Normalizing Flow Maps combines few-step generation with exact likelihoods, NeurIPS oral

sedielem · x · 2026-10-20

• Paper: "Normalizing Flow Maps" (arXiv:2610.22711) by Francesco M. Ruscio, Avishek Joey Bose, Michael Bronstein, Alexander Tong, et al., accepted as an oral at the NeurIPS 2026 STODY workshop.
• Core idea: Combines few-step transport from Flow Maps with tractable likelihoods from Normalizing Flows, introducing exactly invertible pushforward maps that learn the integrator of an underlying dynamical system while supporting efficient exact likelihood evaluation by construction.
• Training: Two complementary consistency axes enable (self) distillation objectives — pointwise matching (ℓ2-based flow-map style regression losses) and distributional matching (aligning pushforward distributions with marginals of a prescribed forward process).
• Results: State-of-the-art sample quality among parameter-matched normalizing flows on CIFAR-10 and ImageNet-64 conditional generation, with likelihood-based RLHF support; also outperforms prior SOTA normalizing flows on equilibrium sampling in Boltzmann Generators.
• Paper, blog, and open-source code are available.

Original post →

More from Research

• Atomwell debuts: 9B foundation model shows clean scaling laws on 80M experimental measurements — chaitjo · 2026-10-20
• SFT vs RLVR: how fine-tuning paradigms shape LLM confidence calibration in high-stakes use — burkov · 2026-10-20
• OpenAI dropped 722 AI-generated math papers in a day; can humans read them? — songyoupeng · 2026-10-20
• Skeptic points to a critical paper everyone should read on fly brain simulations — ShahabBakht · 2026-10-20
• UCSB team wins $1M NSF grant to model higher-order interactions in physical systems with AI — ninamiolane · 2026-10-20
• AI2's MolmoB0T: Zero-Shot Real-World Manipulation from 1.7M Simulated Trajectories — chris_j_paxton · 2026-10-20

Research channel →