Diversity is productive
Quality-Diversity methods preserve high-performing alternatives across niches. The positive lesson for model breeding is direct: do not discard a specialist just because it is not the global champion. It may be the best model for a low-memory device, a rare language, an adversarial edge case, a jurisdictional constraint, or an offline classroom.
This turns diversity into an innovation engine. A future descendant may need exactly the capability that today's single-score leaderboard would have thrown away.
Niche examples
| Niche | Elite artifact |
|---|---|
| Lowest p95 latency | Tiny quantized classifier. |
| Best local privacy | Browser-only summarizer. |
| Best rare-domain accuracy | Specialist adapter. |
| Best abstention behavior | Conservative judge. |
| Best teaching utility | Tutor that explains steps well. |
FUNCTION update_quality_diversity_archive(candidate, archive)
niche = map_to_niche(candidate.descriptors)
incumbent = archive[niche]
IF incumbent IS NULL OR candidate.fitness > incumbent.fitness
archive[niche] = candidate
record_elite_replacement(niche, candidate, incumbent)
END IF
RETURN archive
END FUNCTIONPositive result
A QD archive makes the system richer over time. Instead of one winner and a pile of forgotten experiments, the ecology develops a cabinet of proven options.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.