arXiv:2607.14937 2026 Architecture 2 ideas extracted · analyzed Aug 30, 2026
The paper's transferable asset is not the large DynaMix model, but its reduction to a two-parameter, retrieval-based dynamical update: forecast by combining the current latent state with a nearby observed state and that state's temporal successor. This creates an extremely small zero-shot forecaster whose inductive bias is that local recurrence in state space carries short-term dynamics, while the convex or affine blend can preserve bounded trajectories better than unconstrained autoregressive extrapolation. The most direct neural-network use is a nonparametric residual head or fallback forecaster attached to a learned encoder, with coefficients calibrated globally or in closed form from demonstrations.
Calibrate the two blend coefficients directly from a context trajectory rather than using gradient descent. The one-step prediction problem is a two-variable ridge regression, making per-task adaptation nearly free and suitable for zero-shot or few-shot system identification.
Replace a parameter-heavy recurrent transition, or use this as a fallback, with a two-parameter nearest-neighbor successor blend in latent space. Given a query latent state, retrieve the closest state from an in-context trajectory and combine the query, the retrieved state, and its observed successor; this gives a zero-shot dynamical forecast with almost no trainable transition parameters.