Aligning to What? Limits to RLHF Based Alignment
This work addresses the problem of persistent biases in AI alignment for researchers and practitioners, highlighting incremental limitations in existing RLHF approaches.
The study investigated whether RLHF techniques effectively reduce biases in large language models, particularly against African Americans, and found that current alignment methods are inadequate for mitigating covert biases, with SFT before RLHF even calcifying biases.
Reinforcement Learning from Human Feedback (RLHF) is increasingly used to align large language models (LLMs) with human preferences. However, the effectiveness of RLHF in addressing underlying biases remains unclear. This study investigates the relationship between RLHF and both covert and overt biases in LLMs, particularly focusing on biases against African Americans. We applied various RLHF techniques (DPO, ORPO, and RLOO) to Llama 3 8B and evaluated the covert and overt biases of the resulting models using matched-guise probing and explicit bias testing. We performed additional tests with DPO on different base models and datasets; among several implications, we found that SFT before RLHF calcifies model biases. Additionally, we extend the tools for measuring biases to multi-modal models. Through our experiments we collect evidence that indicates that current alignment techniques are inadequate for nebulous tasks such as mitigating covert biases, highlighting the need for capable datasets, data curating techniques, or alignment tools.