JEPA-PS turns optimizer memory into a measurable quantity: full trajectory pays on CartPole, nearly only the best point on rugged ground
Synopsis
Black-box optimizers are usually asked only whether they beat a baseline, rarely which part of the search history they actually use. JEPA-PS is pretrained with a reward-agnostic objective on random rollouts, learning only to predict the latent representation of the next trajectory point, and at test time reads a short evaluation history and proposes the next candidate in one forward pass. Across 15 control environments and 12 optimizers it does not beat tuned classical search, but ablations show history dependence varies by problem: full history beats best-only by about 22% on CartPole, while on MountainCar the two are nearly identical.
Interpretation
How much history a learned optimizer depends on is not a constant of the architecture but is elicited by the search problem. Classical methods fix memory by design: CEM keeps only the elite set, CMA-ES compresses history into a covariance matrix, and Bayesian optimization retains every observation. On CartPole the full history gains about 22% mean best reward over the best-seen point alone, while on MountainCar, Walker2d and Hopper that gain is near zero.
JEPA-PS separates what was tried from how well it worked, so the history can be subset, reordered or corrupted and the resulting proposal observed. Earlier learned optimizers mostly meta-learn an update rule against a reward-prediction or gradient-matching objective, which makes the role of history hard to isolate. Pretraining uses only parameter vectors and scalar returns, with a loss that predicts the latent representation of the next trajectory point; at test time there are no task descriptions, no reward shaping and no test-time gradients, and one forward pass yields a candidate.
History dependence is associated with fitness-distance correlation (FDC, a measure of whether random samples closer to the optimum tend to score better) in a small sample. The memory requirement of an optimizer had not been linked to an optimizer-independent descriptor of the landscape. Pearson correlation is 0.738 across seven environments and 0.832 for the ordering-sensitivity probe on a five-environment subset; removing CartPole lowers the Pearson value to about 0.50, the Spearman correlation on the same seven environments is 0.571, and the authors treat the association as preliminary.
Perspective
The demonstrated scope is control environments under linear policies, with the main comparison covering 15 environments and 12 optimizers and transfer tests limited to CartPole physics variants; the authors state that JEPA-PS is not a general-purpose optimizer and do not propose an optimizer that exploits the landscape-memory association.
The landscape association rests on seven environments and is strongly influenced by CartPole, which the authors call preliminary; an independent probe based on reward-ordering sensitivity does not corroborate it; and on floor-limited tasks such as MountainCar a near-zero history gain may also mean that neither condition makes progress.
