![[2602.16438] Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16438] Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment
Summary
The paper explores the bias spillover effect in large language models (LLMs), revealing how targeted fairness alignment can inadvertently worsen disparities across multiple sensitive attributes.
Why It Matters
This research highlights the complexities of fairness in AI, emphasizing the need for context-aware evaluation frameworks. As AI systems increasingly influence societal norms, understanding and mitigating bias spillover is crucial for equitable outcomes in technology.
Key Takeaways
- Targeted fairness alignment in LLMs can lead to bias spillover across various attributes.
- Context-aware analysis reveals significant degradations in fairness under ambiguous conditions.
- Improving fairness for one attribute may worsen disparities in others, necessitating multi-attribute evaluation frameworks.
Computer Science > Machine Learning arXiv:2602.16438 (cs) [Submitted on 18 Feb 2026] Title:Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment Authors:Eva Paraschou, Line Harder Clemmensen, Sneha Das View a PDF of the paper titled Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment, by Eva Paraschou and 2 other authors View PDF HTML (experimental) Abstract:Conventional large language model (LLM) fairness alignment largely focuses on mitigating bias along single sensitive attributes, overlooking fairness as an inherently multidimensional and context-specific value. This approach risks creating systems that achieve narrow fairness metrics while exacerbating disparities along untargeted attributes, a phenomenon known as bias spillover. While extensively studied in machine learning, bias spillover remains critically underexplored in LLM alignment. In this work, we investigate how targeted gender alignment affects fairness across nine sensitive attributes in three state-of-the-art LLMs (Mistral 7B, Llama 3.1 8B, Qwen 2.5 7B). Using Direct Preference Optimization and the BBQ benchmark, we evaluate fairness under ambiguous and disambiguous contexts. Our findings reveal noticeable bias spillover: while aggregate results show improvements, context-aware analysis exposes significant degradations in ambiguous contexts, particularly for physical appearance ($p< 0.001$ across all models), sexual orientation, and disability status. We de...
