Xuran Li
Xuran Li
PhD Candidate · CRUISE Lab · UNSW Sydney

Xuran Li

I am a PhD candidate in the CRUISE Lab at UNSW Sydney, supervised by Prof. Flora Salim, Prof. Hao Xue, and Prof. Imran Razzak. My research spans trustworthy machine learning and LLM personalization, focusing on accuracy–fairness alignment, joint robustness and individual fairness, and constraint-based personalization to build reliable AI systems for positive social impact.

Prior to joining UNSW, I completed my master’s degree at the State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, advised by Profs. Peng Wu and Jian Zhang. Following graduation, I worked as a research assistant with Prof. Yi Zeng at the Institute of Automation, Chinese Academy of Sciences, and subsequently with Prof. Xingjun Ma in the FVL Lab at Fudan University.

Download CV Google Scholar GitHub Email

What I work on

Trustworthy Machine Learning and LLM Personalization
Accurate fairness: two similar loan applicants and the four outcomes Accurate Fair, Accurate Bias, Inaccurate Fair, Inaccurate Bias

Accurate Fairness: Aligning Individual Fairness with Accuracy via Siamese In-Processing

A fair model without accuracy is impractical; an accurate model without fairness is untrustworthy. We introduce accurate fairness—aligning the treatment of similar individuals with ground-truth targets—optimized through a Siamese in-processing scheme and measured via a fairness confusion matrix.

Read the AAAI 2023 paper
The robustness–fairness relationship: Robust Fair, Robust Biased, Unrobust Fair and Unrobust Biased

RIF: Characterizing Both Robustness and Individual Fairness under Semantic Preserving Perturbations

Adversarial edits can simultaneously compromise neural network reliability and equity. A robust model without fairness is biased; a fair model without robustness is fragile. We introduce Robust Individual Fairness (RIF)—aligning invariance across similar individuals with robustness against semantic perturbations—and develop RIFair, a black-box attack using decoupled perturbations to reveal failure modes missed by single-axis evaluations.

Read the KDD 2026 paper
Acceptance regions of four evaluation paradigms for personalised responses

Audit-Based, Constraint-Driven LLM Personalized Evaluation

Over-constraining personalization evaluation falsely rejects valid outputs; under-constraining it falsely accepts misaligned ones. We introduce ConQual-Audit, an audit-based, constraint-driven framework that extracts semantic constraints from profiles and references, qualifies them through pruning and recovery, and verifies candidate responses via natural language inference to achieve reliable evaluation.

Read the preprint and get the code

News

2026
Perturbation Effects on Robustness and Individual Fairness accepted at KDD 2026 for oral presentation (Jeju Island, Korea, August 2026). Oral
2026
New preprint on evaluating LLM personalization with semantic constraint verification (arXiv:2606.16368).
2025
Interpreting Deep Neural Networks via Relative Activation-Deactivation Abstractions accepted by ACM TOSEM.
2023
Accurate Fairness: Improving Individual Fairness without Trading Accuracy accepted at AAAI 2023 for poster presentation (Washington, DC, February 2023). Poster

Publications

Google Scholar has the citation counts
2026
Perturbation Effects on Robustness and Individual Fairnessfirst author
Xuran Li, Hao Xue, Peng Wu, Xingjun Ma, Zhen Zhang, Huaming Chen, Flora D. Salim
Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026)CCF ACORE A*Oral
Project pageDOIarXivCode
2026
Evaluating LLM Personalization via Semantic Constraint Verificationfirst author
Xuran Li, Guanqin Zhang, Imran Razzak, Hakim Hacid, Eleanna Kafeza, Hao Xue, Flora D. Salim
arXiv preprint arXiv:2606.16368
Project pagearXivCode
2025
Interpreting Deep Neural Networks via Relative Activation-Deactivation Abstractions
Zhen Zhang, Peng Wu, Yuting Yang, Xuran Li
ACM Transactions on Software Engineering and Methodology (TOSEM 2025)CCF ACORE A*
DOI
2023
RobustFair: Adversarial Evaluation through Fairness Confusion Directed Gradient Searchfirst author
Xuran Li, Peng Wu, Kaixiang Dong, Zhen Zhang, Yanting Chen
arXiv preprint arXiv:2305.10906
Project pagearXivCode
2023
Accurate Fairness: Improving Individual Fairness without Trading Accuracyfirst author
Xuran Li, Peng Wu, Jing Su
Proceedings of the 37th AAAI Conference on Artificial Intelligence (AAAI 2023), 37(12), 14312–14320CCF ACORE A*Poster
Project pagePaperarXivCode
2022
Adversarial Input Detection Based on Critical Transformation Robustness
Jing Su, Zhen Zhang, Peng Wu, Xuran Li, Jian Zhang
Proceedings of the 33rd IEEE International Symposium on Software Reliability Engineering (ISSRE 2022)CCF BCORE A

Code

Every first-author paper ships with its implementation

AccurateFairnessCriterion

Siamese fairness in-processing and the fairness confusion matrix. Pre-trained baseline and Siamese models for Adult, German Credit, COMPAS and Ctrip, so the evaluation reproduces without training.

GitHubPaper page

RIFair

Black-box adversarial framework for robust individual fairness in NLP: Generate–Filter perturbation space, query-based token ranking, decoupled search for RB / UF / UB pairs, and adversarial retraining.

GitHubPaper page

ConQual-Audit

Audit-based evaluation of LLM personalization: construct constraints from the user profile and task reference, qualify them by pruning and recovery, verify responses with NLI. Benchmarks on PrefEval and LaMP.

GitHubPaper page