b0.005 const + twr0.5 Deep-Dive Results
b0.005 const + twr0.5 Deep-Dive Results
This note isolates the within-setting analysis for beta0p005_const_pv0_twr0p5, which remains the second-best completed setting by mean final router-sampled probe accuracy in the current partial snapshot. The current deep dive uses the five completed cells for list_summation, grid_pathfinding, linear_equation_solving, base_conversion, and multidigit_addition. sorting_algorithms / beta0p005_const_pv0_twr0p5 is still running and is omitted.
Snapshot
- completed tasks in this deep dive:
5 - mean final test final-answer accuracy:
1.0000 - mean final test router-sampled probe accuracy:
0.6300 - mean final test posterior-mu probe accuracy:
0.8828 - mean final baseline-normalized reconstruction loss:
1.043901 - mean final validation KL:
2.3
Latent-Space Structure Over Training
This higher-dimensional TWR deep dive is a direct counterpart to the continuous-v1 b0.005 const + twr0.5 note. In the current partial snapshot it no longer looks like a clean improvement over the matched dim=8 predecessor: the completed rows keep perfect or near-perfect final-answer accuracy, but the mean router probe is lower and the matched latent-dimension comparison shows a negative average router-probe delta for this setting.
Caption: Router-sampled latent geometry over training for the completed b0.005 const + twr0.5 runs.
Caption: Selected checkpoints from the same router-sampled latent trajectory view.
Probe, Reconstruction, And KL Dynamics
Across the five completed b0.005 const + twr0.5 tasks, router probe ranges from task-level regressions near the low-0.6 band up to stronger rows above 0.7, while posterior probe stays materially higher than router probe on most completed tasks. Unlike b0 const, the final validation KL values stay very small. The main observational contrast with the continuous-v1 deep dive is that the higher-dimensional version keeps the low-KL character but does not keep the same router-side strength or reconstruction frontier.
Caption: Validation router-probe accuracy, posterior-probe accuracy, baseline-normalized reconstruction loss, and KL over training for the completed b0.005 const + twr0.5 tasks.
