Kai Ye

👤 About Me

1. I am a PhD student in the Department of Statistics at The London School of Economics and Political Science (LSE), advised by Chengchun Shi.
2. I am also the Chief Technology Officer (CTO) at Stats-Powered AI.
3. Prior to this, I completed an MSc in Applicable Mathematics at LSE.
4. Before that, I received a BSc in Mathematics with Finance from The University of Liverpool and Xi'an Jiaotong-Liverpool University.

• AWS Academy Educator

🔬 Research

My current research focuses on the intersection of reinforcement learning and large language models, with particular interest in:

• Reinforcement learning and policy optimization
• Confounded and partially observable environments (POMDPs)
• Human feedback alignment and preference-based learning
• Doubly robust estimation and variance-reduction techniques
• High-dimensional statistics and distributional analysis

📝 Selected Preprints

• 2026
  Research

  Can Tabular Foundation Models Amortize Statistical Inference?

  Ye, K., Gong, S., Zhou, H., Zangirolami, V. and Shi, C.

  TabCon: amortized confidence intervals with tabular foundation models.
• 2026
  Research

  Detecting LLM-Generated Tokens in Human–LLM Coauthored Text

  Lu, Y., Zhou, H., Spill, F., Ye, K., Shi, C. and Zhu, J.

  Token-level detection in text coauthored by humans and LLMs.
• 

  READER: Reasoning-Enhanced AI-Generated Text Detection

  Su, P^*., Ye, K^*., Gong, S., Xu, E., Zhu, J., Livieri, G. and Shi, C.

  READER.
• 

  Kernelized Advantage Estimation: From Nonparametric Statistics to LLM Reasoning

  Gong, S^*., Ye, K^*., Zhu, J., Zhang, X., Zhou, H. and Shi, C.

  KAE.
• 2026
  Research

  Demystifying Group Relative Policy Optimization: Its Policy Gradient is a U-Statistic

  Zhou, H^*., Ye, K^*., Xu, E., Zhu, J., Yang, Y., Gong, S. and Shi, C.

  Demystifying GRPO

📄 Selected Publications

• 

  Robust Reinforcement Learning from Human Feedback for Large Language Models Fine-Tuning (AOAS 2026 · Accepted)

  Ye, K^*., Zhou, H^*., Zhu, J^*., Quinzan, F. and Shi, C.

  VRPO.
• 

  BASIS: Batchwise Advantage Estimation from Single-Rollout Information Sharing for LLM Reasoning (EMNLP (Findings) 2026 · Accepted)

  Gong, S^*., Xu, E^*., Ye, K^*., Livieri, G., Quinzan, F. and Shi, C.

  BASIS: single-rollout information sharing for LLM reasoning.
• 

  A Difference-in-Difference Approach to Detecting AI-Generated Images (CVPR 2026)

  Qi, X^*., Ye, K^*., Shi, C., Yang, Y., Zhu, J. and Zhou, H.

  DID detection for AI-Generated Images.
• 

  Learn-to-Distance: Distance Learning for Detecting LLM-Generated Text (ICLR 2026)

  Zhou, H^*., Zhu, J^*., Xu, E., Ye, K., Yang, Y. and Shi, C.

  Learn-to-Distance.
• 

  Doubly Robust Alignment for Large Language Models (NeurIPS 2025)

  Xu, E^*., Ye, K^*., Zhou, H^*., Zhu, L., Quinzan, F. and Shi, C.

  DRPO4LLM.
• 

  AdaDetectGPT: Adaptive Detection of LLM-Generated Text with Statistical Guarantees (NeurIPS 2025)

  Zhou, H^*., Zhu, J^*., Su, P., Ye, K., Yang, Y., Gavioli-Akilagun, S. and Shi, C.

  AdaDetectGPT.