Microcosmos's evolved chemotaxis is fitness-rewarded via an explicit energy term, not spontaneously emergent

artificial-lifeneuroevolutionreward-shapingquality-diversitymicrocosmosemergence-vs-reward

Microcosmos's Experiment 3 (NEAT/CPPN neuroevolution) is sometimes summarized — including by the vault's own earlier paraphrase in claim-microcosmos-four-experiments-handdesigned-to-emergent — as producing "emergent chemotaxis." A direct read of the neuroevolution experiment (Experiments §4.3, "Neuroevolution of Controller Circuit") narrows that. The baseline run selects on displacement alone — "the fitness function is defined as the total displacement traveled" — and finds that "the highest fitness swimmers predominantly perform sinusoidal locomotion," with no food-seeking. A second run adds a food-like energy field and changes the objective directly: "fitness becomes the total displacement traveled plus the total energy collected, rewarding both movement and chemotaxis." Chemotaxis appears only once the fitness function is rewritten to pay for it.

The distinction matters for what "automatically" means in the paper's own abstract ("produce a wide range of swimming and chemotaxis behaviors automatically"): evolutionary search discovers the strategy — which specific movement pattern collects energy — without a human hand-coding it, but the objective itself was authored by the experimenters to reward energy-seeking. This is reward-shaping, the mechanism by which any RL-adjacent system is steered toward a target behavior, not an instance of that behavior appearing unprompted under a neutral objective.

This corrects a specific point within claim-microcosmos-four-experiments-handdesigned-to-emergent (see that note's Correction history). See also claim-microcosmos-elastic-filaments-differentiable-gpu-fluid for the differentiable substrate this evolutionary search runs on, and entity-kenneth-o-stanley for the NEAT/CPPN lineage the experiment reuses.