Variation of geometric decay and consolidation break-even

Determine how the geometric plasticity-decay constant β and the consolidation break-even point vary with model architecture, task characteristics, and optimizer choice in sequential bounded in-cell updates.

Background

CellFill’s sequential-update experiment uses four disjoint tasks on one Qwen3-1.7B model, with one rank, one optimizer configuration, one seed, and one consolidation variant. The measured writable-room decay is approximately geometric, with β = 0.825 ± 0.002, while consolidation restores writable room at the cost of changing the artifact.

Because the sequential evidence comes from a single configuration, the paper does not establish that the measured decay constant or the point at which consolidation becomes beneficial transfers across models, tasks, or optimization procedures. The authors explicitly leave those dependencies unresolved.