Apr27+Apr30 Tokinv Dynamics

Back to results · Interpretation draft · Previous top-loss dynamics

Note

This is an observational deep-dive for the Apr30 extension loss setting with the highest mean matched strategy router score across the six complete single-task LoRA runs. The selected setting is tokinv_a0p25_b0p01_lin100, which uses normalized inverse-surprisal token-weighted 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 = 0.9176 across its 6 complete single-task runs. This is the top fixed-setting row in the current results snapshot. This page remains single-task-only; analogical steering summaries for this setting are included in results#Analogical-Consistency.

Task	Matched strategy router score	Final answer accuracy	Router probe	Strategy match rate	Posterior-mu probe	Baseline-normalized reconstruction
`base_conversion`	0.9026	0.9980	0.9046	0.9978	1.0000	9.5366
`grid_pathfinding`	0.9204	1.0000	0.9204	1.0000	1.0000	15.2416
`linear_equation_solving`	0.9458	0.9988	0.9472	0.9985	1.0000	8.2572
`list_summation`	0.8948	0.9955	0.8991	0.9952	0.9990	6.0186
`multidigit_addition`	0.9010	0.9999	0.9013	0.9997	1.0000	7.3520
`sorting_algorithms`	0.9411	0.9999	0.9461	0.9947	0.9912	5.9916

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 Apr30 tokinv 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 every cell except a still-high sorting_algorithms cell, so these plots primarily show the timing and geometry of posterior clustering.

Posterior-mu z-space animated training dynamics for the Apr30 tokinv loss setting

Smoothed posterior-mu z-space animated training dynamics for the Apr30 tokinv loss setting

Run-Level Notes

The highest-scoring individual cells are linear_equation_solving at 0.9458, sorting_algorithms at 0.9411, and grid_pathfinding at 0.9204. The lowest-scoring cells are list_summation at 0.8948, multidigit_addition at 0.9010, and base_conversion at 0.9026.

Relative to the earlier tokcc_a0p25_b0p01_lin100 dynamics page, tokinv_a0p25_b0p01_lin100 raises the matched strategy router score on every listed task except base_conversion, where it remains slightly lower. The largest gains are on linear_equation_solving, sorting_algorithms, and multidigit_addition; these gains come with near-ceiling final-answer accuracy and near-ceiling strategy-match rate, so the main observable movement is router-side strategy decodability rather than task accuracy.

The data backing this page are in path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/data/tokinv_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/tokinv_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.