My research connects AI reasoning with optimization: how agents formulate problems, evaluate intermediate decisions, search for better solutions, and reuse experience. I develop process reward models and memory mechanisms for agents, methods for automated algorithm discovery, and systems that connect language to executable optimization models.
A recurring concern is how to tell whether an apparent improvement is real. I study this through reasoning benchmarks, evaluation of generated models, and experiments with executable algorithms and solvers. Depending on the problem, the evidence comes from feasibility checks, objective values, mathematical bounds, expert judgments, or performance on held-out tasks. These signals answer different questions, so I design the evaluation around the claim being tested.
This agenda builds on my work in exact and heuristic optimization and my experience leading industrial AI research. It connects new methods with the constraints of systems that must work repeatedly in practice. Work under anonymous review is described by topic until a title can be made public.
Improve multi-agent reasoning through process feedback, memory, and inference-time search.
Generate algorithms and learn how to navigate difficult search spaces.
Translate specifications into optimization models, test technical reasoning, and build tools for decision support.
Design exact, heuristic, and accelerated methods for constrained decisions.