LoRA patches stay portable across 10 continual-pretraining updates.

PortLLM claimed training-free, data-free transfer of LoRA patches onto updated base models, but only over short horizons and without theoretical grounding. This study runs 10 continual-pretraining steps on Mistral, Gemma, and Qwen and finds portability persists long-run, meaning repeated fine-tuning isn't required when the base model is periodically refreshed (arXiv 2607.20301). The proposed explanation is geometric: near-orthogonality of vectors in high-dimensional space, analyzed through the loss landscape. If you're re-running fine-tunes on every base-model bump, that's a concrete cost argument for testing whether you need to.