Nico Scherf

Active Inference can be framed as a convex MDP, revealing a surprising connection to performative reinforcement learning with robust policy improvement guarantees.

Transformers trained on a simple grid-world learn hidden representations that directly reflect the underlying predictive geometry, offering a glimpse into how neural networks internalize structural constraints.

Even though transformers and echo-state networks can accurately forecast dynamical systems, their internal representations are surprisingly misaligned compared to MLPs and RNNs.