Bayesian Methods Research Group (@bayesgroup) header
Bayesian Methods Research Group (@bayesgroup) profile picture

Research in Bayesian Deep Learning, Reinforcement Learning, Optimization, Structured Prediction, Drug Discovery and more

Joined July 2017
Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)Photo by Bayesian Methods Research Group (@bayesgroup)
Bayesian Methods Research Group retweeted
๐Ÿ’จ Did you know neural networks can behave like literal ideal gases? ๐Ÿ’จ In our #IJCAI2026 paper, we show that at stationarity, scale-invariant NNs surprisingly obey... the ideal gas law ๐Ÿคฏ Thermodynamics helps us to understand how train hyperparams shape final solutions! ๐Ÿงต๐Ÿ‘‡1/8
1
14
3
48
5,895
Bayesian Methods Research Group retweeted
๐ŸŒ๐ŸŒ๐ŸŒ 1/ The original GFlowNets paper tried PPO for sampling, but it failed. We figured out why and fixed it. And now PPO beats all objectives on standard problems including molecular graph generation in large spaces.
10
37
1
340
24,809
Did you like our research on neural networks weights and loss landscapes? Be sure to check out the @weightsummetry #ICML2026 workshop co-organized by our alumni @KateLobacheva
๐Ÿ“ขExcited to announce the Workshop on Weight-Space Symmetries @icmlconf! We welcome 4-page submissions analysing symmetries, their effects on training and model structure, and practical methods to utilize them. Submission Deadline: April 24 (23:59 AoE) #ICML2026
3
743
โš ๏ธ New paper alert!
Excited to present CasTex (#WACV2026) ๐ŸŽ‰ Our text-to-texture method optimizes explicit PBR maps via SDS on cascaded pixel-space diffusion models, avoiding latent artifacts and producing relightable textures ready for production use. Paper & Code โฌ‡๏ธ thecrazymage.github.io/CasTeโ€ฆ
557
Not a mirage but our new paper!
๐Ÿšจ New paper alert! ๐Ÿšจ Our new paper on ArXiv: "MiAD: Mirage Atom Diffusion for De Novo Crystal Generation". We unlock an ability to dynamically add or remove atoms during generation in diffusion models for better materials discovery. ๐Ÿ‘‰ arxiv.org/abs/2511.14426 1/5
3
501
New paper! #NeurIPS2025
๐Ÿš€COSMOS is OUT! @NeurIPSconf 2025! ๐Ÿ“ˆCOSMOSย achieves up toย 2xย faster text generation compared to other diffusion models, utilizing up toย 8x compression in text representations for superior efficiency. ๐Ÿ“„ย Paper:ย arxiv.org/pdf/2506.21170 ๐Ÿ’ปย Code:ย github.com/MeshchaninovViachโ€ฆ (1/6)
1
498
Bayesian Methods Research Group retweeted
While frontier labs are announcing their new models, we also want to be part of this parade. So, weโ€™re happy to announce gfnx โ€“ a JAX-first library with environments and a single-file baseline implementation for GFlowNet research.
4
10
21
1,463
Bayesian Methods Research Group retweeted
1/ Can we efficiently learn the destruction process of diffusion samplers? Can we learn not just the drift, but also the variance for all transition kernels? โ€“ We answer YES in our recent paper โ€œAdaptive Destruction Processes for Diffusion Samplersโ€ (Oral at NeurIPS 2025 FPI Workshop).
1
9
1
17
2,802
Check out our new paper!
(1/n) The usual assumption in GFlowNet environments is acyclicity. Have you ever wondered if it can be relaxed? Does the existing GFlowNet theory translate to the non-acyclic case? Is efficient training possible? We shed new light on these questions in our latest work! @icmlconf
3
500
Check out our new work!
๐Ÿšจ New paper alert! ๐Ÿšจ Our new paper on ArXiv: "DreamBooth DPO: Controlled Optimization of Personalized Diffusion Models" It addresses the core trade-off in personalized T2I: concept fidelity vs. prompt alignment, without any human-curated data ๐Ÿ‘‰ arxiv.org/abs/2505.20975 1/5
2
393
Check out our new work!
๐Ÿšจ New paper alert! ๐Ÿšจ Our new paper on ArXiv: "ImageReFL: Balancing Quality and Diversity in Human-Aligned Diffusion Models". It tackles a key challenge in diffusion models: aligning with human preferences without collapsing diversity ๐Ÿ‘‰ arxiv.org/abs/2505.22569 1/5
3
388
Check our new paper! Catch the presentation at #ICLR2025 by @gritsaev!
1/ GFlowNets are known for training a forward policy to generate complex objects step by step. However, an equally important piece specific to the GFlowNet paradigm is a backward policy, which undoes these steps and plays a crucial role in training.
2
368
Bayesian Methods Research Group retweeted
๐Ÿ‘จโ€๐Ÿ’ผNeural Flow Diffusion Models at #NeurIPS2024 tomorrow! Discover how to build learnable noising processes for straight-line generative trajectories end-to-end and without simulations!๐Ÿคฏ ๐Ÿ“West Ballroom A-D #6809 โฐFri 13 Dec 4:30 pm โ€” 7:30 pm ๐Ÿ”—neurips.cc/virtual/2024/postโ€ฆ
๐Ÿ”ฅ Excited to share our new work on Neural Flow Diffusion Models โ€” a general, end-to-end, simulation-free framework that works with an arbitrary noising processes and even enables learning them! ๐Ÿ“œ: arxiv.org/abs/2404.12940 ๐Ÿงต 1/11
1
11
1
71
5,423
Check out our new paper! To be presented at #NeurIPS2024 by @KateLobacheva this Friday (poster #2408 / Poster Session 5 East / 13 Dec 11 am โ€“ 2 pm PST)
Starting training with a large learning rate benefits generalizationโ€”but why? In our new #NeurIPS2024 paper, we investigate its role in navigating the loss landscape and its effect on feature learning! 1/7 Paper: arxiv.org/abs/2410.22113 Poster: nips.cc/virtual/2024/poster/โ€ฆ
3
433
๐—ก๐—ฒ๐˜‚๐—ฟ๐—ฎ๐—น ๐—™๐—น๐—ผ๐˜„ ๐——๐—ถ๐—ณ๐—ณ๐˜‚๐˜€๐—ถ๐—ผ๐—ป ๐— ๐—ผ๐—ฑ๐—ฒ๐—น๐˜€: ๐—Ÿ๐—ฒ๐—ฎ๐—ฟ๐—ป๐—ฎ๐—ฏ๐—น๐—ฒ ๐—™๐—ผ๐—ฟ๐˜„๐—ฎ๐—ฟ๐—ฑ ๐—ฃ๐—ฟ๐—ผ๐—ฐ๐—ฒ๐˜€๐˜€ ๐—ณ๐—ผ๐—ฟ ๐—œ๐—บ๐—ฝ๐—ฟ๐—ผ๐˜ƒ๐—ฒ๐—ฑ ๐——๐—ถ๐—ณ๐—ณ๐˜‚๐˜€๐—ถ๐—ผ๐—ป ๐— ๐—ผ๐—ฑ๐—ฒ๐—น๐—น๐—ถ๐—ป๐—ด by Grigory Bartosh, Dmitry Vetrov, Christian A. Naesseth arxiv.org/abs/2404.12940
1
1
1
1,052
๐—›๐—ฎ๐—ถ๐—ฟ๐—™๐—ฎ๐˜€๐˜๐—š๐—”๐—ก: ๐—ฅ๐—ฒ๐—ฎ๐—น๐—ถ๐˜€๐˜๐—ถ๐—ฐ ๐—ฎ๐—ป๐—ฑ ๐—ฅ๐—ผ๐—ฏ๐˜‚๐˜€๐˜ ๐—›๐—ฎ๐—ถ๐—ฟ ๐—ง๐—ฟ๐—ฎ๐—ป๐˜€๐—ณ๐—ฒ๐—ฟ ๐˜„๐—ถ๐˜๐—ต ๐—ฎ ๐—™๐—ฎ๐˜€๐˜ ๐—˜๐—ป๐—ฐ๐—ผ๐—ฑ๐—ฒ๐—ฟ-๐—•๐—ฎ๐˜€๐—ฒ๐—ฑ ๐—”๐—ฝ๐—ฝ๐—ฟ๐—ผ๐—ฎ๐—ฐ๐—ต by Maxim Nikolaev, Mikhail Kuznetsov, Dmitry Vetrov, @ai_alanov arxiv.org/abs/2404.01094
2
1,002
Bayesian Methods Research Group retweeted
Did you know that networks trained with different learning rates extract different features (and a different number of them!) from the data? Come by our poster at HiLD Workshop #ICML2024 tomorrow to discuss it with @irsadrtdinov! Paper: openreview.net/forum?id=IID2โ€ฆ 1/6
3
9
1
45
5,335