Learning coefficient estimation on grokked d=8 squared-ReLU checkpoint
157Lambda (λ)SGLD-LLC (locked, gelu cell)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SGLD-LLC (locked, gelu cell)granularity=global model, cost=locked config does not transfer2026.07 | 157 | — | 1.3 | |
| restricted SGLD-LLCgranularity=one MLP block, cost=one estimate2026.07 | 147 | — | 0.99 | |
| SGLD-LLC (calibrated)granularity=global model, cost=5×5 grid + estimate (453 s)2026.07 | 116.7 | — | 0.99 | |
| rate readgranularity=one direction, cost=one 18 s pass, deterministic2026.07 | 1 | 3 | — | |
| restricted SGLD-LLCgranularity=one dead unit, cost=no calibration setting passes2026.07 | 0 | — | — |