continuous_cvae-Apr1-7loss_20260401 Design

Motivation

This collection expands the continuous-CVAE search space to a seven-loss menu spanning normalization, posterior-variance regularization, and token-weighted reconstruction. It also introduces analogical-agreement diagnostics to test whether strategy information transfers across inputs under shared latent samples.

Design

• planned grid: 6 tasks x 7 loss settings x 3 adaptation modes x 1 seed = 126 jobs
• adaptation modes: three tracked modes in the Apr1 report figures
• key metrics:
  • final-answer accuracy
  • router probe accuracy
  • posterior probe accuracy
  • analogical agreement rate
• current live collection status: corrected rerun after archiving the earlier bad-TWR snapshot under path:.runs/synthetic_sequences/disentangled/exp2_continuous_cvae-Apr1-7loss/_archived_bad_twr1_20260405

Research Questions

1. Does reconstruction normalization materially improve router-side probe quality?
2. How do probe quality and analogical agreement move together across tasks and adaptation modes?
3. Which loss settings provide the strongest disentanglement once we require good-enough final accuracy?