Preconditioned Test-Time Adaptation for Out-of-Distribution Debiasing in Narrative Generation

Researchers propose CAP-TTA, a test-time adaptation framework that helps debiased large language models better handle unfamiliar toxic prompts that cause distribution shifts. The method uses context-aware LoRA updates triggered by bias-risk thresholds to reduce toxic outputs while maintaining narrative fluency and reducing computational latency.

• →Debiased LLMs often fail when encountering unfamiliar bias patterns, producing toxic outputs due to distribution shift.
• →CAP-TTA framework performs real-time adaptation using context-aware LoRA updates only when bias-risk exceeds predetermined thresholds.
• →The method achieves significantly lower update latency compared to traditional optimizers like AdamW and SGD.
• →Human evaluation confirms the approach reduces bias while maintaining narrative quality and fluency.
• →The framework mitigates catastrophic forgetting issues common in other debiasing approaches.