FlowOptimizer: Learning to Optimize via Unfolded Flows

FlowOptimizer is a deep unfolded, flow-based learning-to-optimize framework developed by researchers at MIT and Boston University, accepted at ICML 2026. Instead of updating a single point, it represents each optimization step as a learned velocity field acting on a whole population of candidate solutions, conditioned on contextual signals such as objective values, gradients, and population-level statistics, using a permutation-invariant self-attention architecture.

Training proceeds in two phases: an initial simulation-free flow-matching stage learns displacements from starting populations toward improved target populations obtained via sampling, followed by unfolding the model over K iterations and fine-tuning it end-to-end by directly minimizing a weighted sum of per-iteration objective losses. Notably, the whole pipeline is self-supervised, relying purely on objective evaluations without ever needing known solutions.

Across standard non-convex benchmarks and real-world tasks in robotics, power grids, and supply chains, FlowOptimizer beats gradient-based, sampling-based (CEM, CMA-ES, PSO) and prior learning-to-optimize baselines by orders of magnitude in solution quality. It also generalizes from low-dimensional training problems to roughly 10x higher-dimensional ones, and can be understood as a generalization of classical methods like PSO, CEM, and evolutionary strategies under specific configurations.