THE BLACK–WHITE-BOX OPTIMIZATION NETWORK

ICLR 2026 Conference Submission25027 Authors

Keywords: Derivative-free optimization, Combinatorial optimization, Higher-order energy, HUBO, QUBO, Higher-Order Factorization Machines (HOFM), Ising seeding, Tensor-Train (TT)

TL;DR: In this paper, we introduce TTC - a derivative-free optimization framework that couples HOFM surrogates, Ising solvers, and Tensor-Train.

Abstract: We introduce a \textit{Black--White-Box Optimization Network} and its first instance, \textit{Tensor-Train Creator (TTC)}, which couples Ising-style solves, a factorization-machine surrogate, and tensor-train (PROTES) search. Typed couplings, lattice realignment, and warm starts cut oracle calls and time-to-target. On black-box benchmarks and Max-Cut, TTC attains better values under the same evaluation budgets.

Primary Area: optimization

Submission Number: 25027