Learning Boltzmann Generators via Constrained Mass Transport

ICLR 2026 Conference SubmissionAnonymous Authors
samplingBoltzmann generatorsannealing
Abstract:

Efficient sampling from high-dimensional and multimodal unnormalized probability distributions is a central challenge in many areas of science and machine learning. We focus on Boltzmann generators (BGs) that aim to sample the Boltzmann distribution of physical systems, such as molecules, at a given temperature. Classical variational approaches that minimize the reverse Kullback–Leibler divergence are prone to mode collapse, while annealing-based methods, commonly using geometric schedules, can suffer from mass teleportation and rely heavily on schedule tuning. We introduce Constrained Mass Transport (CMT), a variational framework that generates intermediate distributions under constraints on both the KL divergence and the entropy decay between successive steps. These constraints enhance distributional overlap, mitigate mass teleportation, and counteract premature convergence. Across standard BG benchmarks and the here introduced ELIL tetrapeptide, the largest system studied to date without access to samples from molecular dynamics, CMT consistently surpasses state-of-the-art variational methods, achieving more than 2.5× higher effective sample size while avoiding mode collapse.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Constrained Mass Transport (CMT), a variational framework for sampling Boltzmann distributions in molecular systems by imposing constraints on KL divergence and entropy decay between annealing steps. It resides in the Physics-Based Sampling Applications leaf, which contains only two papers total. This sparse positioning suggests the work addresses a relatively specialized niche within the broader sampling literature, focusing specifically on molecular Boltzmann generators rather than general-purpose multimodal sampling. The sibling paper in this leaf addresses different physics constraints, indicating limited direct competition in this exact problem formulation.

The taxonomy reveals that CMT bridges multiple methodological traditions. Its closest conceptual neighbors include Flow-Based Annealing methods (three papers combining flows with annealing schedules) and Tempered Transitions approaches (two papers using geometric annealing with Langevin dynamics). The framework's constraint-based formulation also connects to Optimal Transport samplers (two papers using Wasserstein gradient flows). However, CMT's physics-specific focus distinguishes it from these general-purpose methods: the scope note for Physics-Based Sampling Applications explicitly excludes general Boltzmann sampling methods, placing CMT in a domain-constrained application category rather than among core methodological innovations.

Among twenty-three candidates examined, no refutable prior work was identified across the three contributions. The CMT framework itself was assessed against three candidates with no overlaps found. The connection between constrained optimization and annealing paths examined ten candidates without refutation, as did the ELIL tetrapeptide benchmark. This absence of refutations within the limited search scope suggests either genuine novelty in the specific constraint formulation or that the search did not surface closely related constraint-based annealing work. The benchmark contribution appears particularly novel, being described as the largest system studied without molecular dynamics samples.

Based on examination of twenty-three semantically similar papers, the work appears to occupy a distinct position combining transport-based constraints with physics-specific Boltzmann sampling. The sparse taxonomy leaf and zero refutations across contributions suggest novelty within the examined scope, though the limited search scale means potentially relevant constraint-based or physics-informed annealing methods may exist beyond the top-K semantic matches. The framework's integration of mass transport constraints with variational Boltzmann generation represents a specific methodological combination not clearly anticipated by the surveyed literature.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
23
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: sampling from high-dimensional multimodal unnormalized probability distributions. The field organizes itself around several complementary strategies for navigating complex energy landscapes with multiple modes. Flow-Based and Transport Methods (e.g., Lattice Field Flow[3], Energy-Weighted Flow[18]) learn continuous transformations to map simple base distributions onto target densities, while Diffusion and Score-Based Methods (e.g., Multiscale Denoising Score[7], Particle Denoising Diffusion[30]) reverse stochastic processes to generate samples. MCMC and Tempering Approaches (e.g., Tempered Transitions[2], Tempered Hamiltonian Monte Carlo[4]) employ temperature schedules or parallel chains to escape local modes, and Annealed Importance Sampling techniques (e.g., Sequential Monte Carlo[11], Pigeons[23]) estimate normalizing constants while bridging distributions. Specialized Sampling Frameworks address structured problems such as constrained spaces or tensor decompositions, Application-Driven Sampling Methods tackle domain-specific challenges in physics and engineering, and Methodological Comparisons (e.g., Bayesian Sampling Comparison[5]) evaluate trade-offs across paradigms. Recent work highlights tensions between gradient-based versus gradient-free strategies, the role of learned versus hand-crafted annealing schedules, and the interplay of flow models with tempering ideas. Within Application-Driven Sampling Methods, physics-based applications often impose hard constraints or exploit symmetries, as seen in Constrained Theoretical Spaces[36] and Approximate Symmetries[10]. Constrained Mass Transport[0] sits naturally in this physics-oriented cluster, emphasizing mass conservation and geometric constraints that arise in physical simulations. Compared to nearby works like Constrained Theoretical Spaces[36], which focuses on theoretical guarantees in constrained settings, Constrained Mass Transport[0] appears to integrate transport-based machinery with domain-specific physical constraints, bridging the gap between flow methods and application requirements. This positioning reflects a broader trend of adapting general-purpose sampling tools to respect the structure inherent in scientific problems.

Claimed Contributions

Constrained Mass Transport (CMT) framework

The authors propose a variational framework that generates intermediate distributions under constraints on both the KL divergence (trust-region) and the entropy decay between successive steps. This framework addresses mode collapse and mass teleportation issues in sampling from unnormalized probability distributions.

3 retrieved papers
Connection between constrained optimization and annealing paths

The authors formally characterize how iteratively solving constrained variational problems induces annealing paths (geometric, tempered, and geometric-tempered) that interpolate between a base distribution and the target, as shown in Theorem 2.4.

10 retrieved papers
ELIL tetrapeptide benchmark system

The authors introduce a new molecular benchmark system (ELIL tetrapeptide with d=219 dimensions) that represents the largest and most complex system studied using variational approaches for learning Boltzmann generators purely from energy evaluations without molecular dynamics samples.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Constrained Mass Transport (CMT) framework

The authors propose a variational framework that generates intermediate distributions under constraints on both the KL divergence (trust-region) and the entropy decay between successive steps. This framework addresses mode collapse and mass teleportation issues in sampling from unnormalized probability distributions.

Contribution

Connection between constrained optimization and annealing paths

The authors formally characterize how iteratively solving constrained variational problems induces annealing paths (geometric, tempered, and geometric-tempered) that interpolate between a base distribution and the target, as shown in Theorem 2.4.

Contribution

ELIL tetrapeptide benchmark system

The authors introduce a new molecular benchmark system (ELIL tetrapeptide with d=219 dimensions) that represents the largest and most complex system studied using variational approaches for learning Boltzmann generators purely from energy evaluations without molecular dynamics samples.