When the Chain of Thought and Tree of Thought papers came out, I experimented with a few types of collective intelligence and cognitive strategies in prompt engineering.
I thought they might work to improve the thoughtfulness and diversity of responses.
I also found that simulating multiple perspectives often led to an increased ability to catch and self-correct errors—even in zero-shot prompting, without the need to chain multiple prompts.
Wisdom of the crowd
1. The crowd within
When you do a task, then come back to it some time later, you often get a new perspective simply from letting having the problem marinate in your head subconsciously.
This is ‘the crowd within’, when it seems like you almost have different people in your head generating multiple independent thoughts.
It would likely work well for estimating unknown numbers, if you’re trying to predict something, and decision-making.
Prompt
We're going to use the "crowd within" theory and dialectical bootstrapping to answer a question. This technique leverages collective intelligence and involves participants providing an initial estimate, then refining their estimate by considering potential errors and alternative information. This method aims to increase the accuracy of the final answer through diverse perspectives. For the following question, imagine you're asking a diverse group of people. For each person, provide: 1. Their initial estimate based on their current knowledge. 2. Their revised estimate using dialectical bootstrapping, after a delay of three weeks: - Assume the initial estimate might be wrong. - Consider what information could have been misinterpreted or missed. - Reflect on how this alternative information would change the estimate. - Decide if the initial estimate was an overestimate or an underestimate. - Based on this new perspective, provide their revised estimate. 3. After simulating at least 10 responses, determine the consensus answer - the one that takes into account both initial and revised estimates, aiming for increased accuracy through diverse perspectives.Each participant should provide their initial estimate and then use dialectical bootstrapping to refine their estimate. The question is: [INSERT QUESTION HERE] Here is a sample process: <sample> Question: In what year was the telephone invented? | Person | Initial Estimate | Revised Estimate (Dialectical Bootstrapping after 3 weeks) | |--------|------------------|---------------------------------------------------------------------| | 1 | 1900 | 1890 (Considered the industrial revolution and possible earlier advancements) | | 2 | 1885 | 1875 (Factored in earlier technological innovations missed initially) | | 3 | 1920 | 1910 (Adjusted for historical events that likely sped up discovery) | | 4 | 1895 | 1885 (Revised considering missed scientific milestones) | | 5 | 1910 | 1900 (Adjusted for overlooked historical context) | | 6 | 1870 | 1860 (Factored in underestimated early research and discoveries) | | 7 | 1880 | 1870 (Adjusted considering possible earlier indications) | | 8 | 1930 | 1920 (Considered the potential earlier influences) | | 9 | 1850 | 1840 (Reviewed overlooked data and historical errors) | | 10 | 1940 | 1930 (Adjusted for historical research trends) | </sample> Now, let's apply this method to the following question. Follow these guidelines: 1. Ensure each estimate is thoughtful and considers potential biases. 2. Clearly document both the initial and revised estimates, and explain why. 3. Strive for accuracy through diverse perspectives and critical evaluation. Again, the question is: [INSERT QUESTION HERE]
2. Surprisingly popular
This is an extension of the wisdom of the crowd, based on a finding by scientists at MIT's Sloan Neuroeconomics Lab to improve the results.
It can be used as a chained next step after The Crowd Within or as a standalone step.
Prompt
For the following question, imagine you're asking a diverse group of people. For each person, provide: 1. Their answer to "What do you think is the correct answer?" 2. Their prediction for "What do you think will be the most popular answer?" After simulating at least 10 responses, determine the "surprisingly popular" answer - the one with the largest positive difference between its actual popularity and predicted popularity. The question is: [INSERT QUESTION HERE]
Problem-solving psychology techniques
These methods tap into different cognitive processes, which I've found can also lead to more thoughtful, inductively-reasoned answers.
1. Revealed preference
By creating a persona(s) and have them ‘choose’ between products and explain their choices, it helps understand decision-making processes and user/buyer behavior.
Prompt
Construct a persona of [SPECIFIC BUYER TYPE]. Given the following product options and their prices, have the buyer choose which product they would purchase and explain their decision in detail. Consider their income level and how it affects their choice. Products: 1. [PRODUCT A] - $[PRICE A] 2. [PRODUCT B] - $[PRICE B] 3. [PRODUCT C] - $[PRICE C] Buyer's annual income: $[INCOME]
2. System 1 and 2 multiple agent thinking
System 1 is your gut reaction—fast and automatic, while System 2 is the slow, thoughtful process. It’s somewhat like the wildly popular ‘let’s think step by step’ prompt, but contrasting the two can often force a more logical answer.
I had some (messy) evals for this which I’ll add at an indeterminate time later.
Prompt
Imagine you are thinking from both a System 1 and System 2 perspective. System 1 operates automatically and quickly, with little or no effort and no sense of voluntary control. System 2 allocates attention to effortful mental activities that demand it, including complex computations. Write down both System 1 and System 2's thought processes to the following question, and then reconcile them: [INSERT QUESTION HERE]
3. Fresh perspective approach
Another take on Systems 1 and 2 thinking. Sometimes, you just need to step back to give yourself a different aperture, and you can simulate that with LLMs.
Prompt
1. Provide an initial response to the following problem: [INSERT PROBLEM HERE] 2. Now, imagine you've stepped away from this problem for a week. With this fresh perspective, revisit the problem and provide a new analysis. 3. Compare and integrate both perspectives, highlighting any new insights gained from the "fresh" approach.
More nuanced, thoughtful responses are always useful. If you try any of these, please let me know your results.