ss-disentangled-direct-pretrained-pv-rmkl-sweep-v1 Results

Collection State

planned design: 6 tasks x 2 pv settings x 2 rmkl settings x 1 seed = 24 jobs
completed analyzable cells: 21
failed cells: 3
all failures are sorting_algorithms CUDA OOMs

Completed-Cell Snapshot

mean sampled final-answer accuracy: 0.9997
mean router probe accuracy: 0.4335
mean posterior probe accuracy: 0.4802

Setting-level means across completed cells:

Setting	Final accuracy	Router probe	Posterior probe	Norm recon	KL prior	Posterior variance	Router marginal KL	Total loss
`pv0.0_rmkl0.0`	`1.0000`	`0.4344`	`0.4847`	`0.0119`	`0.0030`	`756.88`	`4277.57`	`0.0122`
`pv0.0_rmkl0.1`	`1.0000`	`0.4343`	`0.4752`	`0.0121`	`0.0025`	`1.00`	`0.00`	`0.0124`
`pv0.1_rmkl0.0`	`0.9990`	`0.4294`	`0.4885`	`0.0119`	`0.0747`	`74.58`	`360.55`	`7.4773`
`pv0.1_rmkl0.1`	`1.0000`	`0.4356`	`0.4734`	`0.0119`	`0.0696`	`0.69`	`0.07`	`0.0941`

Figures

Validation accuracy and alignment dynamics by task and setting. Caption: Accuracy stays high across the completed cells, while router and posterior probe behavior remains task-specific and only modestly sensitive to the PV/RMKL settings.

Objective terms by task and setting. Caption: Baseline-normalized reconstruction and KL-to-prior remain small across settings. Total loss becomes large only when the weighted PV term is active and its target quantity remains large.

Weighted auxiliary PV and RMKL terms by task and setting. Caption: The regularized settings strongly control the targeted auxiliary quantities, often by orders of magnitude relative to the unregularized baseline.

Standardized objective versus auxiliary dynamics by task and setting. Caption: The optimizer clearly responds to the added regularizers, but the representation changes do not translate into a clean global router-probe gain.

Completed-cell final metrics by task and setting. Caption: Final task accuracy remains near-saturated across completed cells, while alignment metrics are modest and task-dependent.