LLM-driven structural growth campaign

Evaluate whether a production LLM can reliably operate the SoftModel’s structural-growth surface end to end, including proposing, testing, and governing topology changes through the reality gate.

Background

The factory lifecycle was validated end to end with a production LLM, including teaching, inference, drift recovery, and version management. Structural-growth tools are exposed through the same interface and are script-validated.

The paper does not yet report an LLM-operated growth campaign, leaving the reliability of autonomous natural-language operation of structural changes unresolved.

References

Finally, the LLM-operated validation covers the factory surface; the growth surface is script-validated and exposed identically, but an LLM-driven growth campaign is future work, as is the exploitation of the self-knowledge signals that the program (App.~\ref{sec:empirical}) shows are real.

— 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 15, paragraph “Limitations”

Fourth, the plasticity lanes compare the growing model against the small fixed control only; a large-from-birth fixed control's keep-learning curve was not run in these lanes and is recorded as an open control (the parameter-matched ceiling exists only in the capacity lane).

— 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 11, subsection “Boundaries”