Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment

Eva Paraschou; Line Harder Clemmensen; Sneha Das

cs.LGFeb 18, 2026

Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment

Authors: Eva Paraschou, Line Harder Clemmensen, Sneha Das

Organizations: Department of Applied Mathematics and Computer Science, Technical University of Denmark, 2800 Lyngby, Denmark · Department of Mathematical Sciences, University of Copenhagen, 2100 Copenhagen, Denmark

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 demonstrate that improving fairness along one attribute can inadvertently worsen disparities in others under uncertainty, highlighting the necessity of context-aware, multi-attribute fairness evaluation frameworks.

Figures & tables

Figure 1. The methodological pipeline of our work consists of extracting the sensitive attributes, assessing the baseline fairness across all sensitive attributes in the pre-aligned LLMs, performing targeted LLMs alignment to eliminate gender bias, and re-assessing all sensitive attributes in the post-aligned LLMs to explore the bias spillover effect.

Panel of Figure 2 — Figure 2. Samples of the four possible combinations of the contexts (disambiguous and ambiguous) and polarities (non-negative and negative) of the BBQ benchmark.

Figure 3. Comparative analysis of the context-unaware results of Mistral 7B (orange), Llama 3.1 8B (blue), Qwen 2.5 7B (purple) across all sensitive attributes. The light shades correspond to the pre-alignment accuracy, while the dark shades to the post-alignment one. The number of asterisks represent the level of significance in the difference.

Panel of Figure 4 — Figure 4. The two new alignment questions that derived from transforming an original alignment question (see 2c ).

Figure 5. Comparative analysis of the ambiguous questions of Mistral 7B (orange), Llama 3.1 8B (blue), Qwen 2.5 7B (purple) across all sensitive attributes. The light shades correspond to the pre-alignment accuracy, while the dark shades to the post-alignment one. The number of asterisks represent the level of significance.

Figure 6. Comparative analysis of the disambiguous questions of Mistral 7B (orange), Llama 3.1 8B (blue), Qwen 2.5 7B (purple) across all sensitive attributes. The light shades correspond to the pre-alignment accuracy, while the dark shades to the post-alignment one. The number of asterisks represent the level of significance.

Sensitive	Mistral 7B			Llama 3.1 8B			Qwen 2.5 7B
Attribute	Agg.	Amb.	Dis.	Agg.	Amb.	Dis.	Agg.	Amb.	Dis.
Gender (align*)	+11.07%	+13.43%	+8.82%	+0.05%	+0.16%	-0.05%	+0.03%	+0.00%	+0.05%
Gender (eval*)	+8.41%	+10.47%	+6.44%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%
Transgender	+8.72%	+6.98%	+10.51%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%
Age	+10.56%	+10.37%	+10.76%	+0.33%	+0.66%	+0.00%	+0.00%	+0.00%	+0.00%
Disability status	+4.95%	+5.32%	+4.58%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%	+0.00%

Table 1. The response rate improvements across the context-unaware (Agg.) setup, as well as ambiguous (Amb.), and disambiguous (Dis.) contexts for Mistral 7B, Llama 3.1 8B, and Qwen 2.5 7B and all sensitive attributes.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	71.64%	93.83%	+22.19%	<0.001
Gender (eval*)	73.68%	94.34%	+20.66%	<0.001
Transgender	72.82%	86.50%	+13.68%	<0.001
Age	65.36%	78.20%	+12.84%	<0.001
Disability status	68.54%	83.07%	+14.53%	<0.001

Appendix

Table 2. Comparative analysis of Mistral 7B alignment accuracy before and after alignment process, across all nine sensitive attributes.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	79.06%	93.66%	+14.60%	<0.001
Gender (eval*)	78.46%	93.65%	+15.19%	<0.001
Transgender	72.22%	83.67%	+11.45%	<0.001
Age	64.34%	72.31%	+7.97%	<0.001
Disability status	69.54%	72.11%	+2.57%	0.007

Appendix

Table 3. Comparative analysis of Llama 3.1 8B alignment accuracy before and after alignment process, across all nine sensitive attributes.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	94.36%	98.39%	+4.03%	<0.001
Gender (eval*)	94.07%	98.44%	+4.37%	<0.001
Transgender	83.33%	85.30%	+1.97%	0.001
Age	80.05%	85.87%	+5.82%	<0.001
Disability status	88.82%	93.38%	+4.56%	<0.001

Appendix

Table 4. Comparative analysis of Qwen2.5 7B alignment accuracy before and after alignment process, across all nine sensitive attributes.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	89.13%	99.37%	+10.24%	<0.001
Gender (eval*)	88.60%	99.37%	+10.77%	<0.001
Transgender	89.03%	93.47%	+4.44%	<0.001
Age	74.63%	76.57%	+1.94%	0.024
Disability status	76.94%	79.14%	+2.20%	0.045

Appendix

Table 5. Comparative analysis of Mistral 7B alignment accuracy before and after alignment process, across all nine sensitive attributes in the ambiguous context.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	72.20%	90.18%	+17.98%	<0.001
Gender (eval*)	69.58%	89.38%	+19.80%	<0.001
Transgender	66.44%	70.37%	+3.93%	0.074
Age	42.07%	48.42%	+6.35%	<0.001
Disability status	50.77%	46.79%	-3.98%	0.011

Appendix

Table 6. Comparative analysis of Llama 3.1 8B alignment accuracy before and after alignment process, across all nine sensitive attributes in the ambiguous context.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	98.70%	99.95%	+1.25%	1.611
Gender (eval*)	99.38%	100%	+0.62%	0.250
Transgender	83.33%	96.99%	+13.66%	0.453
Age	69.39%	77.77%	+8.38%	<0.001
Disability status	91.13%	95.37%	+4.24%	<0.001

Appendix

Table 7. Comparative analysis of Qwen2.5 7B alignment accuracy before and after alignment process, across all nine sensitive attributes in the ambiguous context.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	54.89%	88.29%	+33.40%	<0.001
Gender (eval*)	59.78%	89.35%	+29.57%	<0.001
Transgender	56.15%	79.58%	+23.43%	<0.001
Age	56.19%	79.81%	+23.62%	<0.001
Disability status	60.24%	86.98%	+26.74%	<0.001

Appendix

Table 8. Comparative analysis of Mistral 7B alignment accuracy before and after alignment process, across all nine sensitive attributes in the disambiguous context.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	85.91%	97.14%	+11.23%	<0.001
Gender (eval*)	87.32%	97.92%	+10.60%	<0.001
Transgender	78.01%	96.99%	+18.98%	<0.001
Age	86.47%	96.20%	+9.73%	<0.001
Disability status	88.30%	97.43%	+9.13%	<0.001

Appendix

Table 9. Comparative analysis of Llama 3.1 8B alignment accuracy before and after alignment process, across all nine sensitive attributes in the disambiguous context.

Sensitive attribute	Pre-alignment	Post-alignment	Accuracy	p-value
	accuracy	accuracy	improv.
Gender (align*)	90.01%	96.83%	+6.82%	<0.001
Gender (eval*)	88.77%	96.88%	+8.11%	<0.001
Transgender	70.37%	73.61%	+3.24%	0.003
Age	90.71%	93.97%	+3.26%	<0.001
Disability status	86.50%	91.39%	+4.89%	<0.001

Appendix

Table 10. Comparative analysis of Qwen2.5 7B alignment accuracy before and after alignment process, across all nine sensitive attributes in the disambiguous context.

Sensitive	Total	Mistral 7B			Llama 3.1 8B			Qwen 2.5 7B
Attribute	$N$	Pre	Post	Improv.	Pre	Post	Improv.	Pre	Post	Improv.
		$n$	$n$	%	$n$	$n$	%	$n$	$n$	%
Gender (align*)	3,847	3,441	3,822	+11.07%	3,844	3,846	+0.05%	3,846	3,847	+0.03%
Gender (eval*)	961	880	954	+8.41%	961	961	+0.00%	961	961	+0.00%
Transgender	864	791	860	+8.72%	864	864	+0.00%	864	864	+0.00%
Age	3,680	3,295	3,643	+10.56%	3,668	3,680	+0.33%	3,680	3,680	+0.00%

Appendix

Table 11. The responses provided by the LLMs in the context-unaware setup in each phase (i.e. pre vs post) and the response rate improvements across Mistral 7B, Llama 3.1 8B, and Qwen 2.5 7B models and all the sensitive attributes.

Sensitive	Total	Mistral 7B			Llama 3.1 8B			Qwen 2.5 7B
Attribute	$N$	Pre	Post	Improv.	Pre	Post	Improv.	Pre	Post	Improv.
		$n$	$n$	%	$n$	$n$	%	$n$	$n$	%
Gender (align*)	1,924	1,683	1,909	+13.43%	1,921	1,924	+0.16%	1,924	1,924	+0.00%
Gender (eval*)	480	430	475	+10.47%	480	480	+0.00%	480	480	+0.00%
Transgender	432	401	429	+6.98%	432	432	+0.00%	432	432	+0.00%
Age	1,840	1,640	1,810	+10.37%	1,828	1,840	+0.66%	1,840	1,840	+0.00%

Appendix

Table 12. The responses provided by the LLMs in the ambiguous context setup in each phase (i.e. pre vs post) and the response rate improvements across Mistral 7B, Llama 3.1 8B, and Qwen 2.5 7B models.

Sensitive	Total	Mistral 7B			Llama 3.1 8B			Qwen 2.5 7B
Attribute	$N$	Pre	Post	Improv.	Pre	Post	Improv.	Pre	Post	Improv.
		$n$	$n$	%	$n$	$n$	%	$n$	$n$	%
Gender (align*)	1,923	1,758	1,913	+8.82%	1,923	1,922	-0.05%	1,922	1,923	+0.05%
Gender (eval*)	481	450	479	+6.44%	481	481	+0.00%	481	481	+0.00%
Transgender	432	390	431	+10.51%	432	432	+0.00%	432	432	+0.00%
Age	1,840	1,655	1,833	+10.76%	1,840	1,840	+0.00%	1,840	1,840	+0.00%

Appendix

Table 13. The responses provided by the LLMs in the disambiguous context setup in each phase (i.e. pre vs post) and the response rate improvements across Mistral 7B, Llama 3.1 8B, and Qwen 2.5 7B models.

LLM Alignment

Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.

See all 611 papers

Intersectional Fairness in Large Language Models
Apr 22, 2026Chaima Boufaied, Ronnie De Souza Santos, Ann BarcombAlgorithmic Fairness Language Model Bias Evaluation
University of Calgary, Calgary, Canada
Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study
Mar 25, 2026Nour Bouchouchi, Thibault Laugel, Xavier Renard +3Gender Bias in Language Models LLM Alignment
Sorbonne Université, CNRS, LIP6, F-75005 Paris, France · AXA, Paris, France · Polish Academy of Science, IBS PAN, Warsaw, Poland
Safety Is Not Universal: The Selective Safety Trap in LLM Alignment
Jan 7, 2026Iago Alves Brito, Walcy Santos Rezende Rios, Julia Soares Dollis +2LLM Safety Benchmarks LLM Alignment
Advanced Knowledge Center for Immersive Technologies (AKCIT) Federal University of Goiás (UFG)

Apr 22, 2026J/K move · Enter open · S save

Intersectional Fairness in Large Language Models

Chaima Boufaied, Ronnie De Souza Santos, Ann Barcomb

University of Calgary, Calgary, Canada

Algorithmic Fairness Language Model Bias Evaluation Intersectional Fairness Demographic Bias in Language Models

Open paper AskPDF arXiv

Large Language Models (LLMs) are increasingly deployed in socially sensitive settings, raising concerns about fairness and biases, particularly across intersectional demographic attributes. In this paper, we systematically evaluate intersectional fairness in six LLMs using ambiguous and disambiguated contexts from two benchmark datasets. We assess LLM behavior using bias scores, subgroup fairness metrics, accuracy, and consistency through multi-run analysis across contexts and negative and non-negative question polarities. Our results show that while modern LLMs generally perform well in ambiguous contexts, this limits the informativeness of fairness metrics due to sparse non-unknown predictions. In disambiguated contexts, LLM accuracy is influenced by stereotype alignment, with models being more accurate when the correct answer reinforces a stereotype than when it contradicts it. This pattern is especially pronounced in race-gender intersections, where directional bias toward stereotypes is stronger. Subgroup fairness metrics further indicate that, despite low observed disparity in some cases, outcome distributions remain uneven across intersectional groups. Across repeated runs, responses also vary in consistency, including stereotype-aligned responses. Overall, our findings show that apparent model competence is partly associated with stereotype-consistent cues, and no evaluated LLM achieves consistently reliable or fair behavior across intersectional settings. These findings highlight the need for evaluation beyond accuracy, emphasizing the importance of combining bias, subgroup fairness, and consistency metrics across intersectional groups, contexts, and repeated runs.

Abstract

LLM Alignment

Explore similar work

Intersectional Fairness in Large Language Models

Alignment Reduces Expressed but Not Encoded Gender Bias: A Unified Framework and Study

Safety Is Not Universal: The Selective Safety Trap in LLM Alignment