Mitigating LLM biases toward spurious social contexts using direct preference optimization

Researchers developed Debiasing-DPO, a new training method that reduces harmful biases in large language models by 84% while improving accuracy by 52%. The study found that LLMs can shift predictions by up to 1.48 points when exposed to irrelevant contextual information like demographics, highlighting critical risks for high-stakes AI applications.

• →Large language models show significant sensitivity to spurious contextual information, potentially shifting predictions by up to 1.48 points on assessment scales.
• →Larger AI models sometimes exhibit greater bias sensitivity despite having higher predictive accuracy.
• →Standard bias mitigation techniques like prompts and direct preference optimization prove largely insufficient.
• →The new Debiasing-DPO method reduces model bias by 84% while improving predictive accuracy by 52% on average.
• →Model scaling alone does not naturally produce robustness to spurious contexts, requiring specialized training approaches.