MYND Prompt Phylogeny: A Tree of Prompts, and What Is Still a Placeholder
Two 2023 papers treat prompts as things that evolve. EvoPrompt by Guo and colleagues connects language models with evolutionary algorithms to optimise prompts (arXiv 2309.08532). Promptbreeder from Google DeepMind evolves a population of prompts and also evolves the mutation prompts that change them (arXiv 2309.16797). My repo MYND Prompt Phylogeny asks a related question about history: given a set of prompts, can you draw how they descend from each other?
What is implemented
backend/src/algorithms/phylogenyEngine.ts has a fingerprinter that turns an embedding vector into a short hash with random projections, a locality-sensitive hashing idea, and compares two fingerprints by Hamming distance. The projections are seeded from the index, so the same input gives the same fingerprint. A distance matrix mixes cosine distance with a semantic term.
What is placeholder
The tree builder is labelled UPGMA, but a comment inside it says it is a simplified clustering and that production should use proper UPGMA or neighbour joining. In practice it groups by version and parent links. The fitness function starts from a base score tied to the version number and adds Math.random noise to accuracy, latency, token efficiency and user rating. Those numbers are not measurements.
Why it matters
A tree drawn from random scores looks the same as one drawn from real evaluations, which is the risk. Until the fitness comes from a test set, the picture shows lineage, not quality. Neither paper uses such scores; both evaluate prompts on task benchmarks.
My opinion
My view, not a fact from the code: the lineage view is useful on its own, and the fix is small. Replace the random fitness with a real evaluator, call the clustering what it is, and the project becomes a way to inspect a population like the ones in those papers.
Sources
Guo et al., EvoPrompt: arxiv.org/abs/2309.08532. Fernando et al., Promptbreeder, ICML 2024: proceedings.mlr.press/v235/fernando24a.html. Code: github.com/yethikrishna/mynd-prompt-phylogeny, phylogenyEngine.ts, read on October 10, 2026.