GPT-6 Astra discussions yield optimal regret bound for online inverse linear optimization

Shinsaku Sakaue @shinsaku_sakaueOriginally posted 2026.09.24

Open original on X

After extensive discussions with GPT-6 Astra, I've closed the regret gap in online inverse linear optimization! For utility and actions in the d-dimensional Euclidean unit ball, the optimal regret is Θ(√d). The algorithm is far from efficient, though. 🔗 https://arxiv.org/abs/2609.26978

This post quotes another post: x.com/shinsaku_sakaue/status/2100365620999245842