synthetic-pretrained-disentangled-from-entangled-v2 Design

Summary

synthetic-pretrained-disentangled-from-entangled-v2 is a targeted follow-up to synthetic-pretrained-disentangled-from-entangled-v1, the entangled-initialized Qwen2.5-0.5B disentangling collection. It keeps the same 6 synthetic tasks, lora adaptation, continuous latent setup, and overall training budget as v1, but replaces the single loss configuration with 5 curated loss settings.

Planned design:

• 6 tasks x 5 loss settings x 1 seed = 30 jobs
• initialization from synthetic-pretrained-entangled-v1
• backbone Qwen2.5-0.5B
• continuous latent, dim 8, last-token pooling

Motivation

The main motivation is to test whether stronger effective reconstruction pressure helps the latent variable encode reasoning strategy more cleanly in the entangled-initialized regime.

This follows from two observations in synthetic-pretrained-disentangled-from-entangled-v1:

• posterior-side strategy structure improved on some tasks relative to the direct-pretrained setup, especially base_conversion
• lower baseline-normalized reconstruction loss appeared to coincide with stronger router probe accuracy in the small final-state comparison

That pattern is not strong enough to treat as a conclusion, but it is enough to motivate a focused loss follow-up. The goal of v2 is therefore to test whether reducing or delaying KL pressure improves strategy alignment more effectively than the original constant-beta=0.1 setting.

Loss Settings

The collection uses five curated settings:

1. beta0p1_const_pv0 Baseline reference: beta=0.1, constant schedule, no posterior-variance penalty.
2. beta0p05_const_pv0 Moderate KL reduction with the same constant schedule.
3. beta0p02_const_pv0 Stronger shift toward a reconstruction-dominated regime.
4. beta0p1_lin30_pv0 Same final beta as baseline, but linearly warmed up over 30% of training.
5. beta0p05_lin30_pv0p01 Lower and delayed KL pressure plus a mild posterior-variance penalty 0.01.

All settings keep:

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

Design Intent

This is not a broad factorial sweep. It is a compact diagnostic collection meant to answer three concrete questions:

1. Does lowering beta improve router-side strategy alignment?
2. Does delaying KL pressure via warmup help more than simply lowering beta?
3. Does a mild posterior-variance penalty help once KL pressure is reduced or delayed?

The resulting collection should provide a clearer next decision about whether the main remaining bottleneck is early KL pressure, sampled-latent noise, or something deeper about how the router uses z.