The work tests whether general-purpose global nonlinear optimization can solve difficult combinatorial geometry problems that have recently attracted attention through LLM-driven algorithm discovery, including DeepMind’s AlphaEvolve. Rather than developing specialized algorithms or altering solver internals, the researchers translate a selection of problems from the AlphaEvolve benchmark suite into direct nonlinear programming formulations. They then run these formulations using two established global optimization solvers: commercial FICO Xpress and open-source SCIP.
Both solvers recover previously reported best-known solutions, and in several cases find better ones, including results that surpass recent LLM-driven discoveries. The findings show how much generic nonlinear optimization has advanced: problems that would have challenged general-purpose solvers roughly a decade ago can now be handled with standard solver technology. The study also suggests that LLMs and optimization solvers need not be competing approaches to discovery. Global solvers can serve as powerful tools within LLM-assisted workflows, helping search for, verify, or improve candidate solutions. By revisiting these benchmarks, the researchers make the case that straightforward mathematical formulations and mature optimization software deserve a place alongside specialized algorithm-generation methods.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.