GPT-5.6 closes 30‑year convex‑optimization gap with a single prompt
A targeted prompt guided GPT‑5.6 to improve the convergence bound in a classic convex‑optimization problem from 1/L to 1.5/L, finally closing a three‑decade‑old open gap. The proof was verified by experts, showing large language models can now generate novel, peer‑validated mathematical results.