This work was done during a PhD internship at FAIR NYC, thanks to my amazing supervisors @bdamos.bsky.social, Ricky T.Q. Chen and Brian Karrer.

Special thanks to our core contributors:
@bkmi.bsky.social, Bing Yan, Xiang Fu, Guang-Horng Liu (and of course Carles Domingo-Enrich).

Our evaluation offers a new, challenging, *amortized* sampling benchmark for molecular conformer generation.

The benchmark features real, drug-like molecules from the SPICE dataset, and we hope it drives direct and tangible progress in sampling for computational chemistry (coming soon).

This lets us train conditional diffusion samplers directly from expensive energy functions, namely, state-of-art molecular foundation models, amortizing sampling across thousands of molecules—unlike traditional samplers, which require heavy energy access per new molecule structure, per sample.

We specialize Adjoint Matching—originally designed for reward fine-tuning—to the sampling setting.

By exploiting a factorization of the optimal transition density (a Schrödinger bridge), our new loss enables heavy reuse of simulations and energy evaluations.

This lets us train conditional diffusion samplers directly from expensive energy functions, namely, state-of-art molecular foundation models, amortizing sampling across thousands of molecules—unlike traditional samplers, which require heavy energy access per new molecule structure, per sample.

We specialize Adjoint Matching—originally designed for reward fine-tuning—to the sampling setting.

By exploiting a factorization of the optimal transition density (a Schrödinger bridge), our new loss enables heavy reuse of simulations and energy evaluations.