Synthetic Sequences Experiments

Synthetic sequences are the project's main controlled evaluation setting for latent-strategy disentanglement. The setting combines deterministic tasks with known strategy taxonomies, a stage-1 entangled conditional model that learns a mixture over solution strategies, and a stage-2 disentangling procedure that tests whether the learned mixture can be refactored into an explicit router-generator latent variable model.

Structure

  1. Tasks and data
  2. Entangled baseline
  3. Disentangled factorization
  4. Evaluation and diagnostics
  5. Experiment lines
  6. Documents
Built with LogoFlowershow