Alejandro Ribeiro
University of Pennsylvania, PA·Engineering
Generative artificial intelligence systems can write text, draw images and design molecules, but they produce whatever their training data taught them, including biases, unsafe content and plain falsehoods. The common problem across both projects here is control: how to steer a large pretrained generative model so its output satisfies requirements a user cares about, and how to do so with mathematical guarantees rather than trial and error.
- Fields
Engineering
Computer Science
Mathematics & Statistics
- Topics
generative AI
alignment
large language models
- Grants
- 2 known grants, through Sep 2028
What they work on
- Develop optimization and statistical foundations for aligning generative models by exploiting that the problem is strongly concave in distribution space even though it is highly nonconvex in the parameters of deep networks and transformers.
- Formulate alignment as a reinforcement learning problem with a known value function, and show how this simplification yields tractable algorithms for enforcing fairness, safety, robustness and truthfulness constraints.
- Build a unified treatment of alignment for large language models and generative diffusion models, both posed as matching a generative model to user requirements given a pretrained reference.
- Construct discrete diffusion processes that generate graphs by progressively adding or removing edges, reducing the combinatorial cost of graph-space search compared with the Gaussian diffusion used for images and audio.
- Train generative graph models at small scale and transfer them to much larger graphs using graphon and manifold abstractions and corresponding graphon and manifold neural network architectures.
- Incorporate user-specified structural constraints into graph generation and apply the resulting models to molecular graphs that mimic molecules with known properties for AI-aided drug discovery.
More about the research
One project builds the optimization theory of this steering problem for language and image models; the other extends generative modeling to graphs, where the requirements are structural constraints on the relationships being generated.
Why it matters
Generative AI tools are being built into everyday software while still reproducing biases, unsafe outputs and false content. A rigorous theory of alignment could give developers principled ways to enforce fairness, safety and truthfulness requirements when adapting public pretrained models. Scalable constrained graph generation could speed AI-aided drug discovery by proposing molecules with target properties, with likely effects in communications, robotics, circuit design and health.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Yusu Wang
University of California-San Diego, CA·Computer Science
Explain why trained flow-based and diffusion generative models generalize better than the theoretically optimal solution predicts, using geometry, ODE analysis, manifold learning and deep learning theory, to enable controlled generation of novel samples.
Zaid Harchaoui
University of Washington, WA·Computer Science
Identify the mathematical quantities that drive generalization in frontier generative models, including large language models, using frameworks from probability theory, variational analysis, mathematical statistics and information theory.
Faez Ahmed
Massachusetts Institute of Technology, MA·Engineering·Early-career
Develop diffusion models, generative AI that builds outputs by removing noise step by step, that learn from the full history of design optimization runs rather than only final optimized designs, to generate more precise engineering designs.
The school
University of Pennsylvania: 90 labs on PI Pages, $2.2B a year on research, 750 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.