CVAE Post-Training Methodology for Latent Strategy Disentanglement

This is the landing page for the core CVAE methodology thread. It keeps the methodology readable as one coherent cluster while surfacing the most reusable pieces as linked concept notes.

Core Structure

• Problem setup
• Model components
• Parameterization
• Objective and loss terms

Related Loss Notes

• Distribution-Level Z-Free-Relative Reconstruction Rescaling
• Discrete latent VAEs
• Token-level surprisal-share weighting
• Inter-latent divergence
• InfoVAE connections
• Sequence-level divergence

Current Evaluation Setting

• Synthetic sequences experiments

Summary

The central idea is to factor a frozen entangled conditional generator into:

• a strategy router $p_\phi(z \mid x)$,
• a strategy-conditioned generator $p_\psi(y \mid x, z)$,
• and a training-time inference model $q_\xi(z \mid x, y)$.

The methodology covers both discrete and continuous latent variants, multiple adaptation regimes, and a family of ELBO-adjacent interventions aimed at making the latent variable carry strategy information rather than collapsing.