Mechanism-Guided Recovery from Pruning-Induced Fragility in Compressed Language Models: Insights from Antidepressant Analogs

Background: Aggressive pruning enables efficient deployment of large language models but often induces latent fragility, manifesting as preserved procedural skills alongside eroded factual knowledge and reasoning—a pattern reminiscent of synaptic loss in major depressive disorder under chronic stress.Objective: We sought to compare recovery mechanisms inspired by antidepressant classes in realistically pruned transformers, introducing an isodose framework to equate computational cost and isolate mechanistic effects.Methods: Low-rank adaptation (LoRA) was applied to a severely width-pruned Llama-3.2-1B model (60% MLP neurons removed) exhibiting documented fragility. Three interventions—ketamine-like high-capacity regrowth, SSRI-like low-rank gradual refinement, and neurosteroid-like high-dropout stabilization—were calibrated to equivalent estimated FLOPs (~1.35 × 10¹⁴). Performance was assessed on ARC-Easy and LAMBADA subsets (composite score), with acute relapse simulated via 30% additional pruning and longitudinal durability across eight 10% pruning cycles. A follow-up sweep optimized ketamine-like rank, with generalization tested on alternative pruned architectures.Results: Ketamine-like adaptation yielded superior acute recovery (+2.1% composite) and efficiency (15.4 recovery/PetaFLOP), with lowest relapse drop under equated budgets. SSRI-like showed greatest longitudinal stability (smallest total decline), while neurosteroid-like occupied an intermediate profile. Rank optimization identified moderate capacity (r=32) as optimal (+3.5% recovery, high efficiency), generalizing robustly (+4.0% mean) across width- and depth-pruning regimes.Conclusion: Moderate structural regrowth emerges as efficient and transferable for mitigating compression-induced fragility, highlighting mechanistic trade-offs with implications for robustness-preserving pruning and, speculatively, plasticity-targeted depression therapeutics.