continuous-cvae-Mar20 Design

Motivation

This collection broadens the continuous-CVAE study from earlier sparse sweeps into a multi-task, multi-setting grid. The goal is not just to maintain task accuracy, but to understand which settings make sampled and posterior latents encode strategy information cleanly across tasks and adaptation modes.

Design

• planned grid: 6 tasks x 8 curated loss settings x 2 adaptation modes x 1 seed = 96 runs
• adaptation modes: full and reinit
• primary metrics: sampled final-answer accuracy, sampled-z probe accuracy, posterior-mu probe accuracy, unique-valid strategy rate, invalid strategy rate
• secondary diagnostics: training dynamics for probe quality and answer accuracy

Research Questions

1. How strongly does latent-strategy decodability vary across tasks and loss settings when answer quality is already high?
2. Does full adaptation consistently outperform reinit on probe quality and answer utility?
3. How large is the posterior-versus-sampled gap, and does it appear as a stable pattern or only in isolated runs?