Joint modelling of choice and reaction time for representation learning: an information ceiling

Abstract

Analysis code and tracked results asking whether modelling reaction time alongside choice recovers better item representations than modelling choice alone, in odd-one-out similarity judgement. Includes likelihoods and exact gradients for the LBA, the Racing Diffusion Model and an Advantage Racing Diffusion drift map; a cross-validated model comparison against a regularized choice-only baseline; posterior predictive checks; and a Fisher-information ceiling on how much the reaction-time channel can contribute to the learned representation at all. The headline result is that the ceiling sits below the observed cost of extraction at every training-set size where the comparison is decided, and that a correctly coupled race closes the joint model's deficit to parity without crossing it. PROVENANCE: the analyses postdating the CogSci 2024 paper were carried out with Claude Code, an AI coding agent, under the author's direction; the author has not independently verified the results in detail. The likelihood implementations are validated against arbitrary-precision references, quadrature, simulators and finite differences, and every grid reproduces from recorded seeds — but the reasoning layer has had no independent review. See the "Provenance, and how much of this has been checked" section of the README for what is and is not verified. Reproduce rather than trust.