From left to right: Zachary Ives, Zixuan Yi and Ryan Marcus. Photo by Sylvia Zhang
In any business, time is money. So it’s hardly surprising that many sectors have embraced systems powered by artificial intelligence (AI) and machine learning (ML): these tools promise to make time-consuming processes more efficient.
Zixuan Yi, a second-year doctoral student in Computer and Information Science (CIS) at the PRECISE (Penn Research In Embedded Computing and Integrated Systems Engineering) Center, is working to further improve AI and ML performance.
Yi’s research tackles a thorny problem in data management: query optimization, the task of quickly retrieving data relevant to a user — or AI agent — request.
“By bridging the gap between learning methods and real-world system constraints, I aim to create adaptive, automated solutions that continuously optimize performance and meet evolving user needs,” says Yi, who came to Penn Engineering in 2023 after graduating from Tsinghua University and relocating from Beijing.
Working with her advisors Ryan Marcus, Assistant Professor in CIS, and Zachary Ives, Adani President’s Distinguished Professor and Department Chair, Yi recently co-authored a paper introducing LimeQO, which optimizes multiple queries collectively rather than in isolation, unlocking substantial performance improvements across large workloads.
Yi will present the paper, the first to integrate both neural and linear methods in query optimization, at SIGMOD 2025, one of the world’s premier data management conferences.
“LimeQO represents a significant shift in learned query optimization,” says Marcus. “Yi’s work is especially interesting because it shows how a relatively old AI concept (i.e., low-rank matrix completion) can drastically decrease the amount of data required for training learned systems.”
A New Approach
Prior applications of machine learning to query optimization focused on individual queries, often struggling with high overhead costs and inconsistent performance improvements. Yi’s research takes a different approach by optimizing entire query workloads simultaneously.
“Our key insight is that queries within the same workload often share similarities,” says Yi. “Rather than optimizing them one at a time, we leverage these similarities using low-rank matrix completion, a technique borrowed from recommendation systems.”
The motivation for the paper comes from real-world industry needs. “Many large-scale database systems, such as those used by major tech companies like Meta and Microsoft, struggle with regression issues — where performance is not always guaranteed to improve — and high computational overhead,” explains Yi. By recognizing patterns in query workloads, LimeQO efficiently reduces these overhead challenges, making database optimization more scalable and adaptable.
One of Yi’s biggest contributions is viewing learned query optimization as a so-called “offline exploration problem.” Her work essentially moves costly exploration tasks from the critical “hot” path of production database servers to cheaper and more readily available general computing resources. This allows a database to execute learning tasks “in the background,” without interfering with regular operation.
“We want to guarantee that no query regresses or gets worse,” Yi says. “By framing the problem through this new lens, we created an approach that not only improves efficiency, but also ensures reliability in real-world deployment.”
In other words, LimeQO can actually accelerate customers’ database interactions. “When customers are running the workload queries again and again, they might notice that the queries are running faster and faster,” says Yi. “That’s because our workload-driven query optimization is doing the heavy lifting under the hood.”
Optimizing Databases Through Learning
One of Yi’s other projects focuses on history-based cardinality estimation, which refers to estimating the number of records a system will return before executing a query.
In collaboration with SystemsResearch@Google, she proposed a method where machine learning systems are trained in real time from previously executed queries, rather than having to go through a costly and time-consuming training process. This approach significantly lowers overhead, making the technique more viable for real-world database deployment.
Yi’s findings challenge two common misconceptions: first, that production queries are always unique — indeed, according to a recent paper from the Amazon Web Services Redshift team, 50% of database queries are repeated, and 80% vary only in constant numbers, meaning the differences between queries are trivial for computers to handle. This repetition can be exploited to continuously improve estimation accuracy without significant upfront costs.
Second, Yi’s work undercuts the widespread notion that machine learning models for databases must be large and complex in order to be accurate. As her projects demonstrate, lightweight, incrementally updated models can maintain efficiency while improving estimation accuracy. By addressing these challenges, Yi’s work paves the way for more adaptive, efficient and scalable query optimization.
The Road Ahead: Optimizing Language Agents
Yi, Ives and Marcus are currently working on a new project that focuses on language agents, AI systems where Large Language Models (LLMs) interact with user input and the environment to complete tasks. Their research aims to reduce errors made by these systems and increase efficiency when the systems are combined with external tools such as web searches, database queries and application programming interfaces (APIs).
Just as Yi found ways to simplify database queries, she and her advisors have been developing shortcuts for language agents to break down complex tasks. “We frame this problem as tree-planning, treating the agent’s workflow as a sequence of sub-goals organized into a tree structure,” says Yi. “Our aim is to optimize this structure so that the agent can plan and execute tool-usage steps more effectively.”
Today, many if not most AI agents use a “chain-of-thought” approach, continuously generating tokens in an unbroken stream, similar to a human’s internal monologue. In contrast, Yi’s approach enables AI agents to test multiple overlapping chains-of-thought at the same time.
This tactic strikes a balance between language models providing the right answers and not using too much energy. “Yi’s work on language agents is seeking a structured, scalable approach that improves both accuracy and efficiency,” says Ives. “This type of innovative thinking really contrasts with the more ad-hoc chain-of-thought processes used today.”
Moving forward, Yi is eager to explore real-world applications of language agents that interact reliably with external tools. “I’m excited to collaborate with industrial teams that deal with complex AI-driven products and advance the capabilities of autonomous agents,” she says.
Through advances like LimeQO, Yi continues to push the boundaries of modern data management, paving the way for smarter, more efficient database systems.
You can learn more about Yi’s research on her personal website. If you would like to collaborate with her on a project related to AI and database management, she can be reached by email.