Research
Working papers
-
Partner UED for cooperative multi-agent learning.
We treat unsupervised environment design as a problem over partner policies in cooperative multi-agent learning, rather than over level layouts. The adversary samples co-players from a learned population, and the ego learns to coordinate with whatever partner it draws. We compare planning-based and policy-gradient adversaries against a domain-randomization baseline, then ask whether the limit in this regime comes from the search procedure or from the partner-pool formulation itself. The answer should not depend on which cooperative benchmark we run on.
-
Finite-size scaling of Schelling-type segregation.
We perform a finite-size scaling study of the Schelling segregation model and its multi-radius generalizations. We measure critical exponents at four radii on lattices up to 2048 x 2048, verify universality across radii, and demonstrate hyperscaling through data collapse. The result establishes a noise floor for the order parameter and clarifies which earlier claims survive at large n.
-
Adaptive sensing driven by prediction error.
We decompose the prediction error of a frozen-encoder world model into three additive terms: model floor, online noise gain, and event mass, and gate an expensive classifier on the residual. The detector ties a constant-velocity Kalman on bouncing balls and reaches 94 percent of every-frame accuracy at one-quarter compute on tracking clips, with the online noise estimate generalizing to real Big Buck Bunny scene cuts.
Publications
Forthcoming.