Yin Lin (林茵)
R&D Engineer · Alibaba Token Foundry
Research Interest: Foundation Model Systems & Data Management
R&D Engineer · Alibaba Token Foundry
Research Interest: Foundation Model Systems & Data Management
Hi, my name is Yin Lin. I am a senior algorithm engineer at Alibaba Token Foundry. My research focuses on making foundation models adaptable to diverse domains by building systems that can acquire domain knowledge, learn from experience, and turn model capabilities into useful work.
This includes research on agent harnesses—how agents manage context, use tools, orchestrate subagents, verify their work, and keep learning—and AI-powered data management, exploring how foundation models can support data processing, analytics, and reasoning across domains.
Before joining Alibaba, I earned my Ph.D. from the University of Michigan, Ann Arbor, where I was advised by Dr. H. V. Jagadish. My thesis research focused on data equity systems.
If you are interested in collaboration, feel free to reach out!
Ph.D., Computer Science and Engineering
B.S., Computer Science
Senior Algorithm Engineer, Alibaba Token Foundry
Research Intern, Data Management, Exploration and Mining (DMX)
Summer Intern, Software Architecture Group
How can foundation model systems keep learning, and how can the tokens they generate contribute to reliable task completion and lasting improvements? I explore these questions through agent learning, harness research, and AI-powered data management.
Recursa learns from domain materials and task experience, turning them into reusable instructions, skills, and memory. Parallel learners propose improvements that are validated and merged into a shared knowledge repository. With the underlying model and harness held fixed, the system adapts by improving the knowledge it brings to future tasks.
I study how agent harnesses turn model capabilities and inference compute into reliable task execution. This includes programmatic context management with Scroll to retrieve and reuse interaction history, verification to check and revise agents' work, and multi-agent orchestration to decompose tasks, coordinate agents, and combine their results.
Natural language queries to databases are often ambiguous. AmbiSQL is an interactive system that detects and resolves ambiguities in text-to-SQL translation, enabling users to clarify their intent and receive accurate query results.
Semantic operators bring AI-powered data transformations to large-scale data processing pipelines. These operators are now part of Data-Juicer, enabling intelligent filtering, enrichment, and transformation of training data for foundation models.
DojoZero is an agent arena where AI agents react to real-time data streams to make predictions for sports events. It serves as a testbed for evaluating agents' capabilities in dynamic, time-sensitive decision-making scenarios.
Exploring novel agent applications for data analytics, including using LLMs for feature engineering and data engineering tasks.
arXiv 2026
arXiv 2026
VLDB 2026 (Tutorial)
SIGMOD 2026 (Demo)
IEEE Data Engineering Bulletin 2025
CoRR, 2024, Arxiv/2412.16864
ICDE 2024
CIDR 2024
SIGMOD 2023 (Best Paper Award)
ACM Computing Surveys
Highly cited survey in the field of AI fairnessVLDB 2022 (Demo)
VLDB 2022
VLDB 2020
CoRR, 2020, Arxiv/2010.08807
MLG@KDD 2020
DEXA 2018
Program Committees / Reviewers: ACL, NeurIPS, TKDE, CIKM, IEEE BigData, AIBSD (AAAI Workshop on AI with Biased or Scarce Data), ReLM (AAAI Workshop on Responsible Language Models)