b0 const Deep-Dive Results
b0 const Deep-Dive Results
This note isolates the within-setting analysis for beta0_const_pv0. The refreshed deep dive now uses all six completed b0 const cells: list_summation, sorting_algorithms, grid_pathfinding, linear_equation_solving, base_conversion, and multidigit_addition. The flattened frame grid uses the nearest available checkpoints to 2%, 10%, 20%, 50%, and 100% training progress from the repo-side b0_const_selected_router_frame_progress.csv artifact.
Snapshot
- completed tasks in this deep dive:
6 - mean final test final-answer accuracy:
0.9900 - mean final test router-sampled probe accuracy:
0.8940 - mean final test posterior-mu probe accuracy:
1.0000 - mean final baseline-normalized reconstruction loss:
0.000014 - mean final validation KL:
736285.0
Latent-Space Structure Over Training
The fixed-PCA router-sampled latent projections show task-dependent emergence of strategy structure during training. grid_pathfinding, base_conversion, and linear_equation_solving each end with clearly separated bands or clusters by late training. list_summation broadens from a compact early geometry into a more visibly separated late structure. sorting_algorithms remains the most diffuse row at the end of training. multidigit_addition also contributes latent clouds across the selected checkpoints and ends with visibly separated late-training groups.
Caption: Fixed-PCA projection of router-sampled latent vectors over training for the completed b0 const runs. Columns are tasks, points are colored by inferred strategy label within task, and each frame uses the same task-specific projection basis and axis limits.
Caption: Selected checkpoints from the same fixed-PCA latent trajectory view. Rows are tasks and columns are the nearest available checkpoints to 2%, 10%, 20%, 50%, and 100% training progress.
Probe, Reconstruction, And KL Dynamics
Across the six completed b0 const tasks, final test posterior-mu probe accuracy is 1.0000 on every task and final test final-answer accuracy ranges from 0.9789 to 1.0000. Final test router-sampled probe accuracy ranges from 0.8687 to 0.9263. Relative to the first validation evaluation, router-probe accuracy rises from roughly 0.36-0.48 into the high-0.8 to low-0.9 range while baseline-normalized reconstruction collapses from roughly 13-20 to approximately 1e-5. Final validation KL remains very large, ranging from approximately 2.3e3 to 2.46e6.
Caption: Validation router-probe accuracy, posterior-probe accuracy, baseline-normalized reconstruction loss, and KL over training for the completed b0 const tasks. The completed rows combine rapidly increasing probe accuracy and near-zero reconstruction with large end-of-training KL values.
