Joey Bose

🎉Personal update: I'm thrilled to announce that I'm joining Imperial College London @imperialcollege as an Assistant Professor of Computing @ICComputing starting January 2026. My future lab and I will continue to work on building better Generative Models 🤖, the hardest AI4Science applications in computational biology 🧬and chemistry 🧪, and also a sprinkling of Deep Learning theory 📚 that supports these goals.

@wognsfjq96 @andreamiele_ Thanks for making these connections explicit! In AF3 recycling, there is also no backprop through the iterations, but I do agree that self-conditioning using a standard denoising prediction with stop grad is quite cool!

@bose_joey @andreamiele_ One thing we find interesting about self-conditioning is that it seems to show fixed-point-like behavior even without backpropagating through the iterations or convergence. In practice, standard denoising training with a stop-gradient on the previous prediction seems to be enough

Self-conditioning can be distilled into one-step language generation. We show that self-conditioned flow LMs perform fixed-point iteration, and introduce fixed-point flow maps to compress both the iterations and the entire flow. At one step, FMLM★ achieves 112.5 gPPL at near-data entropy (5.37), substantially improving over FMLM’s 168.3 gPPL at 5.17 entropy. FMLM★ also achieves the best 2–4 step results. 📎 Paper: arxiv.org/abs/2607.00714 ⌨️ Code: github.com/Ugness/self-co…

@wognsfjq96 @andreamiele_ Can you comment on how this relates to known self-conditioning mechanisms like recycling in AlphaFold 3 and Deep Equilibrium model ideas as used in diffusion models? These ideas seem quite related to me but not sure if you see them the same way.

@andreamiele_ Thanks! We believe this structure is universal and can be adapted to many other generative/predictive systems.

We will be presenting our poster today at 4-5pm at the #SPIGM workshop at #ICML2026. If you are working on diffusion for text, either using discrete or continuous space approaches, you will find our results interested. Come chat with us. With @rozonoyer96703 and Jacopo Minniti.

We're presenting DeCAF ☕️: A few-step cofolding model with all-atom flow maps at the GenBio workshop at #ICML2026 as a Spotlight in the morning poster session!! Also, we've now released the code and checkpoints from the paper: code🧑‍💻: github.com/genesistherape… paper🔗: arxiv.org/abs/2606.08375 Come by and say hi 👋!

We're presenting Autoregressive Boltzmann Generators now!!! Come by to #1404 to catch the vibes

Riemannian MeanFlow 🌀🌊 will be at #ICML2026! (don’t worry though, it’s actually a pretty NiceFlow 😎) RMF extends the MeanFlow framework to Riemannian manifolds by learning average velocities directly on the manifold -- this enables fast generation of biomolecules in high dimensions. ⚡️ RMF can generate discrete DNA sequences on the simplex manifold — language flow maps in biology! 🧬 We also show that reward guidance can generate proteins and DNA with structures or functions that we want using fewer steps. 🏆 Huge congrats to @dywoo1247 who led this project and worked incredibly hard to identify the best losses and practices for training these challenging models 🔥 had an awesome time working on this with @hyuunnnnnn, @k_neklyudov, and @sungsoo_ahn_ 🤗 also, if you like this work, check out GFM from Davis et al. who approach the same problem from a different angle :)

@Qu3ntinB @sedielem @icmlconf @BaadeAlan @mathusmassias @ZhengyangGeng @MiaosenWang @thoma_gu @AlexanderTong7 Will be there!

@sedielem @icmlconf @BaadeAlan @mathusmassias @ZhengyangGeng @MiaosenWang @thoma_gu I will be there! @bose_joey @AlexanderTong7

📢 Diffusion circle at @icmlconf 2026: join us Thursday July 9 at 3:30PM at the information desk / job board, we'll head out from there and find a spot to sit. No agenda, just get together and talk shop. Please tell your friends and tag people who might be interested!

🚨 Categorical Flow Maps‼️ Happening NOW, Hall A #2715 with @FEijkelboom @daan_roos_ 🤝🔥 #ICML

Our book "Generative AI and Stochastic Thermodynamics: A Tale of Free Energies" is out this month. Over the coming weeks @sirui_lu97, @wellingmax and I will post about some of the topics it covers. Starting it off: heat and work in latent variable models and variational EM.

Starting the day with some coffee, Diffusion and Flow matching tutorial by @Qu3ntinB and @mathusmassias

I'm attending #ICML2026, presenting a couple papers: 1. "Autoregressive Boltzmann Generators" (Spotlight) arxiv.org/abs/2606.27361 and 2. Few-Step CoFolding with All-Atom Flow Maps---i.e., the DeCAF paper ☕️ (Spotlight) at the GenBio workshop. arxiv.org/pdf/2606.08375 Happy to meet old friends and make new ones! DMs open :)

@danyalrehman17 Awesome work! That is one of the most exciting papers I saw in the last year :)

@FrankNoeBerlin Thanks 🙏. Couldn't agree more about making some progress on BGs!

@bose_joey Making progress on Boltzmann Generators is never stupid 😅 Congratulations!

Can we break free from the tyranny of diffusion and flow matching for atomistic molecular generation 👀? Turns out we can reliablyn scale an Autoregressive model with simplest and perhaps stupidest approach 😅. Check out the new #ICML2026 spotlight paper Autoregressive Boltzmann Generators in the quote tweet.

@jonkhler Thanks for the praise 🙏. Still surprised that this worked as well it did!

@bose_joey Great stuff! :) remember I was daydreaming about something like this close to ten years ago but told myself: ah no, that's never feasible! Great to see proven otherwise!

@giannis_daras Riding the LLM hype train 🚂! (Only for ICML)

@bose_joey That feels like betrayal to me 💀

@giannis_daras I said what I said 😎

@bose_joey Tyranny of diffusion? Who are you and did you do to Joey?