Diverge-converge cycles in LLMs
The Double Diamond is, roughly, a design framework consisting of 4 steps:
- Diverging on problem (Discover). Explore widely to gather a broad range of insights and challenges related to the problem.
- Converging on problem (Define). Analyze and synthesize the gathered insights to define a clear and specific problem statement.
- Diverging on solution (Develop). Brainstorm and generate a diverse array of possible solutions for the identified problem.
- Converging on solution (Deliver). Evaluate and refine the proposed solutions to select the most effective and viable one.
You can prompt an LLM to go through these steps to solve a very briefly formulated problem. This potentially increases the quality of solutions, but more importantly, makes the decision process very explicit and human-interpretable.
For the divergent steps, I set a higher value for the temperature
parameter, which should make the sample more diverse.
I am generally a very convergent thinker, so seeing the diverging steps laid out explicitly really expands the set of options I even consider.
I'd love to see an eval of this compared to a simple Chain-of-thought prompt.