QLoRA Rank Induces Acquisition-Retention Tradeoff in Factual Learning

Analysis of Qwen3-4B reveals that increasing QLoRA rank improves factual acquisition but degrades out-of-domain performance. Higher ranks facilitate better paraphrase generalization for new facts, while full fine-tuning maintains better general capabilities but reaches a lower ceiling for factual acquisition.