Reproduce “HumT DumT: Measuring and controlling human-like language in LLMs” on OpenResearch

About this paper

Myra Cheng, Sunny Yu, Dan Jurafsky

Researchers at Stanford University developed HUMT and SOCIOT, computational metrics to measure human-like tone and associated social perceptions in large language model outputs, and DUMT, a method to reduce it. They empirically found that users generally prefer less human-like LLM responses, and that highly human-like language in LLMs often correlates with social perceptions such as low status and femininity.

Key insights

• Human-like language in LLMs largely emerges from post-training processes like instruction-tuning and Reinforcement Learning from Human Feedback (RLHF), rather than solely from pre-training data.
• Users generally disprefer highly human-like language from LLMs, a consistent finding across various preference datasets, with preferences varying by topic (e.g., strong dispreference for sensitive topics).
• High human-like tone in LLM outputs strongly correlates with specific social perceptions, including warmth, social closeness, and femininity, and shows a strong negative correlation with status, indicating potential for reinforcing stereotypes and fostering overreliance.

Summary by alphaXiv.