synthetic-disentangled-Mar19 Design

Motivation

This collection is a sparse synthetic sweep designed to ask a focused question: which auxiliary loss terms actually change latent-strategy alignment once answer accuracy is already easy to maintain?

Design

• tasks:
  • sorting_algorithms
  • grid_pathfinding
  • multidigit_addition
• planned cardinality: 63 jobs across 21 curated settings
• major setting families:
  • beta-only baselines
  • token-weighted reconstruction
  • posterior-entropy-only
  • router-marginal-KL-only
  • router-support-only
  • occupancy-oriented composites

Research Questions

1. Does token-weighted reconstruction materially improve strategy alignment?
2. Do occupancy-oriented auxiliaries improve latent semantics on their own?
3. Is final task accuracy already saturated enough that alignment becomes the more useful selection metric?