Reproduce “Feynman-Kac-Flow: Inference Steering of Conditional Flow Matching to an Energy-Tilted Posterior” on OpenResearch

About this paper

Konstantin Mark, Leonard Galustian, Maximilian P.-P. Kovar, Esther Heid

Researchers from TU Wien developed Feynman-Kac steering for Conditional Flow Matching (FK-CFM), extending a robust inference-time steering framework to a class of generative models often used for faster, higher-quality generation. This method notably improved the generation of chirality-aware transition states in chemical reactions, reducing the Persistent Chirality Error from 9.0% to 0.7% and the Thresholded Chirality Error from 27.8% to 5.8%.

Key insights

• Feynman-Kac steering can be successfully extended to Conditional Flow Matching by re-formulating the deterministic flow as an SDE with injected diffusion, which is critical for enabling robust particle resampling without compromising the original flow's marginal distributions.
• Effective steering of flow matching particles can be achieved using intermediate reward functions, particularly a one-shot Euler approximation of the final reward, which is computationally efficient and leverages the pre-trained velocity field.
• The FK-CFM framework is adept at isolating specific modes in high-dimensional, multimodal distributions and can enforce complex geometric and chemical constraints, addressing challenges in guided generative modeling.

Summary by alphaXiv.