When an AI evolves ideas over many rounds, does how it picks which ideas to build on matter most?

This explores which ways of choosing the 'parent' ideas that seed the next round of hypotheses in an evolutionary-style search produce the best hypotheses. The corpus has no head-to-head test of selection rules (top-scorer only, tournament, diversity-weighted and so on). It does have several notes on the machinery around selection, and they point in a consistent direction.

The closest match is HypoEvolve, which treats hypothesis development as a generational genetic algorithm. Its main move is to separate what each agent does scientifically from how the group coordinates, so each collaboration rule can be named, changed and compared (How do collaboration rules shape hypothesis quality?). It beat six baselines on drug repurposing. But the note describes the framework making such comparisons possible, and doesn't report which rule won. Mind Evolution is the other evolutionary example. It uses LLM-written crossover and mutation across separate 'islands' of candidates, and it beats best-of-N and sequential revision, solving over 98% of planning tasks (Can evolutionary search beat sampling and revision at inference time?). Islands keep sub-populations apart so one early leader can't take over the whole pool. That makes diversity, and not just picking the top scorer, the likeliest lever on selection. The note credits the whole design, though, and doesn't isolate islands as the cause.

A selection rule can only be as good as the score beneath it, and this is where the corpus is most direct. Language models are good at proposing valid candidates but can't reliably judge how good those candidates are or how uncertain the judgment is (Can language models reliably judge their own candidate quality?). If the model rates its own hypotheses and the best-rated ones become parents, you risk breeding from confident-sounding but weak ideas. The proposed fix is to pair the model with a surrogate fitted to real experimental data, which gives selection an uncertainty estimate to work with.

Other notes describe ways to get a sturdier score. Tree search ranks solution paths by how they turn out, which produces quality signals without human labels or the model's own opinion (Can tree search replace human feedback in LLM training?). Spark-to-Paper isolates model judgment from deterministic, executable checks and asks for the evidence specification before results are seen (Can separating judgment from verification improve research paper reliability?). An agentic reviewer that spends extra compute checking proofs line by line caught errors that passed human review (Can inference scaling help reviewers catch errors humans miss?). Any of these could serve as the fitness judge in a hypothesis search, though that pairing is an inference from the notes, not something they test.

So the library's answer is that no parent-selection strategy is shown to win. The evidence favors two things: keep the pool diverse, as island models do, and choose parents using an external or verifiable score, not the generator's self-assessment. The missing experiment is the one that holds everything else fixed and varies only how parents are picked.