Latent Dynamics / Strategy-Aligned Factorization

Current Conference-Paper State

Source: path:paper/conference-paper/figures/disentanglement.tex, input from path:paper/conference-paper/sections/04-experiments.tex.

Label: fig:disentanglement.

Asset: path:paper/figure-production/latent_dynamics/random_init_multitask/dist0p25_tokcc0p75_a0p25_b0p01_lin100/static/output/posterior_multitask_static.pdf.

Caption headline: Our methodology recovers a strategy-aligned latent with stable cross-input semantics.

The current figure shows PCA projections of latent samples for the Summation task across training checkpoints. Router-sampled latents and posterior-side latent means move from overlapping strategy labels to visibly separated regions as training progresses.

Writing Role

This figure supports the core positive result that base-directed reconstruction learns a strategy-aligned router-generator factorization, rather than merely preserving sampled behavior.

Stale Or Open Items

The caption still contains a draft note that the production setting is provisional. The note should be removed or resolved once the task and objective setting are finalized.