Michal Balcerak

Kempner Research Fellow @ Harvard University | energy-based models, generative modeling & reasoning

[Michal Balcerak @X_MichalB]

Generative Modeling: What's After Flow Matching?

Flow Matching lacks explicit modeling of scores on the data manifold.
Introducing Energy Matching [NeurIPS 2025] unlocking exciting new inference-time capabilities!
Paper: https://t.co/LYfiU4Cn6B
Code: https://t.co/oyeOpx8lUH

[Michal Balcerak @X_MichalB]

Happy to join Harvard University’s @KempnerInst as a Research Fellow!

I’ll continue to explore energy-based models for generation and reasoning. Excited about the future of the field!

If you work on similar questions, let’s connect here or at NeurIPS in Sydney (presenting GEM).

[Michal Balcerak @X_MichalB]

@soranghadri @UniofOxford congratulations !

[Michal Balcerak @X_MichalB]

@willdepue Great point! We worked on this here for continuous spaces: https://t.co/LYfiU4BPh3, and here for discrete spaces: https://t.co/jchfGywCwJ. Exciting to use test-time adaptation!

[Hannes Stark @HannesStaerk]

Monday Starkly Speaking: we have a session with @X_MichalB explaining the Jordan–Kinderlehrer–Otto (JKO) scheme based on the "Energy Matching" paperhttps://arxiv.org/abs/2504.10612

On Zoom 12pm ET / 6pm CEST: https://t.co/1qQiUveH7S

[Michal Balcerak @X_MichalB]

What’s after discrete diffusion? 💎🧬

Graph Energy Matching (GEM): molecular graphs sampled via transport-aligned discrete proposals.

SOTA among graph EBMs. Matches/surpasses diffusion/flow baselines at generation. Supercharges inference-time design!

https://t.co/BmAEq6NbvL

[Sean Welleck @wellecks]

Excited to announce our workshop on flow-based generative models at CMU:

Frontiers of Flows for Generative AI
March 26-27, Pittsburgh PA

https://t.co/U52Mx5vIYf

We have an amazing lineup of featured talks, panel discussions, and lightning talks. Registration is now open!

[Michal Balcerak @X_MichalB]

@rishabh16_ Great. Looking forward! See you.

[Michal Balcerak @X_MichalB]

@soranghadri @antonio_terpin @bjoern_menze @BogenspergerLea @menze_group It handles them effectively: the first derivative (gradient) points directly toward the target distribution, while the second derivative vanishes, meaning there's no curvature to trap samples. This differs from the scenario you're assuming, where both derivatives vanish.

[Michal Balcerak @X_MichalB]

Generative Modeling: What's After Flow Matching?

Flow Matching lacks explicit modeling of scores on the data manifold.
Introducing Energy Matching [NeurIPS 2025] unlocking exciting new inference-time capabilities!
Paper: https://t.co/LYfiU4Cn6B
Code: https://t.co/oyeOpx8lUH

[Michal Balcerak @X_MichalB]

@ourtown2 @antonio_terpin @bjoern_menze @BogenspergerLea @menze_group Hi Brian- thanks. However, I don't agree with these points. For example, regarding 'semantic curvature ignorance': we explicitly train the landscape to have minimal curvature during the early stages of generation, allowing arbitrary curvature afterward.

[Michal Balcerak @X_MichalB]

@soranghadri @antonio_terpin @bjoern_menze @BogenspergerLea @menze_group Hi Soran. Thanks. Ideally, we use sampling-free losses in the warm-up, since it is easier to compute. OT is the gold standard - samples do not get lost off-manifold in local wells as there is little curvature to trap them. I imagine there can be more ideas.

[Michal Balcerak @X_MichalB]

@AlexiGlad Thanks! The future is an EBM! Here also our recent paper on combining optimal transport theory with EBMs. This combination drastically improves the quality of generation: https://t.co/LYfiU4BPh3

[menze_group @menze_group]

Our 'Glioma Optimizing a Discrete Loss' (GliODIL) method is now in Nature Communications!

🧠 It leverages reaction-diffusion priors with MRI + FET-PET data to map patient-specific glioma spread and guide radiotherapy planning.

[Michal Balcerak @X_MichalB]

@ykilcher Thanks for having me! Great discussion.
📄 Paper: https://t.co/v8mHGnBlLP
💻 Code and tutorial: https://t.co/oyeOpx8lUH

[Yannic Kilcher 🇸🇨 @ykilcher]

Come chat with us tonight on Discord about Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling
6pm UTC, be there: https://t.co/DBnBpDjLND

[Michal Balcerak @X_MichalB]

@StatMLPapers Official GitHub (with an additional tutorial): https://t.co/oyeOpx7O59

[Michal Balcerak @X_MichalB]

@bronzeagepapi Official GitHub (with an additional tutorial): https://t.co/oyeOpx7O59

[Michal Balcerak @X_MichalB]

@fly51fly Official GitHub (with an additional tutorial): https://t.co/oyeOpx8lUH

[Michal Balcerak @X_MichalB]

@StatMLPapers Official GitHub (with an additional tutorial): https://t.co/oyeOpx7O59