Disentangled Synthetic Sequences Reports

Collections

• synthetic-disentangled-Mar19 Sparse Mar 19 loss sweep over three synthetic tasks, with token-weighted reconstruction emerging as the strongest positive intervention on selected tasks.

• continuous-cvae-Mar20 Mar 20 continuous-CVAE collection report emphasizing the posterior-vs-sampled gap and strong full-adaptation performance.

• continuous_cvae-Apr1-7loss_20260401 Apr 1 seven-loss continuous-CVAE sweep note, including the caveat that the token-weighted reconstruction implementation was incorrect in this collection.

• continuous_cvae-Apr22-beta-schedule-followup_20260422 Apr 22 beta-schedule follow-up to the Apr1 continuous-CVAE sweep, showing delayed linear schedules as the strongest overall candidates once router probe is paired with final-answer accuracy.

• continuous_cvae-Apr24-reconstruction-stack_20260424 Apr 24 reconstruction-stack follow-up that keeps the Apr22 grid and compares raw, distribution-normalized, example-surprisal-normalized, and normalized token-weighted reconstruction variants; the completed 234/234 aggregate has curated figures, observational results, deep-dive result notes, and an agent-drafted interpretation pending human review.

• continuous_cvae-Apr26-reconstruction-stack-followup_20260426 Apr 26 targeted reconstruction-stack follow-up that keeps the Apr24 grid and closes the main unresolved objective gaps: cc-TWR-only alpha probes, example-inverse-surprisal plus cc-TWR, low-beta raw controls, and a beta-0.1 continuation of distribution-normalized plus cc-TWR.

• continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426 Unified Apr24+Apr26 reconstruction-stack report comparing all 22 loss settings by matched strategy router score, with raw baselines separated and selected Apr26-minus-Apr24 matched contrasts.

• paper-selected-reconstruction-stack-losses_20260426 Paper-facing selected loss-setting subset with an Apr27 partial 180/210 paper basic-grid snapshot, curated C3/C5 figures, and an agent-drafted interpretation pending human review.

• exp2_discrete_cvae-Apr3-7loss Apr 3 and Apr 7 discrete-CVAE executive report on normalization, token weighting, and JSD interventions.

• ss-disentangled-direct-pretrained-v1 Apr 13 direct-pretrained diagnostic note showing saturated task accuracy but weak latent-strategy alignment under a reconstruction-plus-KL objective.

• ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1 Apr 14 completed follow-up sweep showing that explicitly controlling PV and RMKL diagnostics does not yet produce a reliable router-side alignment gain.

• synthetic-pretrained-disentangled-from-entangled-v1 Apr 14 entangled-initialized counterpart to the direct-pretrained collection, showing strong posterior-side gains on some tasks but only partial router-side improvement.

• synthetic-pretrained-disentangled-from-entangled-v2 Apr 15 follow-up loss sweep whose local artifacts suffered naming collisions; the surviving analysis is still useful directionally but the collection is superseded.

• synthetic-pretrained-disentangled-from-entangled-v3 Apr 16 narrowed successor sweep testing lower beta, longer warmup, and a matched mild-PV comparison after the v2 artifact issue.

• synthetic-pretrained-disentangled-from-entangled-v4 Apr 17 low-KL follow-up that adds a zero-beta anchor, a smaller constant-beta setting, and a matched mild-PV comparison while the last few cells are still in flight.

• synthetic-pretrained-disentangled-from-entangled-v5-twr Apr 18 token-weighted reconstruction follow-up that reuses the strongest nonzero-beta regimes from v4 and compares twr=0.25 versus 0.5 in an early partial snapshot.

• from-entangled-v2/continuous-v1 Apr 20 partial snapshot of the new entangled-v2-source continuous rerun, analyzed with the earlier from-entangled figure layouts but without direct numerical comparison to the older collections.

• from-entangled-v2/continuous-v2 Apr 21 matched follow-up that increases continuous_latent_dim from 8 to 64 while keeping the same entangled-v2-source continuous loss menu, currently in a partial 33/36 snapshot.

• from-entangled-v2/continuous-v3-beta-schedule Apr 23 schedule-only follow-up importing the Apr22 six-setting beta-schedule menu into the entangled-v2-source LoRA rerun, with delayed-linear schedules leading the new single-mode sweep on the agreed composite ranking.

• continuous_from_entangled_v2_v4_reconstruction_stack_qwen25_05b_20260426 Apr 26 qwen-0.5B from-entangled reconstruction-stack report that imports the Apr24 loss menu into the pretrained LoRA-only setting; all 78/78 cells are complete and the first-pass observational results note is available.

• from-entangled-v2/paper-method-grid-multi-task-z64-qwen25_20260501 May 1 compact multi-task paper-method grid plus TOK-CC/dist-weight follow-ups: 21/25 cells have final metrics, the Qwen2.5 1.5B paper grid is terminal partial with four failed tokcc-family cells, and the z8/z64 Qwen2.5 0.5B follow-up comparison is integrated into the report.

• from-entangled-v2/additional-model-families_20260724 Extension of the pretrained multi-task experiments beyond Qwen. SmolLM2-360M provides the first usable replication; Pythia-410M is retained as a source-learning diagnostic, with OLMo 2 1B and Gemma 3 270M approved next.

• from-entangled-v2/synthetic-pretrained-disentangled-from-entangled-v2-reconstruction-token-aggregation-sum-ablation-multi-task-z64-qwen25-05b_20260503 May 3 focused sum-token aggregation ablation for Qwen2.5 0.5B multi-task z64 from-entangled-v2 runs, comparing common cells against the prior sequence-mean paper-grid reference.

• from-entangled-v2/tokcc-hparam-followup-multi-task-z64-qwen25_20260503 May 3 targeted 26-cell TOK-CC alpha/beta follow-up for Qwen2.5 0.5B and 1.5B, designed to separate seed noise, beta sensitivity, alpha sensitivity, and the stabilizing effect of a small distribution-normalized reconstruction component.

Notes

• Collection-private report figures live inside each collection directory.