Energy-based models are a common way to represent distributions over discrete structures such as molecular graphs. However, a new arXiv preprint (arXiv:2610.19946) highlights a subtle but important mismatch: an energy function that is accurate at equilibrium does not necessarily produce accurate samples when the sampler is stopped after a fixed number of steps.
The authors point out that in practice, generation is always finite-time — samplers cannot run forever. Therefore, they argue, the energy should be learned with this constraint in mind. The title of the work, "Learning Graph Energies for Finite-Time Generation," suggests the proposed solution is to train the energy model directly for the finite-time sampling procedure.
The abstract does not provide details on the specific training objective or experiments, but the core insight is clear: for practical molecular generation, what matters is the quality of samples at the stopping time, not just the theoretical equilibrium distribution. This could have implications for how energy-based models are evaluated and trained in generative chemistry.