Tag
This paper proposes CAP-TTA, a test-time adaptation framework that uses preconditioned LoRA updates triggered by bias-risk scores to mitigate toxicity and bias in large language models during narrative generation, achieving faster optimization and better fluency than standard baselines.