Exploring Safety-Utility Trade-Offs in Personalized Language Models

TL;DR

This paper investigates how personalized large language models exhibit biases affecting safety and utility, revealing significant performance variance across user identities and proposing mitigation strategies.

Contribution

It quantifies personalization bias in LLMs along safety and utility axes and evaluates its impact across multiple models, introducing mitigation approaches.

Findings

LLMs show significant safety-utility trade-off variance based on user identity

Personalization bias affects models like Llama, Mistral, GPT-3.5, and GPT-4o

Mitigation strategies such as preference tuning can reduce personalization bias

Abstract

As large language models (LLMs) become increasingly integrated into daily applications, it is essential to ensure they operate fairly across diverse user demographics. In this work, we show that LLMs suffer from personalization bias, where their performance is impacted when they are personalized to a user's identity. We quantify personalization bias by evaluating the performance of LLMs along two axes - safety and utility. We measure safety by examining how benign LLM responses are to unsafe prompts with and without personalization. We measure utility by evaluating the LLM's performance on various tasks, including general knowledge, mathematical abilities, programming, and reasoning skills. We find that various LLMs, ranging from open-source models like Llama (Touvron et al., 2023) and Mistral (Jiang et al., 2023) to API-based ones like GPT-3.5 and GPT-4o (Ouyang et al., 2022), exhibit significant variance in performance in terms of safety-utility trade-offs depending on the user's identity. Finally, we discuss several strategies to mitigate personalization bias using preference tuning and prompt-based defenses.