from-entangled-v2 continuous-v3 beta-schedule Results

Back to design

Collection State

• planned design: 6 tasks x 6 beta schedules x 1 LoRA mode x 1 seed = 36 jobs
• completed runs: 36/36
• locally analyzable runs: 36/36
• best overall settings by mean router probe x final answer accuracy: beta0p01_d50_lin50, then beta0p1_d50_lin50

Caption: Every task-by-setting cell is present exactly once in the discovered aggregate.

Global Schedule Ranking

Using the agreed collection-level mean over all 36 runs, the six settings sort as follows:

Observed ranking facts:

• beta0p01_d50_lin50 ranks first on the agreed composite with mean router probe 0.8398 and mean final accuracy 0.9985.
• beta0p1_d50_lin50 is close behind and differs only slightly on the composite (0.8368).
• beta0_unnorm has a higher mean router probe than beta1_d50_lin50, but its lower mean final accuracy leaves it behind on the agreed composite.
• beta0p1_cyclin25_w12p5 has near-ceiling mean final accuracy while also having the lowest mean router probe and the lowest mean posterior probe in the collection.

Caption: Mean router probe x final answer accuracy, router probe, final answer accuracy, and posterior-mu probe for each schedule setting.

Delayed-Linear Versus Cyclical Comparison

The matched delayed-linear versus cyclical comparisons are:

Observed pairwise facts:

• At both beta=0.01 and beta=0.1, delayed linear has the higher mean router probe, the higher mean posterior probe, and the higher mean router probe x final answer accuracy.
• The cyclical schedule has slightly higher mean final-answer accuracy in both matched pairs, with mean deltas of about 0.0015.
• The delayed schedules also have lower mean baseline-normalized reconstruction values in both matched pairs.

Caption: Mean delayed-minus-cyclical deltas over tasks for the main final metrics.

Caption: Per-task delayed-minus-cyclical deltas for the two matched beta pairs. Positive cells favor delayed linear on the listed metric.

Task-Level Structure

The highest-router setting is not identical across tasks:

Observed task-level facts:

• beta0p01_d50_lin50 is the highest-router setting on list_summation, sorting_algorithms, and base_conversion.
• beta1_d50_lin50 is the highest-router setting on grid_pathfinding and multidigit_addition.
• beta0_unnorm is the highest-router setting on linear_equation_solving.
• The weakest composite cells are concentrated in the cyclical beta=0.1 setting and the sorting_algorithms / beta0_unnorm cell.

Caption: Final-answer accuracy, router probe, posterior probe, router-times-accuracy, unique-strategy rate, and baseline-normalized reconstruction by task and setting.

Caption: One point per run, showing where high router probe and high final-answer accuracy coincide and where they separate.

Dynamics Snapshot

Observed dynamics facts from the validation curves:

• The delayed-linear schedules stay among the highest validation router-probe trajectories on most tasks.
• The cyclical beta=0.1 setting remains below the delayed schedules on both validation router probe and validation posterior probe on several tasks.

Caption: Validation final-answer accuracy, router probe, and posterior-mu probe over training for each task and schedule setting.

Top And Bottom Runs By The Agreed Composite Score

Top runs:

Bottom runs: