EMNLP2026 · Main
Anhao Zhao.
Ph.D. Student at PolyU
Research interests:
I am Anhao Zhao, a joint Ph.D. student at the NLP Group of The Hong Kong Polytechnic University & EIT NLP of Eastern Institute of Technology, Ningbo, fortunately supervised by Dr. Xiaoyu Shen and Prof. Wenjie Li.
Research
My research focuses on LLM post-training and LLM efficiency, with the long-term goal of building general-purpose models that advance reasoning, interaction, and inference efficiency.
LLM post-training
I study how supervised fine-tuning (SFT), reinforcement learning (RL), and on-policy distillation (OPD) can develop stronger reasoning and agentic capability. My work explores SFT on self-generated data for efficient reasoning [On-Policy SFT], RL for adaptive streaming reasoning [AdaSR], more stable and effective OPD [PowerOPD, KAT], and token-efficient supervision through selective distillation [DIAL-OPD].
LLM efficiency
I explore efficiency at both the token and architecture levels. At the token level, I study alternatives to explicit reasoning tokens [Latent CoT Survey] and reasoning while reading to reduce response latency [StreamingThinker]. At the architecture level, I investigate token-adaptive computation [SkipGPT], the practical inference speedups of model pruning [Beyond FLOPs], and visual cache reuse and sparse interactions for efficient multimodal inference [miniReranker].
📬I am open to collaborations and discussions. Please feel free to reach out to me if you are interested in my research or any relevant topics.
News
Gave a talk on Multi-Teacher and Cross-Tokenizer On-Policy Distillation at DINQ Session 1 in Beijing.
Five papers accepted to EMNLP 2026: 3 Main Conference papers and 2 Findings papers 🎉!
SkipOPU accepted to ACM Transactions on Reconfigurable Technology and Systems (TRETS) 🎉!
Got one paper accepted by ICML 2026🎉!
Got one paper accepted by CVPR 2026 Spotlight🎉!
Got one paper accepted by ICLR 2026🎉!
Earlier news
Attended EMNLP 2025 in person for the first time — a truly exciting experience 🎉
Started my Ph.D. study at the NLP Group @ PolyU & EIT NLP, supervised by Dr. Xiaoyu Shen and Prof. Wenjie Li.
Got one paper accepted by EMNLP 2025🎉!
Released our new survey on Latent Chain-of-Thought Reasoning.
Got one paper accepted by ACL 2025🎉!
Got one paper accepted by ICML 2025🎉!
Got one paper accepted by EMNLP 2024🎉!
Publications
Most recent publications on Google Scholar.
* indicates equal contribution
Conference & Journal Papers
EMNLP2026 · Main · Oral
EMNLP2026 · Main · Oral
Escaping the KL Agreement Trap in On-Policy Distillation
Haoran Xin, Anhao Zhao, Ying Sun, Jin Li, Xiaoyu Shen, Hui Xiong
EMNLP2026 · Findings
Post-Training Shifts Confidence: A Three-Stage Analysis of How SFT, RL, and OPD Shape CoT Calibration
Shuhao Li, Guodong Du, Anhao Zhao, Wanyu Lin, Tianyu Yuan, Xiaoyu Shen
EMNLP2026 · Findings
ACM TRETS2026 · Accepted
SkipOPU: An FPGA-based Overlay Processor for Large Language Models with Dynamically Allocated Computation
Zicheng He*, Anhao Zhao*, Xiaoyu Shen, Chen Wu, Lei He
CVPR2026
ICLR2026
EMNLP2025 · Main
ACL2025 · Findings
ICML2025
EMNLP2024 · Main
KBS2024
ArXiv Preprints
arXiv2026
DIAL-OPD: Learning More from Fewer Tokens in On-Policy Distillation
Anhao Zhao, Haoran Xin, Junlong Tong, Yingqi Fan, Xuan Lu, Ping Nie, Wenjie Li, Xiaoyu Shen†
arXiv2026
AdaSR: Adaptive Streaming Reasoning with Hierarchical Relative Policy Optimization
Junlong Tong, Wenqi Xu, Yingqi Fan, Anhao Zhao, Xuan Lu, Yang Tan, Xiaoyu Shen
arXiv2026
miniReranker: Efficient Multimodal Reranking through Visual Cache Reuse and Interaction Sparsity
Yingqi Fan, Xuan Lu, Anhao Zhao, Junlong Tong, Ping Nie, Kai Zou, Yunpu Ma, Wei Zhang, Xiaoyu Shen
arXiv2026
ProactiveLLM: Learning Active Interaction for Streaming Large Language Models
Junlong Tong, Yao Zhang, Anhao Zhao, Yingqi Fan, Yunpu Ma, Xiaoyu Shen
arXiv2026
Decoupling KL and Trajectories: A Unified Perspective for SFT, DAgger, Offline RL, and OPD in LLM Distillation
Anhao Zhao, H Xin, Yingqi Fan, Junlong Tong, Wenjie Li, Xiaoyu Shen
arXiv2026
arXiv2026
arXiv2026
Talks
Multi-Teacher and Cross-Tokenizer On-Policy Distillation
DINQ Session 1 · Beijing, China ·
Service
Reviewer/Program Committee Member:
ICLR27, ICLR26, CVPR26, ECCV26, ICML26, NeurIPS26
Teaching Assistant:
COMP1010_26271_C: Computational Thinking and Problem Solving
COMP 5311: Internet Infrastructure and Protocols, Fall 2025, PolyU
COMP 5532: DIGITAL TWINS & VIRTUAL HUMAN, Spring 2026, PolyU