Ph.D. Student, School of Computer Science, Shanghai Jiao Tong UniversityHi! I am Xuan Gong (Xander Gong), a first-year Ph.D. student in the School of Computer Science at Shanghai Jiao Tong University.
My research focuses on the optimization and alignment of foundation models, with an emphasis on improving the reasoning, factuality, agentic capabilities, safety, and science applications of large language models and vision-language models (maybe world models in the future).
If you are interested in partnering on research projects, offering internship opportunities or exchange programs, I would be thrilled to connect with you.
Education
-
Shanghai Jiao Tong UniversitySchool of Computer Science
Ph.D. StudentSep. 2025 - present -
Tongji UniversityB.S. in Computer Science (GPA rank top 5%)Sep. 2021 - Jul. 2025
Honors & Awards
-
National Scholarship of China2022
-
Outstanding Graduate of Tongji University2025
-
First Prize of The 14th National University Student Mathematics Competition Non-Mathematics Category (Shanghai Region)2022
News
Selected Publications on Google Scholar (view all )
RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks
Hanbo Huang, Yiran Zhang, Hao Zheng, Xuan Gong, Yihan Li, Lin Liu, Zhuotao Liu, Shiyu Liang# (# corresponding author)
ICML 2026
RLCracker studies adaptive reinforcement-learning attacks against LLM watermarks, exposing watermark vulnerabilities under learned black-box attack policies.
RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks
Hanbo Huang, Yiran Zhang, Hao Zheng, Xuan Gong, Yihan Li, Lin Liu, Zhuotao Liu, Shiyu Liang# (# corresponding author)
ICML 2026
RLCracker studies adaptive reinforcement-learning attacks against LLM watermarks, exposing watermark vulnerabilities under learned black-box attack policies.
VCORE: Variance-Controlled Optimization-based Reweighting for Chain-of-Thought Supervision
Xuan Gong, Senmiao Wang, Hanbo Huang, Ruoyu Sun, Shiyu Liang# (# corresponding author)
ACL 2026 Main
VCORE introduces variance-controlled optimization-based reweighting for chain-of-thought supervision, improving how reasoning traces contribute to model training.
VCORE: Variance-Controlled Optimization-based Reweighting for Chain-of-Thought Supervision
Xuan Gong, Senmiao Wang, Hanbo Huang, Ruoyu Sun, Shiyu Liang# (# corresponding author)
ACL 2026 Main
VCORE introduces variance-controlled optimization-based reweighting for chain-of-thought supervision, improving how reasoning traces contribute to model training.
Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
Xuan Gong, Hanbo Huang, Hao Zheng, Yiran Zhang, Wenbin Dai, Weishu Zhao, Shiyu Liang# (# corresponding author)
CompLearn Workshop @ ICML 2026
This work introduces reflection anchors for propagation-aware visual retention, targeting long-chain multimodal reasoning where visual evidence must remain reliable across extended inference.
Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
Xuan Gong, Hanbo Huang, Hao Zheng, Yiran Zhang, Wenbin Dai, Weishu Zhao, Shiyu Liang# (# corresponding author)
CompLearn Workshop @ ICML 2026
This work introduces reflection anchors for propagation-aware visual retention, targeting long-chain multimodal reasoning where visual evidence must remain reliable across extended inference.
From Parameters to Prompts: Understanding and Mitigating the Factuality Gap between Fine-Tuned LLMs
Xuan Gong*, Hanbo Huang*, Yiran Zhang*, Shiyu Liang# (* equal contribution, # corresponding author)
ICASSP 2026
We revisit how supervised fine-tuning affects factual knowledge in LLMs, revealing a factuality gap between known and unknown knowledge. This gap can be mitigated at inference via in-context learning (ICL) or out-of-distribution prompts. Our theoretical and empirical results show that test-time prompts can overshadow fine-tuning data, suggesting ICL can compensate for poor fine-tuning and should be considered in evaluating fine-tuning strategies.
From Parameters to Prompts: Understanding and Mitigating the Factuality Gap between Fine-Tuned LLMs
Xuan Gong*, Hanbo Huang*, Yiran Zhang*, Shiyu Liang# (* equal contribution, # corresponding author)
ICASSP 2026
We revisit how supervised fine-tuning affects factual knowledge in LLMs, revealing a factuality gap between known and unknown knowledge. This gap can be mitigated at inference via in-context learning (ICL) or out-of-distribution prompts. Our theoretical and empirical results show that test-time prompts can overshadow fine-tuning data, suggesting ICL can compensate for poor fine-tuning and should be considered in evaluating fine-tuning strategies.
RLSpoofer: A Sample-Efficient Black-Box Spoofing Attack for Stress-Testing LLM Watermarks
Hanbo Huang, Xuan Gong, Yiran Zhang, Hao Zheng, Wenbin Dai, Jie Kuang, Shiyu Liang# (# corresponding author)
Trustworthy AI for Good (AI4GOOD) Workshop @ ICML 2026
This workshop paper studies sample-efficient black-box spoofing attacks for stress-testing the robustness of LLM watermarks.
RLSpoofer: A Sample-Efficient Black-Box Spoofing Attack for Stress-Testing LLM Watermarks
Hanbo Huang, Xuan Gong, Yiran Zhang, Hao Zheng, Wenbin Dai, Jie Kuang, Shiyu Liang# (# corresponding author)
Trustworthy AI for Good (AI4GOOD) Workshop @ ICML 2026
This workshop paper studies sample-efficient black-box spoofing attacks for stress-testing the robustness of LLM watermarks.
DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination
Xuan Gong, Tianshi Ming, Xinpeng Wang, Zhihua Wei# (# corresponding author)
EMNLP 2024 Main
We propose DAMRO, a training-free method to reduce object hallucination in LVLMs by filtering misleading high-attention background tokens using the ViT CLS token. DAMRO significantly improves hallucination control on models like LLaVA and InstructBLIP across multiple benchmarks.
DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination
Xuan Gong, Tianshi Ming, Xinpeng Wang, Zhihua Wei# (# corresponding author)
EMNLP 2024 Main
We propose DAMRO, a training-free method to reduce object hallucination in LVLMs by filtering misleading high-attention background tokens using the ViT CLS token. DAMRO significantly improves hallucination control on models like LLaVA and InstructBLIP across multiple benchmarks.