continuous_cvae-Apr1-7loss_20260401 Results

Collection State

• planned design: 6 tasks x 7 loss settings x 3 adaptation modes x 1 seed = 126 jobs
• completed runs: 126/126****
• analogical-summary runs: 88/126

Figure-Driven Summary

Caption: The refreshed scripted analysis now covers the full 126/126 Apr1 grid and makes adaptation mode an explicit paneling axis.

Caption: Final accuracy stays broadly high, while router probe, posterior probe, and normalized reconstruction vary much more strongly across the seven-loss menu.

Caption: The improved scatter design makes the main Apr1 frontier easier to read: loss setting is encoded by color, task by marker, and the legend now sits outside the plotting area.

Caption: Posterior-side probe quality is often at or above router-side probe quality, but the gap depends on both setting and adaptation mode.

Caption: Answer accuracy remains high over much of the grid, which reinforces that probe quality rather than task accuracy is the sharper diagnostic axis in Apr1.

Caption: Aggregating over tasks makes the broad loss-setting ranking easier to read, while the overlaid task markers show where individual tasks depart from the mean.

Caption: The exact Apr1 twr1 contrasts are the two normalized matched pairs, and the live collection reflects the corrected rerun rather than the archived bad-TWR snapshot.

Caption: This high-level dynamics summary complements the per-mode task grids by showing the mean training trajectories for accuracy, router probe, posterior probe, and normalized reconstruction within each adaptation mode.

Top-2 Deep Dives

• Selection rule: keep settings with mean sampled_test_final/final_answer_accuracy >= 0.95, then rank them by mean sampled_test_final/router_sampled_z/linear_probe_accuracy.
• Selected pair: beta0p1_norm_pv0_twr1, beta0p1_norm_pv0_twr0
• Deep-dive notes: b0p1-norm-pv0-twr1-deepdive-results, b0p1-norm-pv0-twr0-deepdive-results
• The live Apr1 collection is the corrected rerun; the retired bad-TWR snapshot is archived under path:.runs/synthetic_sequences/disentangled/exp2_continuous_cvae-Apr1-7loss/_archived_bad_twr1_20260405.

Caption: Analogical agreement varies strongly across the grid and remains useful as a secondary latent-structure diagnostic on the settings where analogical summaries were generated.

Caption: Router probe and analogical agreement remain positively associated across the available Apr1 analogical subset.

Additional Figures

• per-adaptation alignment dynamics:
  • figures/dynamics_alignment_metrics_full.png
  • figures/dynamics_alignment_metrics_lora.png
  • figures/dynamics_alignment_metrics_reinit.png
• per-adaptation objective dynamics:
  • figures/dynamics_objective_terms_full.png
  • figures/dynamics_objective_terms_lora.png
  • figures/dynamics_objective_terms_reinit.png

Safe Reading

• The strongest Apr1 conclusion remains about how the loss menu changes latent-quality diagnostics, but the live collection should now be read as the corrected rerun rather than through the earlier bad-TWR caveat.
• The refreshed figure suite sharpens that reading by separating adaptation modes cleanly and by pairing the detailed per-task grids with one aggregate dynamics summary over tasks.
• The new deep-dive notes isolate the two best settings under the agreed accuracy-floor-plus-router-probe rule and make their latent trajectories directly inspectable.
• Analogical agreement continues to track router-probe quality on the subset where analogical summaries are available, which supports its use as a semantic-stability diagnostic rather than as a replacement for the main disentanglement metrics.