Automating Heuristic Design with MIT CSAIL Graduate Student Pantea Karimi
Audrey Woods, MIT CSAIL Alliances | February 17, 2026
Behind modern computing systems lies a web of heuristics—fast decision rules that govern video transmission, network behavior, and infrastructure performance. Heuristics act as shortcuts, enabling large platforms like ChatGPT and Zoom to serve millions of customers efficiently by making quick decisions based on connectivity, storage, or user needs. When watching Netflix, heuristics determine the video quality to send to a customer based on their internet speed. In cloud load balancing, a heuristic routes incoming requests to the least busy server. These algorithms are powerful but tricky and imperfect. A badly designed heuristic can lead to enormous losses in a billion-dollar system, but improving them is intensive, difficult, and resource-heavy, especially at scale.
MIT CSAIL PhD student Pantea Karimi is working to change that. Her research aims to systematically analyze, explain, and automate heuristic design, transforming a process that's historically relied on expert trial-and-error into one that's automated, robust, accessible, scalable, and efficient.
SOLVING THE PROBLEMS THAT BOTHER HER
Karimi started studying at MIT during the COVID pandemic in her home country of Iran, where a poor internet connection made video calls with her PhD advisor, MIT CSAIL Associate Professor Mohammad Alizadeh, a frustrating experience. “I was talking with my advisor about the projects I could work on, and the internet was really choppy. He asked me, ‘What do you want to work on? What are your interests?’ And I said: ‘You know what? Maybe we should fix this problem that we're having right now.’” This led her to develop a chain of video compression analytics and congestion control, all of which required heuristics to operate. But validating those heuristics was hard. "At the back of my mind, there was always this doubt: What if I'm doing it wrong? What can I do better? How can I be sure that we are at the limit of what we can do?” That doubt excited her for her new research direction: tools for analyzing heuristics.
FINDING WHERE AND WHY HEURISTICS FAIL
During her time at Microsoft Research, Karimi helped build XPlain, an analysis tool that evaluates heuristic algorithms and identifies weaknesses. "Basically, [XPlain] takes a heuristic in some domain-specific language format and analyzes the behavior of the heuristic,” alerting the user to edge cases where the algorithm might underperform. XPlain was a success, but required expertise to encode the heuristic in the domain-specific language to be useful. Building on this work at Microsoft Research, Karimi began working on MetaEase, a tool that analyzes heuristics directly from code, offering developers a more accessible way to analyze, optimize, and iterate over heuristic algorithms. MetaEase was recently accepted to a top systems conference and is now available as open source.
AUTOMATING HEURISTIC DESIGN WITH AI
At this point in her research, Large Language Models had emerged, raising the question of how much of this process could be fully automated. “There was this explosion of LLM research,” she says, but "LLMs on their own have a very hard time understanding what's happening [with heuristics]. They don't directly have access to the optimization tools, so it's hard for them to know what went wrong just based on a few examples and why exactly." Karimi and her fellow collaborators at Microsoft found that by empowering LLMs with the heuristic tools they’d developed, LLMs could be trained to design more robust heuristic algorithms. This research enabled developers to leverage LLMs' capabilities to navigate real-world trade-offs—latency, cost, and fairness—while systematically improving worst-case behavior.
But LLMs weren’t yet acting like independent collaborators. She and her colleagues wanted a human-like AI system that follows the same steps a researcher would: observe performance metrics, analyze what causes issues, propose a hypothesis, test a change in simulation, and iterate until it improves. Working in a group led by Professor Alizadeh and MIT Professor Hari Balakrishnan (Karimi’s secondary advisor), Karimi helped develop Glia, a “human-inspired AI for automated systems design and optimization.” Utilizing LLMs in a multi-agent, human-inspired workflow, Glia incorporates “multiple entities. You have a supervisor, you have a researcher, you have a note-taker that you can query and retrieve information." Instead of single-shot generation, Glia also iterates as researchers do. "Glia has sometimes found solutions better and faster than what human experts have found, and that is quite exciting."
Overall, Karimi’s research aspires to, as her PhD proposal states, “transform heuristic design from an intuition-driven craft into a systematic, tool-supported, and increasingly automated engineering discipline.” This will not only address issues like cost and performance, but also the enormous challenge of scalability. For large systems like Google, improving heuristics can be a gigantic and time-consuming process. Karimi says, “A key takeaway is that improvements found in smaller, more tractable settings can carry over to larger deployments—so teams can solve the easier problem first, and scale up with more robust solutions.”
LOOKING AHEAD
Karimi is excited to enter industry research when she graduates. Whether she joins a startup or an established tech company, her “next step would be to find a large-scale problem and implement my ideas to automatically improve heuristics there.”
Overall, her motivation remains personal. "I usually work on the stuff that annoys me. For the heuristic analysis and design, the fact that I was a PhD student and it took me almost two years to design a better heuristic for video congestion control really annoyed me. That was such a waste of time and resources for a very niche project." She envisions a future where this burden lifts. "A lot of computer science is heuristic design that could be offloaded to an AI research collaborator." That shift could free other computer scientists to focus on solving the real-world technical problems that bother them as well.
Learn more about Karimi on her website or CSAIL page.
