Applied Scientist@Amazon

I am an applied scientist in Kiro Science, an agentic coding IDE. Previous to that, I work in Rufus LLM Modeling Team, Amazon’s agentic AI shopping assistant. My working topic involves LLM Post-training, LLM Agent for Coding & Personalize Recommendation, Verifiable Coding Environments, Reward Hacking Detection, Large-scale Codebase Understanding and Planning.

My PhD research centers on graph ML and pragmatic reasoning with LLMs (LLM better capture but may not understand both human language and program language), complemented by industry internship experience in generative recommendation and search systems.

I believe that exploiting better underlying structure of data can help LLMs achieve better scaling behavior — for example, program dependency graphs in codebases for coding agents, discourse structure in long-context conversations, and webpage/document connections for deep research agents.

Awards

  • WSDM 2024 Best Paper Honor Mentioned Award (First Author) (3/615)
  • CIKM 2021 Best Short Paper Award (First Author) (1/626)
  • ICML 2024 Spotlight (First Author) (335/9473)
  • EMNLP 2024 best paper recommendation (Second Author)
  • NeurIPS 2024 Top Reviewer Award (Both Main track & DB Track) (0.64%)

Before transitioning to industry role, I spent five years researching the fundamental modeling capabilities on networks. During this journey, three moments stand out as sources of profound inner fulfillment.

  • Pragmatic Reasoning with LLMs LLM better capture but not understand human language, it is interesting to find potential way for LLMs, to **understand the implied meaning of human language**.

  • Pioneering Graph Foundation Models (GFMs) [Research Statement] [Talk Slides]
    GFM is a naive yet intriguing research topic that intersects network science, geometry, and neural-symbolic reasoning, providing a complementary perspective to the existing foundation model paradigm. GFMs may hold future potential for applications in code graphs, scene graphs, and higher-level intelligence, such as analogical reasoning.

  • Serving as the organizer of the [Learning on Graph Conference 2024/25]
    LoG is an exciting new conference facilitating graph and geometry community engagement. Being part of this initiative is exciting.