arXiv paper evaluates LLMs as numerical optimizers (arXiv:2610.37932v1)
This new arXiv paper (arXiv:2610.37932v1) develops an evaluation framework to study large language models' ability to produce feasible, near‑optimal solutions for mathematical optimization problems (convex optimization, fractional programming, constrained projection). It finds LLMs can capture useful structure and sometimes generate near‑optimal outputs, but reliability falls with problem dimension and numerical/formatting errors are common; iterative refinement improves validity but persistent numerical and structural failures limit use as standalone optimizers.