acceptodds
Withdrawn from ICLR 2027

BIT: Boltzmann-Inspired Training for One-Step Molecular Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular forces provide rich local information about the Boltzmann distribution. To harness this information for direct one-step molecular generation, we introduce Boltzmann-Inspired Training (BIT), an attraction–repulsion framework with adaptive training targets. BIT projects reference forces onto atom-pair directions and uses their alignment with differences in pairwise distances between generated and reference conformations to reweight the local attraction. It combines the resulting target with a data-only target using a weight that varies across generated conformations, and trains the generator by regression. On seven MD17 molecules, BIT reduces mean pooled-distance total variation distance (TVD) by 55% and mean atom-pair TVD by 24% relative to the same generator trained without forces. On average across molecules, BIT also achieves lower pooled-distance TVD and distance-histogram mean absolute error (MAE) than AvgFlow and ET-Flow. At inference, the trained generator produces each conformation with a single network evaluation, while force information is used only during training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular conformation sampling, One-step generative modeling, Boltzmann distribution, Training-time force guidance, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_HRJoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BIT, a training method for one-step molecular conformation generators that uses reference forces from MD data. It builds on attraction–repulsion (drifting) training in pairwise-distance space and makes two changes: (1) it reweights the attraction toward each reference conformation using the alignment between its projected force and the distance difference to the generated sample, and (2) it mixes the force-informed and data-only targets per sample, using a statistic derived from a kernel score identity. Forces are only used during training, and sampling takes one network evaluation. On seven MD17 molecules, BIT reduces mean pooled-distance TVD from 0.1056 (no force) to 0.0472, and also outperforms AvgFlow (0.0709) and ET-Flow (0.0752) on average, with additional results on atom-pair TVD, joint distance MMD, geometric state occupancy, ablations, and sampling time.

Critical strengths
  1. Using forces to improve a one-step generator at training time only is a clean and practical idea: sampling stays fast (about 7 ms for 1,000 conformations) and needs no force or energy calls.
  2. The controls are well designed. Shuffling the forces among references raises TVD from 0.0457 to 0.1375, which shows that the gain really comes from correctly paired force information.
Critical weaknesses
  1. BIT does not outperform simply reusing the training data. Drawing 1,000 frames from the training set gives a lower mean TVD (0.0412) than BIT (0.0472) (Appendix C.5). So the results show that forces help the generator fit the training distribution better, but not that they help it go beyond the training data, which is the main promise of using physical information.
  2. The evaluation focuses on marginal pairwise-distance distributions. On metrics closer to the Boltzmann distribution, the gains are small or mixed: state-occupancy TVD improves only 4.7% over the no-force version and is worse than training-set resampling on 3 of 7 molecules, and ET-Flow has lower joint-distance MMD on 3 of 7 molecules. There is no evaluation of energies or free energies of the generated conformations.
  3. The baseline comparison may not be fair. BIT’s hyperparameters are tuned per molecule, while AvgFlow and ET-Flow use default configurations from codebases designed for multi-molecule datasets. Force-based methods such as PSM are not compared under the same protocol, and their published TVD (0.0430 for PSM-P) is lower than BIT’s.
  4. The experiments are limited to seven small MD17 molecules, each trained separately on one contiguous segment of 5,000 frames. It is unclear whether the method scales to larger or more flexible molecules.
  5. The adaptive mixing rule is weakly justified. It is derived only under strong assumptions (Appendix B.3), the statistic can vanish even when distributions differ (Appendix B.2), and it depends on the orientation of reference conformations because no rotational alignment is applied (Appendix A.2).
Questions for authors
  1. Can you show a setting where BIT outperforms training-set resampling, for example with fewer training frames or with training frames that cover the distribution poorly? This is where force information should help most.
  2. Can you evaluate the energy distribution of generated conformations, or compare against force-based methods such as PSM under the same protocol?
  3. How were AvgFlow and ET-Flow tuned for single-molecule training, and do they improve with per-molecule tuning comparable to BIT’s?
  4. In ethanol, Ungated (0.0898) is worse than No-force (0.0754). Can you explain when force reweighting hurts, and how sensitive BIT is to g and c?
  5. How does the rotation dependence of D(x) affect results, and would rotationally aligning the references change performance?
Reviewer_UBJ6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Boltzmann-Inspired Training (BIT), a method that leverages force information from reference conformations to improve the training of one-step molecular conformation generators. Representing molecular structures via interatomic distances, the approach employs an attraction-repulsion training framework; it dynamically reweights the attraction terms for reference conformations based on the alignment between projected reference forces and distance discrepancies. Furthermore, it utilizes a local statistic—inspired by a kernel-score identity—to adaptively blend a data-driven objective with a force-correction objective. The generator is trained via stop-gradient regression, requiring only a single forward pass during inference, with no need for additional force or energy queries. The authors trained and evaluated the model on seven molecules from the MD17 dataset. The primary empirical contribution of this work is the improved matching of distance distributions for the static conformations of small molecules.

Critical strengths

Forces at reference conformations carry local physical information that coordinates alone do not capture, so using them to train generative models is a well-motivated idea. BIT adds this force information to the training objective while still generating samples in a single step at inference. Ablations using the same architecture show that including forces improves how well the generated distance distributions match the reference, which supports the value of this approach.

Critical weaknesses

My main concern is that the paper does not clearly establish the practical need for BIT. In the current setup, each molecule already has 5,000 reference conformations with forces, and BIT is trained separately for the same molecule and conditions. It is therefore unclear what additional scientific capability BIT provides beyond directly using or resampling the available data.

The results show clear improvements over the matched No-force baseline, but do not demonstrate that these gains translate into more accurate physical observables, better coverage of important conformations, or lower overall computational cost.

One-step molecular sampling is certainly a meaningful topic, but the current validation does not yet establish a clear practical or methodological advantage beyond improved fitting on per-molecule MD17 benchmarks.

Questions for authors
  1. Given the same reference conformations and forces, what concrete benefit does BIT provide over directly using or resampling the training data, in terms of accuracy, coverage, or computational cost? If the contribution is mainly methodological, what demonstrates value beyond per-molecule MD17 distance fitting?

  2. Please consider a controlled equilibrium benchmark such as a 1D double well or 2D Müller–Brown potential, where the Boltzmann distribution is known accurately. Comparing BIT, No-force, and training-data resampling on full distribution recovery, basin populations, and free-energy differences would more directly test whether force information provides benefits beyond data reuse.

Reviewer_LgY7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Boltzmann-Inspired Training (BIT), a framework for one-step molecular generation that incorporates molecular forces into training through force-aligned attraction-repulsion targets and adaptive weighting. The proposed approach is tested on seven small MD17 molecules, where BIT improves distance-distribution matching over force-free training. The framework outperforms competitive baselines: AvgFlow and ET-Flow, on average for selected geometric metrics, while allowing conformations to be sampled far more efficiently at inference without requiring force calculations.

Critical strengths
  1. The improvement over the same generator is substantial and reasonably well-isolated. Mean pooled-distance TVD nearly halves and mean atom-pair TVD decreases by 30-40%. BIT also improves over independently-tuned ungated force reweighting on every molecule. Full-distance MMD improves over no-force on all seven molecules, so the evidence extends beyond the pooled histogram. Three training seeds, generation-seed replication, and histogram sensitivity checks also show a somewhat generalizable learning (at least to this test set).

  2. The implementation offers direct sampling with recorded forces that are only needed during the training process. The paper includes training-resampling, reference-sampling, geometric occupancy, and nearest-training-conformation controls. The paper is also clear and upfront about the adaptive rule's assumptions and representation limitations.

Critical weaknesses
  1. The methodological novelty is limited, and the proposed adaptive gate remains insufficiently validated. A closely related work, Drifting to Boltzmann, already uses force-alignment weighting and stop-gradient regression for molecular generation. The authors acknowledge the original drifting work in their literature review but this one does not appear present. Although BIT differs in its force projection and residual-based adaptive mixture, the current ablations do not isolate these contributions. Comparisons with a matched force-aligned baseline, a tuned constant-mixture control, and rotation-sensitivity experiments are needed to establish the gate's practical value.

  2. The experiments do not convincingly establish improvements in molecular ensemble generation or computational efficiency. Evaluation is limited to seven small MD17 molecules trained independently, with some evidence for occupancy improvements. Training-set resampling also achieves better pooled TVD than BIT, raising questions about improvements beyond the stored training ensemble. Importantly, structural validity and stability metrics are also not evaluated and provided, despite being reported in closely-related work. Stronger ensemble-level diagnostics, structural-validity checks, and matched quality-cost comparisons are required.

Questions for authors
  1. Does the proposed method provide a meaningful advantage over closely-related approaches? Can you compare BIT against a matched force-aligned drifting baseline and a tuned constant-mixture variant, using the same architecture and training budget? If improvements fall within experimental uncertainty, that questions the scale of the contributions.

  2. Is the adaptive gate robust to reference orientations? Can you independently rotate reference conformations and their paired forces, both at fixed generated queries and during matched training, and measure the resulting changes in the gate and sample quality? This would help clear up whether the gate provides improvements or whether it depends on arbitrary coordinate choices.

  3. Do geometric improvements translate into better molecular ensembles and generalize beyond small molecules? Can you evaluate bond validity, steric clashes, and joint torsional-population errors, including comparisons against training-set resampling and learned baselines? More importantly, the seven small MD17 molecules are insufficient to establish generalizability or scalability. Evaluation on a wider range of small molecule datasets that go beyond just 7 molecules is needed to demonstrate a functional approach.

  4. Does BIT maintain its efficiency advantage when accounting for total computational cost? Can you report training, hyperparameter-selection, and inference costs under matched settings, including AvgFlow's teacher-generated data and ET-Flow at multiple NFE budgets? Comparing joint quality and structural validity against total computational cost would establish whether BIT offers a meaningful efficiency advantage beyond the reported inference-speed comparisons.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.