Work

History dependence and individual predictive differences in rat working memory

Year: 2026
Type: preprint
Abstract: Recent stimuli bias judgments in working memory. Such effects can arise when a noisy memory is combined with a prior or replaced by a past stimulus. We compared these accounts using ordered rat choices, forward prediction and model recovery. Inference used 15 analysis units after confirmed duplicate sessions were removed. History improved prediction over a fixed prior. Substitution beat the Boboeva fixed-marginal Bayesian augmented with lapse and bias, but had no overall advantage over a running-prior Bayesian. Which model predicted a unit better was consistent across interleaved halves of its sessions: 15/15 units kept the same preferred model with fixed fits and 14/15 with independent half fits (Spearman ρ=0.96 for both). The preferences largely survived refits that matched the lapse rate (12/15 units) or added slow choice drift (15/15). They did not persist between early and late halves of training (7/15; ρ=0.41). In 20 simulations from one fitted hybrid, changing only the trial design gave a central 90% range of 0.70 to 0.87 for ρ. The lower observed correlation suggests that preferences changed across training beyond what this design-only simulation predicted, with no consistent direction of change. Both models failed at least one predictive check in 12 of 15 units. Three-way recovery with a hybrid failed, and the tested alternative designs did not separate all families. The evidence supports history dependence and unit-level differences in which model predicts better, consistent within a period of training but changing across training. It does not select substitution as the general explanation, and neither model fully reproduces the behavior.
Language: English
Field: Psychology
Open Access status: green
OpenAlex ID: W7236697262
Cites:
Cited by:
Related to: 10