Raw-Baseline Latent Deep Dive

Back to results

This deep dive records the behavior of the standard raw ELBO baselines in the completed Apr24 grid.

The three raw baselines have the highest mean final answer accuracy in the loss-setting summary: raw_b1_lin100 has 0.9916, raw_b1_d50_lin50 has 0.9914, and raw_b1_const has 0.9913. Their matched strategy router scores are much lower: 0.4432, 0.4396, and 0.4419.

The low matched scores are driven primarily by router probe values near 0.45 rather than strategy match rate. Mean strategy match rates for raw baselines are 0.9882, 0.9879, and 0.9884, while mean router probe values are 0.4485, 0.4448, and 0.4470.

Posterior probe is also lower for raw baselines than for normalized settings. Raw posterior probe means range from 0.5817 to 0.7294, while the normalized settings in the top group are at or near 0.998.

The bottom six individual runs by matched score are all raw baselines. They include all three adaptation modes and all three raw schedules.

Open empirical checks:

• Inspect raw-run latent snapshots family by family to determine whether the router probe is low because latent clusters are absent, label-swapped, or not aligned with the intended strategy probes.
• Compare posterior probe trajectories with KL trajectories for raw baselines to determine whether lower posterior decodability appears early or after beta pressure increases.
• Re-run a raw baseline at lower beta to separate the effect of raw reconstruction scaling from the effect of beta magnitude.