On the Role of Citations in Preference Data

Yu Hou; Hal Daumé; Rachel Rudinger; William Walden

cs.CLJul 6, 2026

On the Role of Citations in Preference Data

Authors: Yu Hou, Hal Daumé, Rachel Rudinger, William Walden

Organizations: University of Maryland · Johns Hopkins University

Abstract

Many NLP tasks require systems to provide attribution in their outputs--i.e. citations to grounding sources. Attribution serves as a bulwark against model hallucination and as a means for users to verify the credibility of model outputs. Yet, it is unclear how humans and LLMs evaluate citations when comparing outputs, a process central to reward modeling and modern LLM post-training. This paper studies the role of citations in the preferences of human judges and four open-source LLMs within the context of scientific question answering, leveraging mixed effects models to investigate the influence of citations on pairwise judgments. Among our key findings are (1) that humans prefer more diverse citations but fewer overall, and (2) that LLMs show some citation-related preferences compared to humans, despite lacking access to the sources, but these preferences depend on the data and specific models. We further discuss the implications of our findings for preference data collection.

Figures & tables

Figure 1: A pair of citation-attributed model responses to a question from SciArena Zhao et al. (2025a) . We use mixed-effects models to understand the impact of citations on human and LLM preferences on such pairs.

Figure 2: Model ranking correlation between Elo Ranking (SciArena) and Random-Effects Ranking (Human).

Panel of Figure 3 — Figure 3: SciArena : Estimated fixed effects on human votes and LLM judgments ( $\Delta\text{Win Rate}$ ). Sub-figures (a)–(c) show citation-related predictors, while (d) and (e) show non-citation control variables for effect size comparison. Error bars represent 95% CIs. Significance: *** $p<0.001$ ; ** $p<0.01$ ; * $p<0.05$ .

Panel of Figure 4 — Figure 4: ResearchQA : Estimated fixed effects on human votes and LLM judgments ( $\Delta\text{Win Rate}$ ). Sub-figures (a)–(c) show citation-related predictors, while (d) and (e) show non-citation control variables for effect size comparison. Error bars represent 95% CIs. Significance: *** $p<0.001$ ; ** $p<0.01$ ; * $p<0.05$ .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

	Qwen	Gemma	Llama	Skywork
CoT	.80 (90%)	.67 (84%)	.72 (88%)	.89 (95%)
LD	.96 (98%)	.93 (97%)	.93 (97%)	.98 (99%)

Appendix

Table 1: Agreement between each prompt variant and the original prompt, reported as Cohen’s

\kappa

with raw agreement in parentheses. LD: Length Debiasing.

Judge	Predictor	$\Delta$ Win Rate	95% CI	$p$
Human	Citation Diversity	+1.78 ∗	[+0.06, +3.51]	0.043
	Citation Count	-3.55 ∗∗∗	[-5.34, -1.75]	$<$ 0.001
	Citation Age	+0.66	[-0.33, +1.66]	0.193
	Response Length	-1.21	[-3.40, +0.99]	0.283
	Response Content	+4.21 ∗	[+1.00, +7.43]	0.010
Llama	Citation Diversity	+7.47 ∗∗∗	[+5.55, +9.39]	$<$ 0.001

Appendix

Table 2: Estimated fixed effects on human votes and LLM judgments for SciArena , reported as changes in win rate (percentage points) for a one standard deviation increase in each standardized predictor. This table reports the exact estimates underlying Figure 3 . Significance:

{}^{***}p<0.001

;

{}^{**}p<0.01

;

{}^{*}p<0.05

Judge	Predictor	$\Delta$ Win Rate	95% CI	$p$
Human	Citation Diversity	+11.65 ∗∗∗	[+5.31, +17.99]	$<$ 0.001
	Citation Count	-6.47 ∗	[-12.88, -0.05]	0.048
	Citation Mismatch	-10.08 ∗∗	[-17.01, -3.15]	0.004
	Response Length	+21.41 ∗∗∗	[+13.86, +28.96]	$<$ 0.001
	Response Content	-0.22	[-8.66, +8.22]	0.959
Llama	Citation Diversity	+7.76 ∗∗∗	[+3.99, +11.54]	$<$ 0.001

Appendix

Table 3: Estimated fixed effects on human votes and LLM judgments for ResearchQA , reported as changes in win rate (percentage points) for a one standard deviation increase in each standardized predictor. This table reports the exact estimates underlying Figure 4 . Significance:

{}^{***}p<0.001

;

{}^{**}p<0.01

;

{}^{*}p<0.05

Preference Modeling

See all 45 papers

Abstention vs. Hallucination: Benchmarking LLM Source Attribution for Scientific Citations
May 3, 2024Deepa Tilwani, Yash Saxena, Seyedali Mohammadi +5Citation Verification RAG Evaluation
Artificial Intelligence Institute, University of South Carolina, Columbia, South Carolina, USA · University of Maryland, Baltimore County, Baltimore, Maryland, USA · IAIRO, India +1
Do You Need a Frontier Model as a Citation Verifier? Benchmarking Rubric LLMs for Deep-Research Source Attribution
Jul 9, 2026Ethan Leung, Elias Lumer, Corey Feld +3LLM-as-a-Judge Citation Verification
Commercial Technology and Innovation Office, PricewaterhouseCoopers, U.S.
Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation
Sep 1, 2026Yixuan Liu, Lin Chen, Zhuoqi Liu +2AI-Assisted Scientific Research
1Network Science Institute, Northeastern University · Department of Physics, Northeastern University · Department of Electrical and Computer Engineering, Northeastern University +1

May 3, 2024J/K move · Enter open · S save

Abstention vs. Hallucination: Benchmarking LLM Source Attribution for Scientific Citations

Deepa Tilwani, Yash Saxena, Seyedali Mohammadi, Ankur Padia+4

Artificial Intelligence Institute, University of South Carolina, Columbia, South Carolina, USA · University of Maryland, Baltimore County, Baltimore, Maryland, USA · IAIRO, India+1

Citation Verification RAG Evaluation Citation Hallucination LLM Abstention

Open paper AskPDF arXiv

Large language models (LLMs) increasingly generate citation-backed responses, yet citation hallucination remains a major challenge for trustworthy scientific information access. We introduce REASONS, a benchmark of 12,723 sentence-level citation instances spanning 12 arXiv subject categories, designed to evaluate scientific citation attribution under varying evidence conditions. We propose a dual-metric framework consisting of Abstention Rate (AR) and Hallucination Rate (HR) to characterize the trade-off between reliability and responsiveness. Using author-attribution and title-attribution tasks, we evaluate proprietary and open-source LLMs under zero-context, metadata-augmented, cascaded metadata-augmented prompting (CMP), retrieval-augmented, and adversarial settings. Advanced RAG lowers HR relative to Naive RAG (65.4% vs. 87.6%) but reduces AR from 5.0% to 0%. Under adversarial metadata, several systems exceed 85% HR, while retrieval-augmented variants frequently maintain near-zero abstention. Human evaluation of 1,000 outputs (

κ=0.78

) finds a 12.7:1 ratio of factual hallucinations to acceptable paraphrases. Our findings demonstrate that citation attribution systems should be evaluated not only for correctness but also for their ability to abstain appropriately under uncertainty. REASONS provides a benchmark and evaluation framework for studying attribution reliability in citation generation.

Abstract

Preference Modeling

Explore similar work

Abstention vs. Hallucination: Benchmarking LLM Source Attribution for Scientific Citations

Do You Need a Frontier Model as a Citation Verifier? Benchmarking Rubric LLMs for Deep-Research Source Attribution

Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation