Xiaoyang Liu
I am a third-year PhD candidate advised by Prof. Tao Luo at School of Mathematical Sciences, Shanghai Jiao Tong University (SJTU).
Since July 2026, I have been working as a Post-Training Research Intern at the Wizard Intelligence Learning Lab (WILL).
Research
LLM · Lean · Post-Training · AI4Math
I work on large language models and post-training, with a focus on agentic systems, data synthesis, and model evaluation. My research connects LLMs with Lean to advance automated theorem proving and mathematical formalization.
Agentic Systems
Developing LLM-based agentic systems for theorem proving and mathematical formalization, with the goal of supporting mathematical research through reliable automation.
Data Synthesis
Synthesizing high-quality training data and fine-tuning LLMs through expert iteration in ATLAS, and neuro-symbolic data generation and curriculum learning in DSR.
Model Evaluation
Developing semantic and structural evaluation with ASSESS, actionable error diagnosis and correction with FormalRx, and adversarial specification testing with VeriScale.
Publications
* Equal contribution. † Corresponding author.
ICML 2026 · AI4Math Workshop
VeriScale: Adversarial Test-Suite Scaling for Verifiable Code Generation
An adversarial framework that expands and reduces test suites to expose unsound or incomplete formal specifications in verifiable code generation.
ICML 2026
Decompose, Structure, and Repair: A Neuro-Symbolic Framework for Autoformalization via Operator Trees
A neuro-symbolic framework that structures autoformalization with operator trees and repairs failed Lean outputs.
ICML 2026
FormalRx: Rectify and eXamine Semantic Failures in Autoformalization
A framework for detecting, examining, and rectifying semantic failures in autoformalized statements.
ICLR 2026
ASSESS: A Semantic and Structural Evaluation Framework for Statement Similarity
An evaluation framework that combines semantic and structural signals to compare formal statements.
NeurIPS 2025
ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data
A data-centric approach to theorem autoformalization through lifting, augmentation, and synthesis.
ICML 2025 · AI4Math Workshop
Generalized Tree Edit Distance (GTED): A Faithful Evaluation Metric for Statement Autoformalization
A tree-edit-distance metric designed to evaluate the structure and meaning of autoformalized statements.