Post-Training and Reinforcement Learning in LLMs
Post-Training and Reinforcement Learning in LLMs
I work on post-training in LLMs, including supervised fine-tuning, RL and OPD methods.
I work on post-training in LLMs, including supervised fine-tuning, RL and OPD methods.
Instruct, not Answer: Using Instructions as Privileges in On-Policy Context Distillation [arXiv]
Hantao Yu, Sandy Han, Udaya Ghai, Ferhat Erata, Joe Lilien, Aman Goel, Ali Torkamani
In Submission
I work on understanding the tradeoffs among expressivity, efficiency and parallelism in LLM architectures.
On the Computational Hardness of Transformers [arXiv]
Barna Saha, Yinzhan Xu, Christopher Ye, Hantao Yu
Symposium on Theory of Computing (STOC), 2026
Two Heads are Better than One: Simulating Large Transformers with Small Ones [arXiv]
Hantao Yu, Josh Alman
Spotlight at Neural Information processing Systems (NeurIPS), 2025
Fast Attention Mechanisms: a Tale of Parallelism [arXiv]
Jingwen Liu, Hantao Yu, Clayton Sanford, Alexandr Andoni, Daniel Hsu
Neural Information processing Systems (NeurIPS), 2025
Fundamental Limitations on Subquadratic Alternatives to Transformers* [arXiv]
Josh Alman, Hantao Yu
International Conference on Learning Representations (ICLR), 2025
Theoretical Computer Science
I work on algorithms for tensor computation and learning theory. I have worked on making fast matrix multiplication algorithms more practical, and understanding the hardness of dynamic programming given different recursive relationships.
Alexandr Andoni, Justin Dallant, László Kozma, Hantao Yu
In Submission
Improving the Leading Constant of Matrix Multiplication* [arXiv]
Josh Alman, Hantao Yu
Symposium on Discrete Algorithms (SODA), 2025
Tensor Ranks and the Fine-Grained Complexity of Dynamic Programming* [arXiv]
Josh Alman, Ethan Turok, Hantao Yu, Hengzhi Zhang
Innovations in Theoretical Computer Science (ITCS), 2024
Robust Empirical Risk Minimization with Tolerance [arXiv]
Robi Bhattacharjee, Max Hopkins, Akash Kumar, Hantao Yu, Kamalika Chaudhuri
International Conference on Algorithmic Learning Theory (ALT), 2023
Active Learning Polynomial Threshold Functions* [arXiv]
Omri Ben-Eliezer, Max Hopkins, Chutong Yang, Hantao Yu
Neural Information processing Systems (NeurIPS), 2022