Expecting temporal stability reconciles the serial dependence paradox

<p>Serial dependence, the bias of perceptual and cognitive judgements toward recent experience, is ubiquitous yet contested, attributed either to sensory merging or to biased decisions. A central puzzle is that neuroimaging consistently finds neural representations repelled from recently seen features even as behavioural reports are attracted toward them. We propose that both follow from a single principle: an expectation that input is temporally stable, a sensible prior given the autocorrelation of the sensory diet. This expectation acts opposingly at two stages. At encoding it dampens neurons tuned to the predicted feature, which repels neural representations from it; while at readout a preparatory bias toward that feature produces behavioural attraction. Because these two processes are parallel, rather than sequential, the framework reconciles these opposing signatures and explains the paradoxical effects on precision, where neural fidelity falls near the previous feature while behavioural precision rises at both the previous and the opposing values. Our proposal also distinguishes serial dependence from the superficially similar sensory adaptation, which shows the opposite neural signature (sharpening, rather than dampening), and operates over a different exposure timescale. Together, we term this framework the stable world theory, which recasts serial dependence as an adaptive feature of predictive processing rather than a flaw.</p>