Apr27 Top Loss Dynamics

Note

This is an observational deep-dive for the Apr27 paper basic-grid loss setting with the highest mean matched strategy router score across the six complete single-task LoRA runs. The selected setting is tokcc_a0p25_b0p01_lin100, which uses normalized concentration-corrected token reconstruction with alpha 0.25, beta 0.01, and a linear beta warmup over the full epoch.

Selection

The selected loss setting has mean analysis/router_probe_times_strategy_match_rate = 0.8811 across its 6 complete task runs. This is the top row in the aggregate loss-setting summary. This page remains single-task-only; the recovered multi-task dynamics and best-setting latent views are in multitask.

Task	Matched strategy router score	Final answer accuracy	Router probe	Strategy match rate	Posterior-mu probe	Baseline-normalized reconstruction
`base_conversion`	0.9153	0.9952	0.9214	0.9934	1.0000	8.8885
`grid_pathfinding`	0.9131	0.9997	0.9201	0.9924	1.0000	10.5553
`linear_equation_solving`	0.8675	0.9966	0.8779	0.9882	1.0000	9.2168
`list_summation`	0.8834	0.9526	0.9308	0.9491	0.9972	7.8331
`multidigit_addition`	0.8437	0.9981	0.8454	0.9980	1.0000	6.4418
`sorting_algorithms`	0.8635	0.9981	0.9068	0.9523	0.9875	3.9407

Validation Dynamics

The validation dynamics panel plots validation final-answer accuracy, router probe, strategy-match rate, matched router score, posterior-mu probe, baseline-normalized reconstruction, and KL prior by epoch progress for each complete single-task run.

Router-Sampled Z-Space

The router-sampled z-space frames plot PCA projections of sampled prior z values, colored by inferred generation strategy. Rows are the complete single-task LoRA runs and columns are selected training checkpoints.

Router-sampled z-space animated training dynamics for the top Apr27 loss setting

Posterior-Mu Z-Space

The posterior-mu z-space frames plot PCA projections of posterior mean vectors, colored by ground-truth strategy. Posterior probe accuracy is near ceiling for most cells, so these plots primarily show the timing and geometry of posterior clustering.

Posterior-mu z-space animated training dynamics for the top Apr27 loss setting

Smoothed posterior-mu z-space animated training dynamics for the top Apr27 loss setting

Run-Level Notes

The highest-scoring individual cells are base_conversion at 0.9153, grid_pathfinding at 0.9131, and list_summation at 0.8834. The lowest-scoring cells are multidigit_addition at 0.8437, sorting_algorithms at 0.8635, and linear_equation_solving at 0.8675.

The data backing this page are in path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/data/tokcc_a0p25_b0p01_lin100_final_metrics_by_task.csv and path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/data/tokcc_a0p25_b0p01_lin100_task_checkpoints.csv. The generation script is path:experiments/synthetic_sequences/disentangled/analysis/run_continuous_cvae_apr27_top_loss_dynamics_20260429.py.