Daniel Corva's paper restores curiosity in Gaussian AI agents

In a new paper posted to arXiv, Daniel Corva tackles a fundamental problem in active inference: why do linear-Gaussian state-space models lose all incentive to gather information? Recent work had shown that under standard formulations, the epistemic term of the expected free energy becomes constant, flattening the agent into a Kalman filter with a fixed gain sequence—no action could make it curious again. The only known workaround required multiplicative control in the dynamics, leaving the observation side unexplored.

Corva shows that a minimal state-dependent observation noise—where the sensor's covariance R(x) varies with the state x, such as accuracy degrading with range—is sufficient to break the deadlock. When R is evaluated at the predicted mean, the posterior covariance and Kalman gain become action-dependent. Under mild rank and non-degeneracy conditions, epistemic value is no longer constant. This is a constructive proof that the Bar-Shalom-Tse dual effect can appear in the agent's maintained covariance, meaning actions now shape future estimate quality, not just the state.

The paper is notable for its rigor: an executable witness accompanies the theorem—any fixed filter that matched the agent's beliefs would refute both theorem and witness. The companion Python library cpomdp v0.4.2 automatically detects the linearization incompatibility and raises a typed IncompatibleLinearizationError. This work offers a precise, observation-side characterization of curiosity in Gaussian agents, bridging dual control theory and active inference for the first time.