Ravi Mangal
Colorado State University-Fort Collins, CO·Computer Science
Large language models can now write source code from a description, but the code they produce often contains subtle bugs that look plausible and pass casual inspection. In an industrial software setting that is a security problem: quietly wrong code can introduce vulnerabilities.
- Fields
Computer Science
Engineering
- Topics
large language models
code generation
program analysis
- Grants
- 1 known grant, through Sep 2028
What they work on
- Design quantitative metrics for evaluating code generated by large language models using symbolic execution, which explores program paths with symbolic inputs, and Bayesian program analyses that reason probabilistically about program behavior.
- Train a differentiable reward model from these program scores that can assess the quality of partial as well as complete generated code, making semantic feedback usable by gradient-based learning.
- Fine-tune code-generating language models with the semantic reward model, following the reinforcement learning with human feedback (RLHF) recipe but substituting program-analysis scores for human ratings.
- Improve decode-time generation by using the reward model together with similarity-based program ranking to constrain and prune the decoding tree of candidate code continuations.
- Develop semantics-guided evaluation metrics and collect new benchmarks of realistic coding tasks for training and evaluating code language models.
More about the research
A diagnosis of why this happens is that these models are trained, used, and graded on how code looks rather than on what it does, that is, on program semantics. This collaborative project aims to improve the quality of model-generated code by wiring program analysis, the automated study of what a program actually computes, into every stage of the model's life cycle: training, generation, and evaluation.
Why it matters
Code produced by language models is already being used in software development, and subtly erroneous code can expose security vulnerabilities in deployed systems. Scoring generated code by what it actually computes, and using those scores during both training and generation, could reduce such bugs and give the field benchmarks and metrics that reflect realistic coding tasks rather than surface similarity.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Swarat Chaudhuri
University of Texas at Austin, TX·Computer Science
Develop foundation model programming, which writes scientific hypotheses as programs combining symbolic rules with neural modules and foundation models, to extract interpretable concepts from images and video.
Cong Ma
University of Chicago, IL·Mathematics & Statistics·Early-career
Develop robust ranking algorithms for pairwise comparison data that tolerate ill-conditioned sampling of which pairs are compared and relax parametric modeling assumptions.
Narges Norouzi
University of California-Berkeley, CA·Education Research
Model students' problem-solving strategies and algorithmic thinking in computer science courses by integrating fine-grained process data to trace competencies across levels of abstraction.
The school
Colorado State University-Fort Collins: 23 labs on PI Pages, $466M a year on research, 276 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.