In Diversity-Driven Search, Predicting Survival Does Not Improve Selection

Admission is only the first hurdle

Quality-diversity search aims to collect many different, high-performing solutions rather than identify a single winner. That makes it relevant to problems such as discovering distinct ways a system can fail. But an immediate contribution to a collection may not remain useful as the search continues.

In their November 2026 preprint, Corrado Calamita, Matteo Mendicini, Luca Ferrari and Riccardo Bertamini investigate that distinction in Dominated Novelty Search, or DNS. They ask whether information available before a candidate has reproduced can identify parents whose offspring will remain in the archive. Their answer separates prediction from intervention: some signals distinguish lasting contributions, but the tested preferences based on those signals do not improve search.

The experiments use standard optimization and simulated-robotics benchmarks, not a direct evaluation of automated failure discovery.

An archive with recurring competition

DNS maintains a finite collection without dividing behavior space into fixed cells. Each update pools existing archive members with new offspring. Candidates receive a competition score based on their average behavioral distance to their three nearest neighbors with equal or higher objective fitness, using fewer neighbors when necessary. A candidate with no qualifying neighbor receives an infinite score. DNS retains the highest-scoring candidates up to capacity.

This favors behavioral separation relative to equally good or better alternatives. Because the entire pool is ranked again at every update, admission offers no permanent protection: later candidates can displace an earlier contribution.

The researchers distinguish immediate entry from persistence 1,000 generations after an offspring’s creation. Their measurements use uniform parent sampling, allowing them to examine associations before changing selection policy.

They test a planar arm with 128 joint angles, an omnidirectional simulated ant and a simulated two-dimensional walker. The arm’s descriptor records its endpoint; the ant’s records final position; the walker’s records foot contact. Both locomotion tasks use Brax spring physics and 1,000-step episodes.

Measurement runs last 10,000 generations, producing 256 offspring per generation with the same Iso+LineDD variation operator. Archive capacities are 256, 1,024 and 4,096, with five, ten and five random seeds respectively for each task. Statistical analysis treats runs, not millions of correlated parent–offspring events, as the units of evidence.

The same history can point two ways

The clearest signal is lineage depth: the number of reproduction steps from a founder along the base-parent chain. Unlike offspring-success feedback, depth is available before a candidate first reproduces.

At capacity 1,024, shallower parental lineages are associated with immediate offspring entry, while deeper lineages are associated with offspring persistence. Several other available signals also reverse direction between the outcomes.

The paper measures discrimination using AUC, where 0.5 indicates no ranking ability and values below 0.5 indicate an association in the opposite direction. Lineage depth’s persistence AUC increases with archive capacity on all three tasks. On the ant benchmark, its distance from chance rises from 0.097 to 0.198 between capacities 256 and 4,096.

Larger archives turn over more slowly when offspring production stays fixed. However, capacity also changes competition geometry, attainable coverage and the ratio of offspring to slots. The results therefore support an association consistent with a turnover explanation, not an isolated causal effect of turnover.

A predictive signal is not a successful policy

The intervention experiments replace uniform base-parent sampling with a rank-based preference controlled by temperature: lower temperatures create stronger bias.

On the ant benchmark, fitness-based selection is tested at capacities 1,024 and 4,096 for 4,000 generations over ten seeds. Moderate bias matches uniform DNS. At capacity 4,096, the reported projected QD-score is approximately 4.04 million versus 4.02 million for uniform selection. Strong bias instead reduces the score by 64.4% at that capacity, compared with 43.6% at capacity 1,024.

The higher-capacity loss accompanies projected coverage falling from 0.9999 to 0.517 and lower mean novelty, while maximum fitness remains essentially unchanged. The authors interpret this as lost diversity rather than lost peak quality.

Depth-based selection also fails to outperform uniform sampling in ten-seed ant experiments at capacity 1,024 over 10,000 generations. Favoring deeper lineages, weakening that preference or reversing it produces differences within between-seed variation. Uniform coverage is about 0.773, so the null result is not explained by already-complete coverage.

What the findings establish

The study supports measuring admission and retention separately. It does not establish that lineage depth causes durable offspring, or that predictive rankings automatically provide useful search guidance.

Its scope is DNS, one variation operator and three predefined benchmark domains; the parent-selection interventions are confined to the ant task. Adaptive archive capacity remains a proposed research direction, not a demonstrated improvement. The practical lesson is narrower: under these conditions, an informative association was insufficient to beat uniform parent sampling.