b0p1-norm-pv0-twr0-deepdive-results
****— type: report status: draft aliases: [] created: 2026-04-21
b0.1 norm Deep-Dive Results
This note isolates the within-setting analysis for beta0p1_norm_pv0_twr0, selected under the Apr1 deep-dive rule: keep settings with mean final-answer accuracy at least 0.95, then rank by mean final router-sampled probe accuracy. The current note covers all 18 runs (6 tasks x 3 adaptation modes).
Snapshot
- included runs:
18 - mean final test final-answer accuracy:
0.9885 - mean final test router-sampled probe accuracy:
0.7604 - mean final test posterior-mu probe accuracy:
0.9439 - mean final baseline-normalized reconstruction loss:
0.1544 - mean final validation KL:
1.491 - mean first-eval validation router probe accuracy:
0.5650
Latent-Space Structure Over Training
The fixed-PCA router-sampled latent trajectories are organized by adaptation mode and task rather than pooled across the collection. The animated grid shows how separation develops over the shared training timeline, while the three static frame sheets hold the same view at approximately 2%, 10%, 20%, 50%, and 100% training progress for full, lora, and reinit.
Caption: Fixed-PCA projection of router-sampled latent vectors over training for beta0p1_norm_pv0_twr0. Rows are adaptation modes, columns are tasks, and each panel keeps a run-specific projection basis and axis limits across the full trajectory.
Caption: Selected latent-geometry checkpoints for the full rows.
Caption: Selected latent-geometry checkpoints for the lora rows.
Caption: Selected latent-geometry checkpoints for the reinit rows.
Probe, Reconstruction, And KL Dynamics
The task-wise dynamics plot overlays the three adaptation modes for each task. This makes it easier to see whether the later latent separation comes from router-side probe gains alone or from a broader shift in posterior probe quality, reconstruction, and KL.
Caption: Validation router-probe accuracy, posterior-probe accuracy, baseline-normalized reconstruction loss, and KL over training for beta0p1_norm_pv0_twr0, with one line per adaptation mode inside each task row.
Analogical Agreement
Analogical-steering summaries are available for this setting across all task and adaptation-mode cells, so the note includes a task-by-mode analogical-agreement summary alongside the latent and probe views.
Caption: Analogical unique-agreement rate by task and adaptation mode for this loss setting.
