[Paper Review] Flow-based Extremal Mathematical Structure Discovery

Gergely Bérczi, Baran Hashemi+1 more|arXiv (Cornell University)|Jan 25, 2026

Machine Learning in Materials Science0 citations

TL;DR

FlowBoost is a closed-loop, flow-based generative framework that learns to discover rare extremal geometric configurations by integrating conditional flow matching, geometry-aware sampling, and reward-guided fine-tuning, achieving competitive results with substantially less compute and no reliance on LLMs.

ABSTRACT

The discovery of extremal structures in mathematics requires navigating vast and nonconvex landscapes where analytical methods offer little guidance and brute-force search becomes intractable. We introduce FlowBoost, a closed-loop generative framework that learns to discover rare and extremal geometric structures by combining three components: (i) a geometry-aware conditional flow-matching model that learns to sample high-quality configurations, (ii) reward-guided policy optimization with action exploration that directly optimizes the generation process toward the objective while maintaining diversity, and (iii) stochastic local search for both training-data generation and final refinement. Unlike prior open-loop approaches, such as PatternBoost that retrains on filtered discrete samples, or AlphaEvolve which relies on frozen Large Language Models (LLMs) as evolutionary mutation operators, FlowBoost enforces geometric feasibility during sampling, and propagates reward signal directly into the generative model, closing the optimization loop and requiring much smaller training sets and shorter training times, and reducing the required outer-loop iterations by orders of magnitude, while eliminating dependence on LLMs. We demonstrate the framework on four geometric optimization problems: sphere packing in hypercubes, circle packing maximizing sum of radii, the Heilbronn triangle problem, and star discrepancy minimization. In several cases, FlowBoost discovers configurations that match or exceed the best known results. For circle packings, we improve the best known lower bounds, surpassing the LLM-based system AlphaEvolve while using substantially fewer computational resources.

Motivation & Objective

• Motivate and formalize the problem of discovering extremal geometric structures in high-dimensional configuration spaces.
• Introduce FlowBoost as a closed-loop SBO framework combining continuous generation, geometry-aware sampling, and reward-guided optimization.
• Demonstrate FlowBoost on sphere packing, circle packing, Heilbronn problem, and star discrepancy to obtain competitive or improved results.
• Show that domain-specific inductive biases and reward-guided learning can match or exceed LLM-based systems at lower computational cost.

Proposed method

• Adopt conditional flow matching to learn a time-dependent vector field that transports a simple prior to high-quality configurations.
• Incorporate geometry-aware sampling by interleaving flow integration with projection onto the constraint manifold to maintain feasibility.
• Integrate geometric penalties and an overlap energy to enforce constraints during training (soft penalties and projections).
• Use reward-guided fine-tuning (RG-CFM) with a teacher-student setup and importance-weighted updates to bias sampling toward high-reward configurations while avoiding collapse.
• Employ a GAS-based inference procedure with adaptive walls, Gauss–Newton projections, proximal relaxation, and terminal refinement to ensure feasible samples.
• Define problem-specific rewards (e.g., effective radius for packing) and use a weighted flow-matching loss with consistency regularization to maintain diversity while improving performance.

Figure 2 . Comparison of the minimum-excess metric for two local search heuristics (common circle counts). The physics-push heuristic yields consistently higher minimum-excess values than SRP/SRS, indicating worse configurations across the tested regime.

Experimental results

Research questions

• RQ1 Can a flow-based generative model with geometry-aware sampling discover extremal geometric configurations efficiently in continuous spaces?
• RQ2 Does a closed-loop reward-guided SBO framework outperform open-loop generative approaches and LLM-based methods on classic extremal geometry problems?
• RQ3 How do geometric constraints and inductive biases affect sample feasibility and discovery of rare configurations?
• RQ4 To what extent can FlowBoost improve known constructions in sphere packing, circle packing, Heilbronn, and star discrepancy problems?

Key findings

• FlowBoost discovers configurations that match or exceed best known results on several geometric optimization problems.
• For circle packings, FlowBoost improves the best known lower bounds and outperforms the LLM-based AlphaEvolve with lower computational resources.
• In Heilbronn problem instances, FlowBoost improves the minimum triangle area toward best known numerical values.
• For sphere packing in dimension 12, FlowBoost finds denser configurations than some classical heuristics.
• FlowBoost is presented as the first systematic application of flow-based generative models with RL to extremal mathematics, achieving strong results with reduced compute and without LLMs.

(a) $N=55$ : training vs. RG-CFM (pushed).