New paper reveals 'cognitive relapse' when AI models adopt synthetic realities

A new paper from MD Ibrahim Hossain Ridoy (arXiv, July 2026) investigates a fundamental question under the free energy principle: Can a predictive system permanently adopt a synthetic environment as its default hypothesis, displacing the one that first shaped it? The author studies this computational problem—called 'ontological inversion'—through a controlled proxy: a convolutional variational autoencoder (VAE) paired with a recurrent latent predictor. The network's evidence lower bound (ELBO) is mathematically identical to variational free energy. It is first trained on a baseline visual domain, then on a mixed stream where a swept rehearsal ratio r controls how much baseline content persists during transition to a target domain. Representational capacity (what the latent space can discriminate) is tracked separately from default behavior (what the system generates when unconstrained).

Across 90 runs, the two sharply diverge: representational accuracy stays at 0.97–0.998 regardless of r, while default behavior spans nearly the system's entire range depending on r alone. Strikingly, at intermediate r, the system's default output rises toward the target domain, then partially reverts toward the baseline even while training continues unchanged—a structural failure the author terms 'cognitive relapse.' The paper establishes a computational existence proof that resistance to reality adoption is not reducible to learning speed but is a structural property with distinct failure modes. This has direct implications for AI alignment and model editing: even when a model appears to have 'learned' a new environment, its latent beliefs may unexpectedly revert.