An AI system got much better by evolving a whole population of tweaks, but was the population really the secret?

This explores how much of DarwinX's improvement comes from keeping a population of harness variants, as opposed to evolving one lineage. The corpus can't answer that directly. What it has on DarwinX is the headline result: about 17 points of average gain across benchmarks, Terminal-Bench 2.1 rising to 84.7% and WebArena-Infinity reaching 93.0% audit-clean pass@1. It gets there by selecting over a population of harness variants (prompts, tools, skills, control flow) while the model stays frozen Can frozen models improve by evolving their harnesses?. That is a comparison against an un-evolved harness. The material provided doesn't split out how much of the gain is the archive and how much is simply searching at all.

The corpus does say what a fair answer would need. Harness-evolution gains have to be measured against task-level test-time search given the same feedback and the same inference budget, and only the surplus beyond that counts as the harness design's contribution How should we measure gains from automatic harness evolution?. A single lineage is the simplest such baseline. A population also evaluates many more candidates, so archive benefit and extra compute are easy to confuse. The archive's true share is the gap between population and single-lineage search at equal budget. Nothing here says DarwinX ran that comparison, so it's worth checking the paper's ablations before quoting a number.

Two neighbouring results suggest why the answer may not be a clean split. One shows accumulated discovery trees being replayed off-policy as a cheap simulator for scoring exploration policies Can past discoveries train better exploration policies?. In that work, a retained history is a resource a single lineage throws away. The other trained a model on program-evolution operators (Draft, Improve, Debug, Crossover) and found that learning and search gains added together instead of one explaining the other Can training and search gains add together in program evolution?. That result concerns training versus search, not archive versus lineage. It still warns that when components are complementary, asking which one gets the credit can be the wrong question.

There's a third possibility: some of the gain comes from the harness itself, however it was found. A 20B model with a stateful harness that externalized bookkeeping matched frontier models and beat the next open searcher by 11.4 points Can externalized bookkeeping let smaller search agents beat larger ones?. If a good harness carries that much capability, a single lineage might land on much of DarwinX's gain, and the archive's role would be reaching the last stretch reliably. The corpus can't tell you whether that's true. It gives you the right control (matched-budget single-lineage search) and a reason to expect the archive's contribution to be real but entangled with everything else.