Identity-Robust Language Model Generation via Content Integrity Preservation

TL;DR

The paper investigates why LLM outputs exhibit demographic disparities even when underlying knowledge is stable, showing that identity cues influence the decoding process rather than internal representations. It introduces Identity-Robust Generation (IRG), a training-free, three-stage framework that detects and neutralizes non-critical identity information at the query level and optionally personalizes presentation without compromising content integrity. Empirical results across four benchmarks and 18 identities demonstrate substantial reductions in identity-dependent disparities (average around 77%) while preserving task performance and safety, and show robustness to various identity-expression forms. The work provides a practical, deployable approach to mitigate identity-driven quality degradation in LLMs without retraining, advancing fair and reliable user experiences at scale.

Abstract

Large Language Model (LLM) outputs often vary across user sociodemographic attributes, leading to disparities in factual accuracy, utility, and safety, even for objective questions where demographic information is irrelevant. Unlike prior work on stereotypical or representational bias, this paper studies identity-dependent degradation of core response quality. We show empirically that such degradation arises from biased generation behavior, despite factual knowledge being robustly encoded across identities. Motivated by this mismatch, we propose a lightweight, training-free framework for identity-robust generation that selectively neutralizes non-critical identity information while preserving semantically essential attributes, thus maintaining output content integrity. Experiments across four benchmarks and 18 sociodemographic identities demonstrate an average 77% reduction in identity-dependent bias compared to vanilla prompting and a 45% reduction relative to prompt-based defenses. Our work addresses a critical gap in mitigating the impact of user identity cues in prompts on core generation quality.

Figures (5)

• Figure 1: Identity cues in user prompts can lead to divergent factual or utility answers, even for the same objective question. While model internal knowledge remains stable, user identity can skew generation outcomes. We address this by analyzing identity relevance, generating identity-neutral content, and applying controlled personalization with content verification.
• Figure 2: Discrepancy between generation performance and internal knowledge stability. (Left) Generation accuracy on TruthfulQA varies significantly across user identities, with degraded performance compared to the "no user identity" baseline. (Right) In contrast, internal factual knowledge remains robust across groups, suggesting that user identity biases the generation process despite stable internal representations.
• Figure 3: Workflow of identity-robust language model generation (IRG): Our framework decouples identity-irrelevant content retrieval from identity-aware presentation. Stage 1 detects and removes non-critical demographic cues from the user query. Stage 2 generates identity-neutral content to preserve factuality and utility. Stage 3 is an optional step to reintroduce stylistic personalization for the identity while ensuring content integrity.
• Figure 4: Attribute-specific personalization bias (PB) across different identities within six sociodemographic categories. Our identity-neutral generation consistently reduces performance variance compared to vanilla prompting.
• Figure 5: Readability-based personalization strength and identity-induced bias (PB) across datasets. Vanilla prompting shows limited personalization but high bias; style prompting increases personalization at the cost of larger bias. Our method achieves strong stylistic adaptation while substantially reducing bias.