synthetic-pretrained-disentangled-from-entangled-v5-twr Design

Summary

synthetic-pretrained-disentangled-from-entangled-v5-twr is a targeted follow-up to synthetic-pretrained-disentangled-from-entangled-v4. It keeps the same entangled-initialized Qwen2.5-0.5B setup, the same 6 synthetic tasks, and the same continuous-latent lora recipe, but replaces the previous posterior-variance follow-up with a token-weighted reconstruction follow-up.

Planned design:

• 6 tasks x 4 loss settings x 1 seed = 24 jobs
• initialization from synthetic-pretrained-entangled-v1
• backbone Qwen2.5-0.5B
• continuous latent, dim 8, last-token pooling

Motivation

The v2 through v4 from-entangled sweeps increasingly suggested that stronger effective reconstruction pressure matters more than posterior-variance control or router-marginal KL control. In particular:

• lower effective KL pressure often improved router-side strategy alignment more than the PV and RMKL interventions did
• the matched mild-PV comparisons remained mixed or negative overall
• v4 pushed further into the low-KL regime, but it still did so with uniform reconstruction weighting

That motivates a different follow-up: keep the best current nonzero-beta regimes, but reweight the reconstruction objective toward tokens that the frozen z-free baseline finds hard. The goal is to test whether more targeted reconstruction pressure can improve latent strategy structure without needing to drop beta even further.

Loss Settings

The collection uses four settings:

1. beta0p005_const_pv0_twr0p25 Very low constant beta with mild token-weighted reconstruction: beta=0.005, constant schedule, posterior_variance_weight=0.0, token_weighted_reconstruction_weight=0.25.

2. beta0p005_const_pv0_twr0p5 Same very low constant-beta regime with a stronger token-weighted term: token_weighted_reconstruction_weight=0.5.

3. beta0p01_lin100_pv0_twr0p25 Full-training warmup to beta=0.01 with mild token-weighted reconstruction: token_weighted_reconstruction_weight=0.25.

4. beta0p01_lin100_pv0_twr0p5 Same full-training warmup regime with the stronger token-weighted term: token_weighted_reconstruction_weight=0.5.

All settings keep:

• normalize_reconstruction_loss = true
• posterior_variance_weight = 0.0
• router_marginal_kl_to_prior_weight = 0.0
• router_support_weight = 0.0
• inter_latent_divergence_weight = 0.0

Design Intent

This collection is meant to answer four concrete questions:

1. In the strongest current constant low-beta regime, does adding token-weighted reconstruction improve router-side strategy alignment relative to the v4 no-TWR baseline?
2. In the strongest current warmup regime, does adding token-weighted reconstruction improve router-side strategy alignment relative to the v4 no-TWR baseline?
3. Is token_weighted_reconstruction_weight=0.5 more effective than 0.25, or does the stronger weighting mainly distort optimization?
4. Does the token-weighted intervention behave differently in the constant-low-beta and full-warmup-to-0.01 regimes?

Current Known Risks

The collection is still early, so the design should be read with two current practical caveats in mind:

• the collection is only partially complete, so the first analysis pass needs to handle a large number of missing cells cleanly
• sorting_algorithms has been memory-fragile in several earlier from-entangled collections, so it remains a likely operational risk for later cells even though no failure pattern is established yet in this early snapshot