UCLA's AutoOPT Automates Optimization Research with GPT-5.6 and Discovers New Algorithms

UCLA optimization researcher Ernest Ryu's team (work led by Shuvo) introduced AutoOPT, a domain-specific framework built around GPT-5.6 Sol that automates first-order optimization research end to end and has discovered two new algorithms. Ryu estimates that roughly 90% of algorithm design and convergence-proof work in optimization will be automated by AI, and urges researchers to embrace these tools.

Confirmed

• AutoOPT splits the discovery of optimal first-order methods into four stages: numerical design via the BnB-PEP methodology; symbolic discovery of analytic descriptions and convergence proofs using frontier LLMs; formal verification in the Lean 4 proof assistant, among others (m1, truncated at the formalization stage).
• The system discovered two new algorithms: the Lemniscate Accelerated Gradient Method and Analytic ITEM-f, with the lemniscate constant appearing in a convergence rate for the first time (m2).
• Ryu admitted these constructions are "frightening" and that he could not have come up with them himself without the LLM and the domain-specific harness (m6, m7).
• He predicts that designing efficient algorithms and proving convergence guarantees — crafts he spent years honing — are already and will continue to be about 90% automated, and that embracing these tools could make optimization research roughly 10x faster (m3, m4, m7).
• Paper details are covered in Shuvo's thread, while Ryu's long-form posts focus on methodology and reflections on how research will change (m5).

Why it matters

• This is a concrete case of LLM-driven automated mathematical research in optimization theory: not just writing assistance, but end-to-end production of new algorithms and convergence proofs with formal verification.
• The lemniscate constant appearing in a convergence rate for the first time suggests automated pipelines can uncover mathematical structures human researchers had not reached.
• Ryu's "90% automation" judgment comes from a frontline researcher's own reflection, situated in the broader history of science automating parts of research workflows while researchers move to higher-level tasks — offering direct career and methodological guidance for peers.

2026-08-19 ~ 2026-08-19 · 7 related posts

• [source] AutoOPT: Automating Optimization Research with GPT-5.6 Sol — ErnestRyu · 2026-08-19
• [source] Researchers built a domain-specific harness around GPT-5.6 Sol to automate optimization research — ErnestRyu · 2026-08-19
• Optimization researcher: ~90% of algorithm design and convergence proofs will be automated — ErnestRyu · 2026-08-19
• Algorithm Design to Be 90% Automated by AI — ErnestRyu · 2026-08-19
• [source] AutoOPT discovers new optimization algorithms, introducing the lemniscate constant — ErnestRyu · 2026-08-19
• Optimization researcher: I couldn't have derived these new constructions without LLMs — ErnestRyu · 2026-08-19
• Optimization researcher: with LLMs and domain harnesses, we go 10x faster on harder problems — ErnestRyu · 2026-08-19