Expressive Policy

This is an idea that I stumbled upon while reading a paper written by Perry Dong.

EXPO: https://arxiv.org/pdf/2507.07986

Policy Type	Expressiveness
Linear policy	Low
Shallow MLP (1 hidden layer)	Medium
Deep MLP (many hidden layers, nonlinearities)	High
Transformer policy	Very High
Diffusion-based policy	Extremely High

🧨 This limits it to:

• Single-mode behaviors (deterministic or unimodal)