Lingming Zhang
University of Illinois Urbana-Champaign, IL·Computer Science
Software bugs affect millions of people and cost billions of dollars. A leading way to fix them automatically uses agents built on large language models (LLMs), AI systems trained on huge amounts of text and code.
- Fields
Computer Science
Engineering
- Topics
software engineering
large language models
AI agents
- Grants
- 1 known grant, through Sep 2029
What they work on
- Train LLMs jointly to generate synthetic software issues and to resolve them, using a self-play loop in which the model learns to create the issues that most improve its own bug-fixing ability.
- Build live software agents that reason about and modify their own scaffold, the code that directs how they use tools, at runtime to maximize bug resolution.
- Enable agents to reuse and build on their earlier action trajectories so their performance on software tasks improves progressively over time.
More about the research
These agents can call external tools and carry out a chain of actions, called a trajectory, to finish a software task such as fixing a reported issue. They score well on benchmarks. They are held back, however, by weak training data, ad hoc design, and poorly guided exploration. This project aims to let such agents keep improving themselves on their own.
Why it matters
More capable self-improving coding agents could help developers worldwide find and fix real software bugs automatically. That could reduce the costs and harms that bugs cause. Better coding and reasoning in LLMs could also benefit other downstream applications that rely on these models.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Antonio Mastropaolo
William & Mary, VA·Computer Science
Test whether filtering low-quality training examples with task-specific software engineering metrics lowers the computational cost of training large language models for code while preserving their learning ability.
Hari Balakrishnan
Massachusetts Institute of Technology, MA·Computer Science
Test whether AI agents built on large language models can discover better decision rules, or heuristics, for networked systems in settings where goals are measurable, using concrete case studies and benchmark evaluation.
Milos Gligoric
University of Texas at Austin, TX·Computer Science
Develop a language and framework for writing fine-grained tests that check individual statements and code fragments inside functions, and evaluate it on open-source projects.
The school
University of Illinois Urbana-Champaign: 41 labs on PI Pages, $907M a year on research, 870 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.