Xiaonan Luo

Computer Science and Engineering

University of Notre Dame

Notre Dame, IN 46556

I am a Ph.D. student in Computer Science and Engineering (CSE) at the University of Notre Dame, where I am fortunate to be advised by Prof. Xiangliang Zhang. Previously, I was a Student Researcher at Bake AI, and I received my B.S. in Data Science and Computer Science from the Hong Kong University of Science and Technology (HKUST).

My recent first-author work includes SecProbe, an adaptive evaluation framework for coding agents on real-world cybersecurity vulnerabilities; KITE, a method for learning from synthetic data without model collapse; and RefineLab (AAAI 2026), which automatically optimizes benchmark datasets toward explicit quality targets.

Research Interests

I study large language models (LLMs), with a focus on:

Recursive Self-Improvement: systems that improve themselves, from self-evolving research agents to models trained on their own data.
Cybersecurity: evaluating and improving coding agents on real-world vulnerabilities.
Agents: long-horizon LLM agents that learn from experience.
Post-Training: reinforcement learning and adaptive reasoning for more capable, reliable models.
Synthetic Data: generating and curating data that targets a model’s weaknesses without collapse.

selected publications

arXiv
SecProbe: Adaptive Evaluation of Coding Agents on Cybersecurity Vulnerabilities

Xiaonan Luo, Yue Huang, Kehan Guo, and 9 more authors

arXiv preprint arXiv:2609.33763, 2026

Abs arXiv Bib HTML PDF Website

SecProbe is an adaptive evaluation framework for coding agents on cybersecurity vulnerabilities. It combines Item Response Theory with on-demand synthesis of repository-scale vulnerability-repair tasks, using observed performance to estimate agent ability and select or generate informative tasks. Across 353 tasks spanning six languages and 151 CWE types, nine frontier models reach at most 28.33% success, and SecProbe matches baseline ability estimates with up to 29.5% fewer tasks.
@article{luo2026secprobe, title = {SecProbe: Adaptive Evaluation of Coding Agents on Cybersecurity Vulnerabilities}, author = {Luo, Xiaonan and Huang, Yue and Guo, Kehan and He, Ping and Zou, Chuan and Gao, Chujie and Li, Lichi and Ma, Yuchen and Xu, Zhangchen and Chen, Zichen and Han, Yufei and Zhang, Xiangliang}, journal = {arXiv preprint arXiv:2609.33763}, year = {2026}, }
arXiv
Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning

Xiaonan Luo, Yue Huang, Kehan Guo, and 4 more authors

arXiv preprint arXiv:2607.17043, 2026

Abs arXiv Bib HTML PDF

We find that model collapse in iterative instruction tuning is a polarization of competence: using DINA-based skill diagnosis, we show that synthetic training can reinforce strong skills while degrading weakly supported ones. We propose KITE, a two-stage synthetic-data framework that couples weakness-guided candidate generation with boundary-aware uncertainty curation, enabling more stable multi-generation improvement than strong synthetic-data baselines across reasoning benchmarks.
@article{luo2026kite, title = {Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning}, author = {Luo, Xiaonan and Huang, Yue and Guo, Kehan and He, Ping and Zou, Chuan and Hua, Ting and Zhang, Xiangliang}, journal = {arXiv preprint arXiv:2607.17043}, year = {2026}, }
AAAI
Better Datasets Start From RefineLab: Automatic Optimization for High-Quality Dataset Refinement

Xiaonan Luo^*, Yue Huang^*, Ping He, and 1 more author

In AAAI Conference on Artificial Intelligence, 2026

Abs arXiv Bib HTML PDF

RefineLab is a framework that formulates QA dataset refinement as an integer linear program, automatically selecting coverage, difficulty, and factual-consistency edits under a token budget to align datasets with explicit quality targets. Its modular refinement operators yield stable evaluation improvements on benchmarks including MMLU, GSM8K, and HumanEval.
@inproceedings{luo2026refinelab, title = {Better Datasets Start From RefineLab: Automatic Optimization for High-Quality Dataset Refinement}, author = {Luo, Xiaonan and Huang, Yue and He, Ping and Zhang, Xiangliang}, booktitle = {AAAI Conference on Artificial Intelligence}, year = {2026}, }
NeurIPS
AdaReasoner: Adaptive Reasoning Enables More Flexible Thinking

Xiangqi Wang, Yue Huang, Yanbo Wang, and 4 more authors

In Advances in Neural Information Processing Systems, 2025

Spotlight

Abs arXiv Bib HTML

AdaReasoner introduces an adaptive reasoning framework that enables large language models to dynamically adjust their reasoning strategies based on problem complexity and context.
@inproceedings{wang2025adareasoner, title = {AdaReasoner: Adaptive Reasoning Enables More Flexible Thinking}, author = {Wang, Xiangqi and Huang, Yue and Wang, Yanbo and Luo, Xiaonan and Guo, Kehan and Zhou, Yujun and Zhang, Xiangliang}, booktitle = {Advances in Neural Information Processing Systems}, year = {2025}, note = {Spotlight}, }