We released a revised preprint of Distribution-Aware Algorithm Design with LLM Agents, joint work by Saharsh Koganti, Priyadarsi Mishra, Pierfrancesco Beneventano, and Tomer Galanti.
The paper studies learning when the output is executable solver code rather than a predictor. In this setting, correctness is not enough: the learned solver should generalize in both solution quality and runtime on future instances drawn from the same deployment distribution.
The central abstraction is a solver hint: reusable distribution-specific structure inferred from samples and compiled into specialized solver code. Empirically, LLM code agents synthesize high-quality solvers across structured combinatorial optimization distributions while running much faster than broad generic baselines.
Read the preprint or Tomer Galanti's accompanying explainer, Distribution-Aware Programming.