About the session
Combinatorial optimization problems arise in domains such as viral marketing, computational sustainability, and healthcare. Their NP-hard nature makes exact algorithms impractical, leading to reliance on handcrafted heuristics or neural combinatorial solvers. Heuristics require significant expert effort and often remain sub-optimal, while neural solvers depend on NP-hard ground-truth data, generalize poorly out of distribution, and require GPU-based inference, raising scalability and sustainability concerns.
We reframe combinatorial optimization as a problem of autonomous algorithm discovery. Instead of predicting solutions, we design an AI agent that synthesizes executable heuristics directly. Leveraging large language models for code generation and evolutionary search for refinement, the agent proposes, evaluates, and iteratively improves candidate algorithms as explicit programs. Crucially, this process does not rely on optimal supervision, but improves solution quality through objective-driven feedback.
The result is transparent, human-readable code that runs efficiently on CPUs, eliminating neural inference costs. By discovering algorithmic structure rather than fitting parameters, the approach promotes interpretability, cross-domain robustness, and computational sustainability, offering a scalable pathway toward self-improving combinatorial optimization solvers.
