On the flow matching interpretability

2025-10-24

AI Summary

Current flow-matching generative models lack interpretability in their intermediate dynamics. To address this, we propose a physics-guided flow-matching framework that explicitly constrains each step of the generative flow to correspond to the thermal equilibrium evolution of a 2D Ising model along a continuous annealing path—thereby endowing latent-space trajectories with rigorous physical semantics. Our method adopts an encoder–flow-matching network–projector architecture to realize temperature-driven diffusion modeling in a continuous latent space. Experiments demonstrate high physical fidelity across multi-scale lattices, significantly faster generation than conventional Monte Carlo methods, with accelerating speedup as system size increases. The core contribution is establishing a mathematically grounded mapping between generative dynamics and statistical mechanical equilibrium states, introducing a novel paradigm for interpretable generative modeling grounded in physical principles.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Generation

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Generative models based on flow matching have demonstrated remarkable success in various domains, yet they suffer from a fundamental limitation: the lack of interpretability in their intermediate generation steps. In fact these models learn to transform noise into data through a series of vector field updates, however the meaning of each step remains opaque. We address this problem by proposing a general framework constraining each flow step to be sampled from a known physical distribution. Flow trajectories are mapped to (and constrained to traverse) the equilibrium states of the simulated physical process. We implement this approach through the 2D Ising model in such a way that flow steps become thermal equilibrium points along a parametric cooling schedule. Our proposed architecture includes an encoder that maps discrete Ising configurations into a continuous latent space, a flow-matching network that performs temperature-driven diffusion, and a projector that returns to discrete Ising states while preserving physical constraints. We validate this framework across multiple lattice sizes, showing that it preserves physical fidelity while outperforming Monte Carlo generation in speed as the lattice size increases. In contrast with standard flow matching, each vector field represents a meaningful stepwise transition in the 2D Ising model's latent space. This demonstrates that embedding physical semantics into generative flows transforms opaque neural trajectories into interpretable physical processes.

Problem

• Lack of interpretability in flow matching generative models
• Opaque intermediate steps in neural generation processes
• Mapping flow trajectories to physical equilibrium states

Innovation

• Flow steps constrained to physical distribution trajectories
• Encoder-flow-projector architecture ensures physical constraint preservation
• Vector fields represent interpretable Ising model transitions