continuouscvae-Apr22-beta-schedule-followup20260422
continuous_cvae-Apr22-beta-schedule-followup_20260422
Status: complete - 108/108 runs completed and locally analyzable.
This collection is the Apr22 random-init continuous-CVAE follow-up to continuous_cvae-Apr1-7loss_20260401. It keeps the same six tasks, the same three adaptation modes, and the same continuous-latent setup, but replaces the earlier seven-loss menu with six pure beta-schedule settings. The current aggregate shows that delayed linear schedules lead the agreed router probe x final answer accuracy ranking, with beta0p1_d50_lin50 first overall and beta0p01_cyclin25_w12p5 second.
- design
- results
- profile
- attention_pattern
- beta0p01-full-schedule-deepdive-results
- full-vs-reinit-dynamics-deepdive-results
analysis.mddeferred until results review
Latest additions: the V3 attention-pattern report now includes clearer base-model surprisal scaling, split task-wise alignment figures, hard-attention summaries, bootstrap-significance heatmap annotations, and per-task token-aligned example panels for the best full-mode loss setting.
Artifacts
- Analysis script:
path:experiments/synthetic_sequences/disentangled/analysis/run_continuous_cvae_apr22_beta_schedule_followup_20260422.py - Aggregate directory:
path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr22-beta-schedule-followup_20260422 - Repo-side report:
path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr22-beta-schedule-followup_20260422/report.md - Findings:
path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr22-beta-schedule-followup_20260422/findings.json - Figures directory:
path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr22-beta-schedule-followup_20260422/figures
Context
- Direct design predecessor: continuous_cvae-Apr1-7loss_20260401 design
- Direct results predecessor: continuous_cvae-Apr1-7loss_20260401 results
- Layout reference: from-entangled-v2 continuous-v2 design
- The next intended step is a focused deep dive into the top two schedules under the agreed
router probe x final answer accuracyranking rule.
Runs
- Slurmkit collection:
path:.slurmkit/collections/exp2_continuous_cvae-Apr22-beta-schedule-followup.yaml - Run count:
108/108complete - Raw run directory:
path:.runs/synthetic_sequences/disentangled/exp2_continuous_cvae-Apr22-beta-schedule-followup - Date range:
2026-04-22