I'm a 5th-year Ph.D. candidate at Institute for AI Industry Research (AIR), Tsinghua University, advised by Prof. Yunxin Liu, working closely with Prof. Ting Cao. I received my B.E. in Electronic Engineering, also from Tsinghua University in 2022. I was at EESAST and NICS-EFC as an undergraduate.
I work on self-evolving physical AI (or, PSI, physical self-improvement). I see my previous systems work as a foundation to accelerate the evolution loop. I hope to bridge algorithm breakthroughs with the underlying infrastructures.
Homepage · Google Scholar · Zhihu 知乎 · Rednote 小红书 (looking forward to 1k followers 😆)
First / co-first:
- OxyGen (NeurIPS 2026): unified KV Cache management for VLA inference under multi-task parallelism
- Cosmos Lite: efficient inference and parallel rollout for Cosmos 3 robot policies on consumer-grade GPUs
- ActProbe: a lightweight action-space probe for early failure detection of robot policies
Collaborations:
- Zetta ζ (1k+ stars🔥): an efficient closed-loop embodied harness for self-evolving physical intelligence
- Zeva: in-context causal learning (ICCL) for generalizable embodied manipulation
- EmbodiSkill (integrated into Microsoft/SkillOpt): a training-free skill-aware framework for self-evolving embodied agents
- Embodied.cpp (100+ stars): a portable inference runtime for embodied AI models on heterogeneous robots
First / co-first:
- Vec-LUT (MobiSys 2026 Best Paper Award Runner-Up🏅): vector table lookup for parallel ultra-low-bit LLM inference
- FlexNN (MobiCom 2024): an adaptive memory management framework for DNN inference under memory constraints
(back to the days when human had to handwrite 10k LoC C/C++ for one paper)
Collaborations:
- Personal LLM Agents (500+ citations🔥): a survey and position paper, where I led the Efficiency section
- Time's Up (EMNLP 2025): an empirical study of LLM reasoning ability under strict output length constraint