OpenNovelty.org

FALCON: Few-step Accurate Likelihoods for Continuous Flows

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.0Download Report PDF
Generative ModelsFlow MatchingBoltzmann GeneratorsAI for Science
Abstract:

Scalable sampling of molecular states in thermodynamic equilibrium is a long-standing challenge in statistical physics. Boltzmann Generators tackle this problem by pairing a generative model, capable of exact likelihood computation, with importance sampling to obtain consistent samples under the target distribution. Current Boltzmann Generators primarily use continuous normalizing flows (CNFs) trained with flow matching for efficient training of powerful models. However, likelihood calculation for these models is extremely costly, requiring thousands of function evaluations per sample, severely limiting their adoption. In this work, we propose Few-step Accurate Likelihoods for Continuous Flows (FALCON), a method which allows for few-step sampling with a likelihood accurate enough for importance sampling applications by introducing a hybrid training objective that encourages invertibility. We show FALCON outperforms state-of-the-art normalizing flow models for molecular Boltzmann sampling and is \emph{two orders of magnitude faster} than the equivalently performing CNF model.

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 proposes FALCON, a method for accelerating likelihood computation in continuous normalizing flows (CNFs) used for Boltzmann sampling. It sits within the 'Boltzmann Generators with Continuous Normalizing Flows' leaf, which contains four papers including the original work. This leaf is part of the broader 'Normalizing Flow-Based Generators' branch, which also includes equivariant flows and rigid body flows as sibling leaves. The taxonomy reveals this is a moderately populated research direction within the larger field of fifty papers, suggesting active but not overcrowded exploration of CNF-based approaches for molecular sampling.

The paper's position within the normalizing flow branch distinguishes it from neighboring diffusion-based methods (e.g., 'Energy-Based Diffusion Training', 'Torsional and Internal Coordinate Diffusion') and GFlowNets. The taxonomy shows clear boundaries: normalizing flows emphasize exact likelihood computation and invertibility, while diffusion methods trade off likelihood tractability for flexible denoising processes. FALCON's focus on few-step sampling with accurate likelihoods addresses a computational bottleneck specific to CNFs, contrasting with diffusion acceleration techniques like consistency models or distillation found in sibling branches. The 'Sampling Acceleration and Efficiency Techniques' category exists separately, indicating FALCON bridges architectural innovation with efficiency concerns.

Among twenty-seven candidates examined, the contribution-level analysis reveals mixed novelty signals. The core FALCON method (Contribution A) examined ten candidates with one appearing to provide overlapping prior work, suggesting some precedent for few-step flow acceleration exists within this limited search scope. The hybrid training objective for invertibility (Contribution B) examined seven candidates with one potential refutation, indicating prior exploration of invertibility constraints. The scalable equivariant architecture (Contribution C) examined ten candidates with none clearly refuting it, suggesting this component may be more distinctive. These statistics reflect a focused semantic search, not exhaustive coverage of the flow-based sampling literature.

Based on the limited search scope of twenty-seven top-ranked candidates, FALCON appears to combine known elements—CNF acceleration, invertibility training, equivariant architectures—in a novel configuration targeting a specific computational bottleneck. The taxonomy context shows this work incrementally advances a moderately active research direction rather than opening entirely new territory. The analysis cannot assess whether broader literature beyond the top-K semantic matches contains additional relevant prior work, particularly in adjacent fields like general normalizing flow acceleration or non-molecular applications of few-step flows.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
27
Contribution Candidate Papers Compared
2
Refutable Paper

Research Landscape Overview

Core task: sampling molecular conformations from Boltzmann distributions. The field has organized itself around several complementary directions. Generative Model Architectures for Boltzmann Sampling explores diverse neural frameworks—normalizing flows (including continuous variants like those in Boltzmann Generators[6]), diffusion models (e.g., Torsional Diffusion[4]), and flow matching approaches (Equivariant Flow Matching[10])—each offering different trade-offs between expressivity and computational cost. Sampling Acceleration and Efficiency Techniques focuses on speeding up convergence through methods like parallel tempering (Parallel Tempering Algorithm[20]) and flow perturbation (Flow Perturbation Acceleration[18]). Surrogate Models and Transfer Operators leverage learned approximations of dynamics (Surrogate Model MD[9], Implicit Transfer Operator[14]) to bypass expensive simulations, while Conformational Ensemble Prediction and Analysis emphasizes extracting ensemble properties and validating generated distributions. Domain-Specific Applications and Benchmarks ground the methodology in real molecular systems, Theoretical Foundations provide rigorous underpinnings, and Energy-Weighted and Continuous Flow Training refine how models incorporate energetic information during learning (Energy-Weighted Flow Matching[35]). Recent work has intensified around two contrasting themes: whether to use discrete-time diffusion or continuous-time flows, and how tightly to couple energy functions during training versus sampling. Iterated Denoising Energy[1] and Consistent Sampling Simulation[3] exemplify efforts to improve diffusion-based samplers, while Transferable Boltzmann Generators[5] and Sequential Boltzmann Generators[8] extend flow-based methods to new regimes. FALCON[0] sits within the normalizing flow branch alongside Boltzmann Generators[6], emphasizing continuous normalizing flows for direct Boltzmann sampling. Compared to Transferable Boltzmann Generators[5], which prioritizes generalization across molecular families, FALCON[0] focuses on refining the flow architecture itself for improved sampling fidelity. The interplay between architectural innovation and training strategies—whether to weight samples by energy or rely on post-hoc reweighting—remains an active question shaping how practitioners balance accuracy, efficiency, and transferability across diverse molecular systems.

Claimed Contributions

FALCON: Few-step accurate likelihoods for continuous flows

The authors introduce FALCON, a continuous flow-based generative model that enables few-step sampling while providing fast and accurate likelihood computation for Boltzmann sampling. The method uses a hybrid training objective combining regression loss and a cycle-consistency term to encourage invertibility, making it suitable for importance sampling applications.

10 retrieved papers
Can Refute
Hybrid training objective for invertibility

The authors propose a hybrid training objective that combines a regression loss for stable few-step generation with a cycle-consistency term to encourage invertibility prior to convergence. This design allows the model to be invertible, trainable with regression loss, and compatible with free-form architectures while supporting efficient likelihood evaluation.

7 retrieved papers
Can Refute
Scalable softly equivariant continuous flow architecture

The authors introduce a simple and scalable softly equivariant continuous flow architecture that significantly improves over existing state-of-the-art equivariant flow model architectures. This architectural contribution enables the use of larger and more expressive models for molecular sampling tasks.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

FALCON: Few-step accurate likelihoods for continuous flows

The authors introduce FALCON, a continuous flow-based generative model that enables few-step sampling while providing fast and accurate likelihood computation for Boltzmann sampling. The method uses a hybrid training objective combining regression loss and a cycle-consistency term to encourage invertibility, making it suitable for importance sampling applications.

Contribution

Hybrid training objective for invertibility

The authors propose a hybrid training objective that combines a regression loss for stable few-step generation with a cycle-consistency term to encourage invertibility prior to convergence. This design allows the model to be invertible, trainable with regression loss, and compatible with free-form architectures while supporting efficient likelihood evaluation.

Contribution

Scalable softly equivariant continuous flow architecture

The authors introduce a simple and scalable softly equivariant continuous flow architecture that significantly improves over existing state-of-the-art equivariant flow model architectures. This architectural contribution enables the use of larger and more expressive models for molecular sampling tasks.

Table of Contents