PhD · Pattern Recognition & Intelligent Systems
Huanxuan Liao 廖桓萱
I am a PhD student at the
University of Chinese Academy of Sciences (UCAS)
and the Institute of Automation, Chinese Academy of Sciences
(CASIA),
advised by Prof. Shizhu He and
Prof. Kang Liu.
Previously, I earned my B.Eng. from
North China Electric Power University (NCEPU) (2023),
working with
Prof. Min Shi and Prof. Yun Ju.
Better data & architectures for stronger models and longer horizons.
Research Statement
I study foundation model pretraining and mid-training. I examine how data composition, model architecture, and training schedules shape capabilities across stages. A central question is which training choices remain effective across a scaling ladder of model sizes and token budgets.
- Data and Model Design. I use long-context training to study data mixtures and training schedules. My modeling work covers efficient attention, context and KV-cache compression, and adaptive visual input allocation (LCLM-Horizon, HyCo2, SparK, ResAdapt, AllocEmbed).
Adaptation
- Continual Adaptation. Knowledge transfer, skill retention, and generalization across tasks (TAGI, SHIP, DyPRAG, AAG, NesyCD, SKIntern, SpaRTA).
- Agent Learning. Memory, decision-making, and policy improvement through interaction and feedback (CITI, KATE, BuPO, Tool-RL-Box, OPERATE, Distill Yourself, SLE).
Education
PhD in Pattern Recognition and Intelligent Systems
University of Chinese Academy of
Sciences (UCAS)
2023 - Present
Institute of Automation, Chinese Academy
of Sciences (CASIA)
2023 - Present
B.Eng. in Intelligence Science and Technology
North China Electric Power
University (NCEPU)
2019 - 2023
Graduated with honors, Outstanding Graduate
News
-
Award Selected for the 2026 CAST Young Science and Technology Talent Cultivation Program (Doctoral Student Special Plan).
Paper One paper accepted to ACM TOIS'26. HyCo2 (first author).
-
Paper One paper accepted to EMNLP'26 (Main). BuPO (co-author).
-
Paper One paper accepted to the FAGEN Workshop at ICML'26. Multi-step tool-use RL study (co-author).
-
Paper One paper accepted to AAAI'26. SparK (first author).
-
Award Received the National Scholarship from the Ministry of Education at CASIA.
Earlier news
-
Paper One paper accepted to ACL'25 (Main). ETAPP (co-author).
-
Survey A Comprehensive Survey on Long Context Language Modeling released on arXiv (co-author and core contributor).
-
Paper One paper accepted to NeurIPS'24. TAGI (first author).
-
Paper One paper accepted to CCKS'23. Dynamic Weighted Neural Bellman-Ford Network (first author).
Award Received the Beijing Outstanding Graduate honor.
Publications
View on Google Scholar (* stands for equal contribution)
Loading publications...
Data and Model Design
Training data, evaluation, model efficiency, and input allocation.
Continual Adaptation
Knowledge transfer, skill retention, and cross-task generalization.
Agent Learning
Memory, tool use, and learning from interaction and feedback.
No publications match your filters. Try another research area or search term.
Awards & Honors
Scholarships & Honors
-
Young Science and Technology Talent Cultivation Program (Doctoral Student Special Plan)
China Association for Science and Technology (CAST), 2026
Selected for 中国科协青年科技人才培育工程博士生专项计划.
-
National Scholarship (国家奖学金)
Ministry of Education
CASIA (2025) - NCEPU (2021-2022, 2 consecutive years). Top scholarship in China; 0.2% domestically
-
Climbing Second Prize Scholarship
Institute of Automation, CAS, 2024
Top scholarship in CASIA
-
Beijing Outstanding Graduate Awards
Beijing Ministry of Education, 2023
Top graduate in Beijing
Competitions
-
National Third Prize
Information Security Competition, 2022
-
National Excellence
College Student Innovation and Entrepreneurship Project, 2021
Academic Services
Conference Reviewing
- ACL ARR 2024–2026
- NeurIPS 2025–2026
- ICML 2026
- COLM 2026
- ICLR 2025–2026
- AAAI 2026
Community Contributions
- Awesome-LLM-Long-Context-Modeling > 2.2k
- Distill Yourself Homepage
- LCLM-Horizon Survey Core Contributor
- NVIDIA/kvpress Contributor