Comparison with broader immobilization methods

Establish whether governed plasticity is superior to the broader immobilization family, including progressive columns and frozen-backbone adapter methods, in direct comparisons on capacity-pressing continual-learning tasks.

Background

The paper evaluates governed plasticity against an Elastic Weight Consolidation baseline and reports a favorable result on three of four tested tasks. It explicitly declines to claim superiority over progressive columns and adapter-freezing approaches, leaving that broader comparison unresolved.

This problem is included because the paper’s central contrast is between governance-based plasticity and immobilization-based continual-learning methods, but the empirical evidence does not yet cover the full immobilization family.

References

We do not claim superiority over the broader immobilization family (progressive columns, adapter freezing) --- that comparison remains open --- nor that composition is more effective than widening in general: the measured picture is regime-dependent --- the first-generation bar failed (App.~\ref{app:twodir}), and the 2026 depth mapping locates the compositional advantage precisely where iterated composition is real, at fixed parameter budget, and nowhere else (\S\ref{sec:campaigns-results}).

— SoftModel: A Neural Model That Grows Its Own Topology -- Governed Structural Growth for Continual In-Service Learning  (2608.16409 - Xie, 17 Aug 2026) in Section 1, paragraph “What we claim, and what we do not”; Section 15, paragraph “Limitations”