Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control. arxiv.org/abs/2409.08861

New paper! We cast reward fine-tuning as stochastic control. 1. We prove that a specific noise schedule *must* be used for fine-tuning. 2. We propose a novel algorithm that is significantly better than the adjoint method*. (*this is an insane claim) arxiv.org/abs/2409.08861

This is a really jam-packed and exciting work. I'm extremely proud of Carles Domingo-Enrich for taking on an ambitious project at FAIR, resulting in this absolutely outstanding paper.

Congrats on Carles Domingo-Enrich Ricky T. Q. Chen on such a nice paper using #stochastic #optimal #control for fine-tuning (controlling) pre-trained models (base dynamics) given task-specific criteria (reward / cost)!

Next Monday at 12 pm ET, Carles Domingo-Enrich will join LoGG host Hannes Stärk to discuss stochastic optimal control (SOC) and a new algorithm named Adjoint Matching. 📄 Read the paper: arxiv.org/abs/2409.08861 ⏺️ Tune in live: zoom.us/j/5775722530?p…

On Monday, Carles Domingo-Enrich will present his new "Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control" arxiv.org/abs/2409.08861 Quite fresh off the press! Monday zoom at 9am PT / 12pm ET / 6pm CEST: portal.valencelabs.com/logg

New paper! A taxonomy of loss functions for stochastic optimal control: arxiv.org/pdf/2410.00345 If our recent work on Adjoint Matching (arxiv.org/abs/2409.08861) made you want to learn about deep-learning SOC techniques, check my systematic study on all available losses!

Happy to release a concise review of recent progress in sequence-structure co-generation for protein design. with our amazing intern chentongwang + Sarah Alamdari Carles Domingo-Enrich and Ava Amini

Undergrad internship opportunities at MSR! If you are a rising junior or senior undergraduate student interested in working with me, apply here and mention my name: aka.ms/msr-ugrad. Topics: fine-tuning and inference of generative models for continuous and discrete data.

New paper out! We introduce “Generator Matching” (GM), a method to build GenAI models for any data type (incl. multimodal) with any Markov process. GM unifies a range of state-of-the-art models and enables new designs of generative models. arxiv.org/abs/2410.20587 (1/5)

I’m excited to share our latest work on generative models for materials called FlowLLM. FlowLLM combines Large Language Models and Riemannian Flow Matching in a simple, yet surprisingly effective way for generating materials. arxiv.org/abs/2410.23405 Benjamin Kurt Miller Ricky T. Q. Chen Brandon Wood

If you are a PhD student and want to intern with me or my colleagues at Microsoft Research New England Microsoft Research, please apply at jobs.careers.microsoft.com/global/en/job/…

This paper didn’t get the attention it deserves. Have you ever wondered why RLHF/reward fine-tuning is far less common for image generation models compared to large language models (LLMs)? This paper by Ricky T. Q. Chen and team explains it in details. 🧵

Why permute when you can cheaply permute?

Our MIT class “6.S184: Introduction to Flow Matching and Diffusion Models” is now available on YouTube! We teach state-of-the-art generative AI algorithms for images, videos, proteins, etc. together with the mathematical tools to understand them. diffusion.csail.mit.edu (1/4)

New paper out with FAIR(+FAIR-Chemistry): Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching We present a scalable method for sampling from unnormalized densities beyond classical force fields. 📄: arxiv.org/abs/2504.11713

This work was done during a PhD internship at FAIR NYC, thanks to my amazing supervisors Brandon Amos , Ricky T. Q. Chen and Brian Karrer. Special thanks to our core contributors: Benjamin Kurt Miller Bing Yan Xiang Fu Guan-Horng Liu (and of course Carles Domingo-Enrich)

Reward-driven algorithms for training dynamical generative models significantly lag behind their data-driven counterparts in terms of scalability. We aim to rectify this. Adjoint Matching poster Carles Domingo-Enrich Sat 3pm & Adjoint Sampling oral Aaron Havens Mon 10am FPI

🚀Excited to open source the code for Adjoint Matching --- as part of a new repo centered around reward fine-tuning via stochastic optimal control! github.com/microsoft/soc-…

Against conventional wisdom, I will be giving a talk with particular focus on the "how" and the various intricacies of applying stochastic control for generative modeling. Mon 9:50am Hall 1 Apex #ICLR2025 Also check out the other talks at delta-workshop.github.io!