Chattarin Poungtubtim¹, Timothy F. Brady¹, John T. Serences¹; ¹Department of Psychology, UCSD

Representational drift is a phenomenon in which neural activity associated with the same stimulus gradually deviates from its initial pattern over a period of days and weeks, even after behavioral performance has stabilized. While the cause and functional role of this drift is still debated, recent studies have shown that representation drift is correlated with sparser and more separable patterns of neural activity (Ratzon et al., 2024l Kumar et al., 2025). Thus, drift might result from neural networks moving toward more efficient, sparser solutions that still support adequate task performance. In turn, as the separability of neural representations increases, any noisy perturbations of the representations will be less likely to cause confusion. Therefore, representational drift might also drive networks towards more stable and robust solutions. To test this hypothesis, we trained continuous time recurrent neural networks (ctRNNs) on a visual working memory task that required encoding and storing stimuli over a brief delay period. To ensure stimulus encoding independent of the required response, we presented a cue after the delay period indicating the stimulus-response mapping, and the model produced a recall response indicating the identity of the remembered stimulus. We observe that models drift further from their initial states even as performance remains stable: representational drift. As they drift, their activity becomes more sparse, and has greater separation between different memory representations. Crucially, we show that post-drift network configurations are more robust to increased stimulus noise, internal noise and increases in the duration of the delay period. Overall, we provide theoretical evidence that representational drift can play a functional role in making task representations become more stable, robust, and energy efficient over time.