Autoregressive Boltzmann Generators

A new autoregressive framework called Autoregressive Boltzmann Generators (ArBG) has been proposed to improve the sampling of molecular systems at thermodynamic equilibrium, moving beyond the normalizing-flow paradigm that has dominated the field [1]. The framework, detailed in a preprint posted to arXiv on June 25, 2026, departs from the flow-based Boltzmann Generator (BG) paradigm [1]. Existing BGs combine a generative model with exact likelihoods and an importance sampling correction to rapidly produce uncorrelated equilibrium samples, but they predominantly rely on normalizing flows [1]. These flows face limitations: discrete-time versions suffer from restricted expressivity due to strict invertibility constraints, while continuous-time variants incur computationally expensive likelihoods [1]. ArBG circumvents the topological constraints of flows and enables sequential inference-time interventions [1]. Boltzmann Generators are a class of energy-based models, an approach that applies the canonical ensemble formulation from statistical physics to learn from data [4]. Energy-based models detect latent variables in a dataset and generate new datasets with a similar distribution [4]. The ArBG framework enhances scalability by leveraging architectures effective in Large Language Models (LLMs) [1]. LLMs are typically based on the transformer architecture, which uses a parallel multi-head attention mechanism to contextualize tokens and has been widely adopted for training large models on extensive datasets [3][6]. The researchers empirically demonstrate that ArBG leads to significant improvements over flow-based models across all benchmarks, with particularly strong results in larger peptide systems such as the 10-residue Chignolin [1]. They also introduce Robin, a 132 million parameter transferable model trained with the ArBG framework [1]. Robin improves over the previous state-of-the-art, reducing the zero-shot energy error on 8-residue systems by over 60% [1]. The preprint, which has not yet been peer-reviewed, is available on arXiv, an open-access repository that hosts scientific papers in fields including physics and computer science before journal publication [7]. The code for the project has been released on GitHub [1].