Instead of storing a fixed, pretrained parameter bank the way a standard model or a Mixture-of-Experts architecture does, a hypernetwork learns a general mapping from some input (data, context, or a task description) to the weight values another network should use. A September 2026 Cambridge paper proposed using a compact hypernetwork to turn live, run-time context — facts a user supplies mid-session, corrections they give — into a low-rank modulation of a shared base model's feed-forward weights, so that information could persist and be reused without being re-fed through the prompt on every request. The paper's specific contribution was carrying a Bayesian belief over the hypernetwork's latent code and updating it online as a session proceeds, rather than reading context once and freezing the generated weights.