Lamarckian Weight Inheritance Enables Evolutionary Continual Learning

Submitted to Neurocomputing (Elsevier) — preprint version. Currently under peer review. This paper introduces EvoTransformer with Controlled Online Evolution (COE), a framework combining Lamarckian weight inheritance with multi-objective fitness-gated evolutionary search for continual NLP. The central finding is empirical: across three benchmarks (20 Newsgroups, AG News, CoNLL-2003 NER), five seeds and two task types, no architectural mutation clears the fitness acceptance threshold when child models are randomly initialised. Without Lamarckian weight inheritance, evolutionary continual learning is entirely paralysed under the constrained per-mutant training budgets typical of online deployment. Key results: - Lamarckian inheritance gap reaches +0.137 entity F1 on CoNLL-2003 NER, scaling with task complexity. - COE Full outperforms all baselines including DER++ on old-task retention across benchmarks. - Structural architecture mutation does not contribute measurably beyond learning rate search at the evaluated scales — the system functions as inheritance-enabled local adaptation rather than genuine architecture evolution. These findings explain a common failure mode in prior evolutionary NAS approaches and establish weight inheritance as the enabling mechanism for evolutionary continual learning under realistic resource constraints. This work extends the EvoTransformer architecture (Zenodo DOI: 10.5281/zenodo.16886541) to the continual learning regime.