. While it excels in providing thorough, multifaceted answers to complex questions, it is less suitable for simple queries, urgent responses, highly specialized topics, or situations demanding personalized interaction. Understanding these nuances ensures the effective and appropriate application of AMA Prompting.

Ask Me Anything (AMA) Prompting offers an effective strategy for improving the capabilities of large language models like ChatGPT. By generating multiple prompts targeting different aspects of a query and aggregating the responses, AMA Prompting provides more comprehensive, robust, and accurate answers.

Integrating this technique into ChatGPT enhances its ability to break down complex questions, address them from various perspectives, and synthesize the insights into an informative unified response.

Though not without limitations, AMA Prompting significantly boosts model performance across diverse tasks and models, enabling even smaller open-source LLMs to match or exceed the few-shot capabilities of larger proprietary models. Its simplicity, scalability, and reliability make AMA Prompting a promising prompt engineering paradigm for unlocking more of the latent potential within foundation models.

Share this post