acceptodds
Withdrawn from ICLR 2027

ConformerDiff-Adaptive: A Sampler-Agnostic Wrapper for Per-Molecule NFE Allocation in Conformer Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sampling from diffusion-based molecular conformer generators is expensive: models such as GeoDiff use thousands of denoising steps for every molecule, irrespective of conformational complexity. We propose ConformerDiff-Adaptive, a backbone-frozen wrapper that allocates a global number-of-function-evaluations (NFE) budget heterogeneously across a batch of molecules to maximize aggregate coverage. The wrapper has three complementary components: an exact multi-choice knapsack allocator (which beats a greedy heuristic by 19.7 points on average); a calibrated monotone heavy-atom-count prior that carries the dominant in-distribution signal; and a low-weight LightGBM residual on cheap RDKit features for molecule-specific structure beyond size. On GeoDiff's published test_data_1k benchmark with 3-seed averaged ground truth, our wrapper outperforms the strongest baseline by +23.0 to +35.9 points absolute COV-R across the compute-constrained regime (20-50% of the published NFE budget, all p < 0.001, paired bootstrap). At 80% budget and above the advantage disappears: allocation degenerates to a uniform schedule, and we show this is intrinsic rather than a limitation of our predictor. The wrapper is sampler-agnostic across three substantially different samplers (Langevin dynamics, DDIM-style generalized, and DPM-Solver-2) and backbone-agnostic across GeoDiff and AGDIFF. The pipeline modifies no backbone parameters and requires about 3 GPU-hours of one-time per-deployment calibration. Code, trained predictors and calibration labels are available at https://anonymous.4open.science/r/ConformerDiff-B5F0

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular conformer generation, diffusion models, inference-time compute allocation, sampling efficiency, multiple-choice knapsack

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_ycsiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ConformerDiff-Adaptive, an inference-time wrapper for diffusion-based molecular conformer generation that allocates a fixed global NFE budget non-uniformly across molecules. The method first estimates each molecule's expected COV-R at different NFE levels using a calibrated heavy-atom-count isotonic prior and a low-weight LightGBM predictor, then jointly assigns NFE using an exact multi-choice knapsack formulation. The main contribution is therefore to formulate conformer generation under a shared compute budget as a cross-molecule resource-allocation problem, rather than using the same number of sampling steps for every molecule.

The method is evaluated mainly on GEOM-Drugs, with additional experiments on GEOM-QM9, three different samplers, and two backbones. In the compute-constrained regime, the authors report substantial COV-R improvements over uniform and atom-count-based allocation baselines, while the advantage largely disappears once most molecules reach their coverage saturation region.

Critical strengths
  1. The problem formulation and overall presentation are clear. The paper clearly motivates why a uniform NFE schedule can be inefficient when molecules exhibit different coverage–NFE curves, and it formulates the shared-budget allocation problem in a simple, interpretable way using a multi-choice knapsack objective. The relationship between the predictor, the allocation module, and the final compute budget is easy to follow.

  2. The paper provides useful component-wise analysis. The authors not only report final performance but also analyze which parts of the method drive the gain.

Critical weaknesses
  1. The empirical validation appears limited relative to the breadth of the claimed generality. The largest gains appear in strongly compute-constrained settings with GeoDiff-style generators, while the benefit shrinks substantially in faster-sampling regimes and disappears near saturation.

  2. The formulation fixes the underlying generator and optimizes only how its NFE budget is distributed. Under an actual inference-budget constraint, however, a practitioner may instead choose or train a generator specifically optimized for low-NFE sampling. The current experiments do not establish where heterogeneous allocation lies on this broader compute-quality frontier.

  3. The technical novelty of the predictor appears limited. The method relies on standard surrogate modeling followed by a standard multi-choice knapsack solver.

  4. The allocation objective maximizes aggregate mean COV-R and can therefore improve the average by effectively abandoning difficult molecules under tight budgets. It would be useful to discuss whether this objective matches realistic conformer-generation workloads, where minimum per-molecule quality or worst-case performance may matter.

Questions for authors
  1. How well does the proposed allocation strategy generalize to more modern low-NFE conformer generators, especially flow-matching or distilled models? Because the reported gains shrink with faster samplers and disappear near saturation, additional experiments on genuinely few-step generators would clarify whether the method remains practically useful beyond the GeoDiff-style setting.

  2. Why is heterogeneous NFE allocation the appropriate optimization target under a fixed inference budget, compared with using a generator that is itself trained or designed for low-NFE sampling? Could the authors provide a matched-compute comparison between (i) a high-NFE generator with adaptive allocation, (ii) a low-NFE generator with uniform allocation, and ideally (iii) a low-NFE generator with adaptive allocation?

  3. What is the methodological novelty beyond combining a standard coverage surrogate with a standard multi-choice knapsack solver?

  4. The current objective may improve aggregate mean COV-R by effectively giving up on difficult molecules under tight budgets. This concern is directly supported by the behavior reported in Section 5.2: at 40% budget, 71 of 200 molecules are assigned to the minimum n=500n=500 rung, where their measured COV-R is exactly zero. While this is optimal for the stated aggregate objective, it may not reflect realistic deployment settings in which every molecule is expected to receive at least a minimally useful conformer set. Could the authors quantify this failure behavior across budget levels and evaluate a constrained formulation that enforces a minimum per-molecule NFE or minimum-quality requirement?

  5. For both Tables 2 and 4, the reporting compresses the budget-dependent behavior considerably. Full budget–coverage curves would make the claimed generalization substantially easier to assess. Could the authors provide more fine-grained results across the full budget range, including the explicit numerical improvement at each budget level?

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

The paper studies budget-aware sampling for diffusion-based molecular conformer generation. Instead of assigning the same number of denoising steps to every molecule, the authors propose allocating a fixed total NFE budget across molecules based on how much each is expected to benefit from additional sampling. The diffusion backbone is kept frozen, and the objective is to maximize mean COV-R under the shared budget constraint.

The method formulates the allocation as a multiple-choice knapsack problem, where each molecule selects its NFE from a 9-point grid between 500 and 5000. The problem is solved exactly using dynamic programming. To estimate the coverage at each NFE level, the authors combine isotonic regressions on heavy-atom count, fitted on 150 in-domain calibration molecules, with a LightGBM predictor trained on RDKit features.

On 50 held-out GEOM-Drugs molecules, the method improves COV-R by 23–36 percentage points over an atom-count GP baseline at 20–50% of GeoDiff's published sampling budget. The improvement disappears above roughly 70% budget, which the authors attribute to saturation of the per-molecule coverage curves. They also evaluate the approach with other samplers (DDIM-style and DPM-Solver-2), another backbone (AGDiff), and on GEOM-QM9. The ablations show that the simple atom-count prior accounts for nearly all of the performance of the full method, while the exact knapsack solver substantially outperforms the tested greedy allocation baseline.

Critical strengths
  1. Transparent reporting. The paper explicitly reports that the gains disappear at high budgets, provides multi-seed results alongside single-seed ones, and clearly distinguishes the Oracle-MCKP diagnostic from deployable methods.

  2. A simple, cheap wrapper that composes with existing samplers. It changes no backbone parameters, adds under one second of CPU time per batch, and Table 2 shows that it can be applied on top of faster samplers rather than competing with them.

Critical weaknesses

1. Limited significance. The main experiments use GeoDiff-family backbones with thousands of Langevin steps and the original GeoDiff protocol (δ=1.25\delta = 1.25 Å), while more recent conformer generators that the paper itself cites (e.g., AvgFlow, DiTMC), including few-step ones such as AvgFlow, are not evaluated, so the practical value for such models is not established. On DPM-Solver-2, the gains are smaller than with Langevin dynamics (+7.2 and +0.4 pp, Table 2). At 70% budget, Best-Uniform already matches the full-budget reference, so the gains are confined to lower-coverage operating points whose practical use is not demonstrated.

2. Weak baseline and confounded evidence. The headline gains are measured against AtomCount-MCKP, with much of the difference attributable to GP versus isotonic regression on the same feature: Ours-Prior already achieves 99.7% of the full method's performance. At 30–40% budget Best-Uniform is stronger than AtomCount-MCKP, contrary to the abstract. In Table 5, the coarse 4-point grid prevents the uniform baseline from fully using its budget, so the observed gain may come from better budget utilization through mixed NFE levels rather than from molecule-aware allocation. Only mean COV-R is reported, although at 40% budget about a third of molecules are assigned the lowest NFE level, which yields zero coverage.

3. Thin methodological contribution. In distribution, the method reduces to an isotonic regression on heavy-atom count plus a standard knapsack DP; the 1,077-feature LightGBM adds 0.2 pp. The monotone atom-count assumption is not validated, and atom count may be a poor difficulty axis (e.g., n-hexyl and phenyl substituents have the same heavy-atom count but very different flexibility).

4. Unclear cost accounting. Extrapolating Table 8's wall-clock time (19,782 s for about 200k NFE) to the stated Phase B calibration workload gives roughly 328 GPU-hours, versus the reported ~3, and no break-even analysis is given.

Questions for authors
  1. How does Ours compare with a control that randomly permutes its NFE assignments across molecules (same rungs, same total budget)? For Table 5, how does it compare with a uniform schedule that uses the full budget?
  2. Is the gap to AtomCount-MCKP due to monotonicity or to the estimator? Please show unconstrained binned means of COV-R vs. heavy-atom count, and compare with an equally regularized non-monotone smoother.
  3. How were the ~3 GPU-h (Phase B) and 7 GPU-h (Phase A) figures obtained, given Table 8's throughput? How many molecules are needed for calibration to pay off?
  4. How would reference conformers for calibration be obtained in practice? A small-scale experiment using references prepared with xTB or DFT, including their computational cost, would strengthen the practical case.
Reviewer_hSbbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper tackles the problem of sampling budget allocation for small molecule conformation, using diffusion models, with the goal of maximizing a measure of coverage within a given budget. The main idea involves two steps: fit a regressor that predicts coverage based on input molecule (size and graph), and solve a knapsack problem (exactly, with dynamic programming) to find the optimal budget allocation given their regressor.

Critical strengths
  • The method they propose is simple and efficient (once a set of generations from the generative model are available), and appears to work well for the metric of choice in the paper.
  • Relevant problem, trying to reduce inference costs for a pre-trained model.
Critical weaknesses
  • The method shows limited benefits when preserving the original quality level. The biggest gains by the method w.r.t baselines are achieved at budgets of 20-50%, where the final performance is quite a bit below the best achievable performance. At a 70% budget the "best-uniform" is already at top performance, and from then on there is no benefit of one method over another one. While reducing budget by more than 50% is desirable, it comes with a considerable performance degradation (the method has a lower degradation than baselines, but still fairly below the top performance reached at 70%).
  • I find a few trade-offs that are concerning. While table 1 at a 40% budget shows much better performance for their method over baselines, the paper also reports a significant fraction of molecules that are not "solved" (no generated conformer is close to a reference). While this is acknowledged, it is not reported in the tables for results. For this specific budget, 71 out of 200 molecules get zero hits. It is not clear how this compares to the baselines. While overall coverage is better (measured as the average fraction of references covered per molecule), leaving >30% of the molecules with zero hits is concerning. One could envision other metrics that better characterize this trade-off: in addition to average coverage, also include for how many molecules a certain threshold for the coverage is recovered.
  • Limited methodological novelty. The paper claims that the (molecule size, number of sampling steps) → coverage monotonic regressor, combined with the same knapsack DP, recovers 99.7% of the method's performance. From this, I wonder whether a similar result could be obtained by simply choosing the smaller molecules first, and choosing the smaller n such that coverage is good for them (over a certain threshold). This is a very simple baseline; while the paper presents a somewhat more complex approach (with the GBM and the knapsack problem), how much does it improve over a baseline like this?
Questions for authors

NA

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

The idea of varying the number of integration steps for different molecules is sound to me. However, multiple points make me questioning the relevance and the quality of the paper:

  1. The work from [1] already decreases the NFE to 50, so working with a budget of 1/3k is already based on outdated results.

  2. In the abstract, the authors write "On GeomDiff's published test_data_1k benchmark ... " GeomDiff [2] did not publish any benchmark, instead they used two standard benchmarks in the field, as explained in Section 4.

  3. In main results Table 1 the authors write that the main results were tested on 50 held-out molecules. What molecules are those? the regular test set of GEOM, are 1000 molecules for Drugs and QM9, right? Did the authors subsample that full set?

Overall, I think the authors should proof-read all of their papers' claims before submitting anything.

[1] Frank, Thorben, et al. "Sampling 3D Molecular Conformers with Diffusion Transformers." Advances in Neural Information Processing Systems, (2026).

[2] Xu, Minkai, et al. "Geodiff: A geometric diffusion model for molecular conformation generation." arXiv preprint arXiv:2203.02923 (2022).

Critical strengths
Critical weaknesses

see paper summary

Questions for authors

see paper summary

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

Related papers

Discussion (0)

Sign in to comment.