acceptodds
Under review as a conference paper at ICLR 2027

DRAFT-Solver: Dynamic spaRse Attention via node draFTing

Abstract

Learning surrogate models for steady-state partial differential equations (PDEs) on large unstructured meshes represents a challenging task because relevant interactions may span the entire domain, while accurate predictions often require additional computation only in localized, task-dependent regions. Dense attention captures long-range dependencies but becomes impractical on large discretizations, whereas sparse and multiscale approaches commonly prescribe where fine-scale processing should occur using geometric or physical priors. To address these limitations, we introduce DRAFT-Solver (Dynamic spaRse Attention via node draFTing), a dual-scale architecture that learns its computational allocation directly from data. DRAFT-Solver combines global information exchange through softly constructed latent tokens with fine-grained self-attention over a learned, fixed-budget subset of nodes. Both token routing and node selection are trained end-to-end without predefined surface regions or manually specified masks. Across benchmarks spanning two- and three-dimensional aerodynamics, semiconductor-device simulation, and biomolecular electrostatics, DRAFT-Solver achieves competitive or improved performance relative to graph-, neural-operator-, and transformer-based baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, PDE surrogate modeling, unstructured meshes, sparse attention, adaptive computation, transformers

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_cDmcOpenReview ↗
Rating
4(+3)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DRAFT-Solver, a neural surrogate for steady-state PDEs on unstructured meshes. Each DRAFT block combines two attention pathways:

Coarse pathway. Nodes are softly assigned to K latent tokens through an assignment matrix computed once from the initial embeddings and shared across all blocks. Self-attention is applied among the tokens, and the result is routed back to every node. Fine pathway. A scoring network selects a fixed-budget subset of M = ⌊bN⌋ nodes, using top-M hard routing with a straight-through estimator. Full self-attention is applied among the selected nodes.

Global PDE parameters are injected through FiLM. An optional pairwise ranking loss encourages the scorer to give higher scores to nodes with larger prediction errors.

The method is evaluated on AirFRANS, ShapeNet Car, CMOS semiconductor simulations and SPACIER-M. Each benchmark reports a variant with ranking supervision and one without. The paper also analyses where the fine budget is allocated (Figs. 4–5, App. B–C).

Critical strengths

Broad evaluation with a single architecture. The same design is applied to 2D and 3D aerodynamics, semiconductor-device simulation and biomolecular electrostatics. The reported numbers on AirFRANS and CMOS are strong compared with the baselines listed.

An informative and candid analysis of the learned allocation. The visualisations of the selected nodes (Fig. 4; Figs. 6–9 for twelve samples per dataset) and the cumulative-capture curves (Fig. 5) show consistent, interpretable allocation patterns, such as surface-focused selection on ShapeNet Car. The discussion of when ranking supervision helps and when it hurts (Sec. 5, App. B.3–B.4) is open about the method's limitations.

Critical weaknesses
  1. Limited novelty

The coarse pathway is essentially Transolver's physics-attention (soft slicing into tokens, token attention, routing back with the same weights; Eqs. 3–4), and the fine pathway is standard top-k selection with a straight-through estimator. I do not see a distinctive new idea beyond combining these known components.

  1. Efficiency is never measured

Although the paper is motivated by the cost of attention on large meshes (L37–38), it reports no runtime, memory or parameter count for DRAFT or any baseline, so the central efficiency claim is unsupported.

  1. Missing ablations

There is no coarse-only model and no random-selection fine pathway, so there is no evidence that the learned node selection, rather than the extra attention layer, drives the gains.

  1. Headline results are selected per dataset

The contributions (L69–72) report, for each dataset, whichever variant (with or without ranking) scores better on the test metrics, without a validation-based selection rule. Neither variant is consistently better than the baselines.

Add. In my view, the text in Figure 1 of this paper is far too small, while the font size in Figure 4 is too large. I would like to see the variation in font sizes across the figures kept within a reasonable range; this is an aspect of the paper's overall quality.

Questions for authors
  1. What are the runtime, peak memory and parameter count of DRAFT compared with the baselines?

  2. How does DRAFT compare with a coarse-only model and with random node selection in the fine pathway?

5 comments

We thank the reviewer for the review. We will first answer the questions with all the ablations and efficiency measurements requested, and then go through the raised weaknesses. Unless stated otherwise, the experiments use AirFRANS with the configuration of the paper (no ranking loss, b = 0.09, K = 32) and change only the component under study.

Coarse-only model and random node selection.

VariantVol ↓Surf ↓CL ↓ρL ↑
Transolver0.00250.00800.10260.9977
SpiderSolver0.00170.00430.07410.9988
DRAFT, coarse only (no fine pathway)0.00020.00190.01960.9998
DRAFT, random fine selection0.00020.00130.02130.9998
DRAFT, geometric fine selection (nodes closest to the airfoil)0.00050.00210.03310.9994
DRAFT, without FiLM0.00040.00360.03020.9996
DRAFT (full)0.00020.00120.01690.9998

The reviewer asked whether the gains come from the learned selection or simply from the extra attention layer. The ablation separates the two:

  • Coarse only. Removing the fine pathway increases the surface error by 58% and the CL error by 16%.
  • Random selection. Keeping the extra attention layer but selecting the nodes at random, with the same budget, increases the CL error by 26% (0.0213 vs 0.0169).
  • Geometric selection. Selecting the nodes with a hand-crafted geometric prior, the nodes closest to the airfoil, is the worst choice. Its CL error is about 2× that of the learned variant, even though this is the natural choice for aerodynamics.
  • Conclusion. The extra attention layer alone does not explain the gain: the learned allocation does.

Efficiency: runtime, peak memory and parameter count.

Setup: single NVIDIA H200, AirFRANS sample subsampled to 32,000 nodes, batch size 1, 20 warm-up and 100 timed iterations. Training time is one forward + backward + optimizer step.

ModelParamsTrain time (ms) ↓Train peak mem (GB) ↓Inference time (ms) ↓Inference peak mem (GB) ↓
Transolver3.89M57.13.6220.40.37
SpiderSolver3.99M133.211.0347.71.23
DRAFT4.01M47.14.1315.60.38
  • Model size. The parameter counts are comparable.
  • Against Transolver. DRAFT is faster in both training (1.2×) and inference (1.3×).
  • Against SpiderSolver. DRAFT is 2.8× faster in training and 3.1× faster in inference, and uses 2.7× less training memory.
  • Larger meshes. The advantage holds as the mesh grows. On the full 179k-node mesh, DRAFT remains faster than Transolver in both training (283 vs 313 ms) and inference (92 vs 110 ms), while SpiderSolver runs out of memory even on a 141 GB GPU.

We also report the effect of the fine budget b on accuracy and cost:

bSurf ↓CL ↓Train (ms)Inference (ms)
0.020.00180.022843.214.3
0.050.00110.027144.514.7
0.090.00120.016947.115.6
0.200.00150.022259.720.5
0.400.00220.0149107.537.6

All budgets outperform both baselines on all metrics. b = 0.09 gives the best accuracy–cost trade-off, and up to b = 0.20 DRAFT remains faster than Transolver.

(Continued in the next comment.)

(Continuation of our previous comment.)

Novelty.

We agree that the coarse pathway builds on soft token aggregation in the spirit of Transolver, and that top-k selection with a straight-through estimator is a known tool. Our contribution is not a new attention primitive. It is a new answer to a central question for mesh-based surrogates: where to spend node-level computation.

  • Existing solvers do not learn where to refine. Transolver treats all nodes uniformly through latent tokens. SpiderSolver prescribes the refinement regions from the geometry, through boundary-informed partitions.
  • DRAFT learns it. To our knowledge, DRAFT is the first PDE surrogate that learns, end-to-end and per sample, a fixed-budget set of nodes for additional node-level attention, on top of a global token pathway that updates every node. No surfaces, partitions or masks are specified.

The new experiments show that this design choice, and not the individual components, is what matters:

  • Learned allocation matters. At the same budget, random refinement is worse than learned refinement. Geometry-prescribed refinement is worse still (2× the CL error), even though it is the SpiderSolver-style prior and the natural choice for aerodynamics.
  • One mechanism across domains. Without any dataset-specific masks, the same mechanism learns different, physically meaningful allocations across four domains (Fig. 4): the airfoil surface, especially the leading edge, in aerodynamics, the device interfaces in CMOS, and most atoms in SPACIER-M, where the target is a global sum.
  • Lower cost. DRAFT achieves this while being faster than both Transolver and SpiderSolver, up to the full mesh.

Soft token aggregation and top-k selection are building blocks. Their combination into a learned, budgeted refinement for PDE surrogates is new, and we will make this distinction explicit in the revised paper.

Selection of the headline results.

Reporting both variants on every benchmark is intentional and not a selection procedure. Whether error-ranking supervision helps is one of the questions the paper studies. Its answer depends on the structure of the target, and it can only be observed by reporting both variants side by side.

  • No hidden selection. Every table reports both variants, so the reader always sees the result of each one, including where ranking hurts.
  • The outcome follows a consistent pattern (Section 5). Ranking helps when the metric is local and region-specific: on ShapeNet Car the surface drag depends on the surface region that ranking concentrates on. It does not help when the target aggregates outputs of different physical relevance (AirFRANS, CMOS) or is a signed global sum with cancelling contributions (the SPACIER-M solvation energy). This gives a criterion that can be applied a priori: use ranking when the downstream quantity is local, and end-to-end allocation otherwise.

We agree that the contribution statement (L69–72) mixes the two variants. In the revision, we will rewrite it to state the results of the default model explicitly and to present ranking supervision as an analysed extension, together with the criterion above.

Figure font sizes. Thank you, we will harmonize the font sizes, enlarging the text in Figure 1 and reducing it in Figure 4.

We hope that the new ablations and efficiency measurements address the reviewer's main concerns, and we would be grateful if the reviewer would consider them in the final assessment.

Follow-up

The current PDF does not yet include the changes promised in the rebuttal. As a reminder, ICLR allows authors to upload a revised PDF during the discussion period.

Reviewer_XKZiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DRAFT-Solver, a dual-scale architecture for surrogate modeling of steady-state partial differential equations (PDEs) on large unstructured meshes, aiming to balance global information exchange and fine-grained local modeling. The architecture combines coarse-grained global attention with fine-grained sparse attention. Specifically, it uses a learnable assignment mechanism to aggregate mesh nodes into a small set of latent tokens for efficient global information exchange, while a node-importance scoring mechanism selects a fixed-budget subset of nodes for additional attention to enhance fine-grained feature modeling. The method is evaluated on multiple benchmarks, including AirFRANS and ShapeNet Car, covering 2D and 3D aerodynamics, semiconductor-device simulation, and biomolecular electrostatics. Experimental results demonstrate that DRAFT-Solver can be applied across diverse PDE scenarios and achieves competitive predictive performance against existing baselines.

Critical strengths

This paper employs a learnable node-importance scoring mechanism and fixed-budget node selection to allocate additional fine-grained computation to selected nodes. Compared with refinement strategies relying on predefined geometric priors, this design provides a data-driven approach to computational resource allocation. Visualizations across different physical scenarios, including aerodynamics and semiconductor-device simulation, show that the proposed method learns different node allocation patterns. This offers a promising approach to adaptive feature modeling on complex unstructured meshes and demonstrates potential research value.

Critical weaknesses

W1. The experimental evaluation is not sufficiently comprehensive. While the proposed method combines global latent-token attention, fine-grained sparse attention, and learned node selection, the experiments mainly focus on overall predictive performance and the effects of ranking supervision. Systematic ablation studies of key components, particularly the coarse and fine pathways, are missing. This makes it difficult to assess their individual contributions to the reported performance improvements over the baselines.

W2. Although the paper focuses on efficient information exchange over large unstructured meshes and provides a theoretical complexity analysis, it lacks empirical evaluations of runtime, GPU memory consumption, and scalability across different mesh sizes. Moreover, when the node selection ratio is fixed, the fine-grained pathway retains quadratic complexity with respect to the number of nodes. In particular, on SPACIER-M, approximately 75% of the nodes receive additional fine-grained attention, raising concerns about the practical efficiency gains. Further empirical evidence is needed to substantiate the efficiency and scalability of the proposed method.

Questions for authors

Q1. The appendix provides dataset-specific architectural and ranking-allocation settings, with substantial differences in fine-attention budgets across benchmarks. Notably, the fine-attention budget on SPACIER-M is set to 75%, considerably higher than those of other benchmarks. Could the authors clarify the rationale behind these settings?

Q2. Furthermore, although the proposed method achieves competitive predictive performance across multiple benchmarks, how sensitive is its performance to the proportion of nodes selected for fine-grained attention?

4 comments

We thank the reviewer for the helpful comments, which will make the paper clearer and easier to follow. We ran new experiments for each point and will include them in the revised paper. Unless stated otherwise, the AirFRANS experiments use the configuration of the paper (no ranking loss, b = 0.09, K = 32) and change only the component under study.

W1. Ablation of the coarse and fine pathways.

VariantVol ↓Surf ↓CL ↓ρL ↑
Transolver0.00250.00800.10260.9977
SpiderSolver0.00170.00430.07410.9988
DRAFT, coarse only (no fine pathway)0.00020.00190.01960.9998
DRAFT, random fine selection0.00020.00130.02130.9998
DRAFT, geometric fine selection (nodes closest to the airfoil)0.00050.00210.03310.9994
DRAFT, without FiLM0.00040.00360.03020.9996
DRAFT (full)0.00020.00120.01690.9998
  • Fine pathway. Removing it increases the surface error by 58% and the CL error by 16%.
  • Learned selection. At the same budget, it outperforms both random selection (CL −21%) and a geometric near-wall prior (CL −49%, Surf −43%).
  • Conclusion. The gain does not come from the extra attention layer alone: where the fine computation is placed matters. A hand-crafted prior placing it near the surface is worse than the allocation learned from data.

W2. Empirical runtime, memory and efficiency.

Setup: single NVIDIA H200, AirFRANS sample subsampled to 32,000 nodes, batch size 1, 20 warm-up and 100 timed iterations. Training time is one forward + backward + optimizer step.

ModelParamsTrain time (ms) ↓Train peak mem (GB) ↓Inference time (ms) ↓Inference peak mem (GB) ↓
Transolver3.89M57.13.6220.40.37
SpiderSolver3.99M133.211.0347.71.23
DRAFT, coarse only3.43M34.43.9811.10.37
DRAFT (b = 0.09)4.01M47.14.1315.60.38

DRAFT is faster than both Transolver (1.2× in training, 1.3× in inference) and SpiderSolver (2.8× / 3.1×). It also uses 2.7× less training memory than SpiderSolver.

Scalability with mesh size. We repeated the measurement for increasing mesh sizes, subsampling the same AirFRANS sample to N nodes up to the full mesh (179k nodes), on a single NVIDIA H200 (141 GB) with batch size 1.

Training step time (ms) ↓

ModelN = 16k32k64k128k179k (full mesh)
Transolver3057107205313
SpiderSolver89133323759out of memory
DRAFT304787175283

Inference time (ms) ↓

ModelN = 16k32k64k128k179k (full mesh)
Transolver11204077110
SpiderSolver3348142333out of memory
DRAFT1016296092

Peak training memory (GB) ↓

ModelN = 16k32k64k128k179k (full mesh)
Transolver1.93.67.113.919.3
SpiderSolver4.911.028.686.8> 141
DRAFT2.14.18.115.922.1
  • Runtime. DRAFT is the fastest model at every mesh size, in both training and inference, up to the full mesh.
  • Memory. DRAFT's memory grows linearly with N and stays close to Transolver. SpiderSolver runs out of memory on the full mesh, even on a 141 GB GPU.

On the quadratic term. The fine pathway costs O(M²) in the number of selected nodes. With a fixed ratio b, M = ⌊bN⌋, so this term grows as b²N².

  • Compute. Its constant is small: b² ≈ 0.008 for b = 0.09, i.e. less than 1% of dense attention over all N nodes.
  • Memory. It is computed with fused attention kernels (FlashAttention / memory-efficient attention) that never store the M×M attention matrix, so its memory is only linear in M. This is why the fine pathway adds less than 1 GB of memory even on the full mesh.

(Continued in the next comment.)

(Continuation of our previous comment.)

Q2. Sensitivity to the fine-attention budget.

bVol ↓Surf ↓CL ↓ρL ↑Train (ms)Inference (ms)Train mem (GB)
0 (coarse only)0.00020.00190.01960.999834.411.13.98
0.020.00020.00180.02280.999843.214.34.04
0.050.00030.00110.02710.999844.514.74.08
0.090.00020.00120.01690.999847.115.64.13
0.200.00030.00150.02220.999759.720.54.27
0.400.00030.00220.01490.9998107.537.64.52
  • Robustness. Performance is robust to b: every budget outperforms SpiderSolver by 2–4× on Surf and 3–5× on CL.
  • Cost growth. Small budgets are cheap. Memory grows only mildly with b (from 3.98 GB to 4.52 GB), while runtime grows faster above b ≈ 0.2, as expected from the M² term.
  • Choice of b. b = 0.09 gives the best accuracy–cost trade-off, and up to b = 0.20 DRAFT remains faster than Transolver.

Q1. Rationale for the dataset-specific budgets, in particular 75% on SPACIER-M.

The budgets in Table 7 of the revised PDF (Table 5 in the original submission) were selected empirically per dataset. For SPACIER-M, we trained the ranking variant of the paper with smaller budgets and an otherwise identical configuration:

bR² ↑MAE (kcal/mol) ↓MAPE (%) ↓
0.100.9927211.257.51
0.350.9947213.027.81
0.750.995203.717.08

b = 0.75 is the best setting on all three metrics. We attribute this to the physics of the task, which differs from the other benchmarks:

  • Global target. The solvation free energy is a signed global sum of long-range Coulombic contributions over all atoms.
  • Distributed difficulty. The relevant interactions are not concentrated in a small region, such as a boundary layer or a device interface, but spread over the molecule. Dense node-level interaction over most atoms is therefore beneficial.

In contrast, AirFRANS and CMOS have spatially localized difficult regions, so small budgets (9–10%) are sufficient. We will add this rationale and the SPACIER-M budget sweep to the paper.

We hope these results address the reviewer's concerns and would be glad to clarify any remaining points. Additionaly we kindly ask to the reviewer if they are satisfied with the review and the new experiments to increase their final score, thanks again!

AuthorsOpenReview ↗

Revised pdf

FYI we did upload a revised pdf with all the requested analysis highlighted in blue. You can also see the general comment that summarizes it, thanks in advance!

Thank you again for your review, since the deadline is approaching could you please give us a feedback on the experiments addressed! I think they really improved the readibility and the performance of the model and its piece thanks to the ablations proposed!

You can also read them already implemented in the paper pdf.

Reviewer_u6F6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DRAFT-Solver, a dual-scale attention architecture for learning PDE surrogate models on large unstructured meshes. The proposed model combines a coarse pathway, which aggregates mesh nodes into latent tokens for global information exchange, with a fine pathway, which applies additional self-attention to a learned subset of nodes.

Critical strengths
  1. The motivation is clear and practically relevant.
  2. The empirical results are encouraging across different physical domains.
  3. The analysis of learned node allocation is informative.
Critical weaknesses
  1. I would like to better understand the contribution of each architectural component.
  2. The computational efficiency claims need stronger empirical support.
Questions for authors
  1. Could the authors provide a component-wise ablation of coarse attention, fine attention, learned node selection, and FiLM conditioning?
  2. How does learned top-k selection compare with random selection and geometry-based selection when the number of fine-attention nodes is fixed?
  3. What are the actual training-time, inference-time, and peak-memory costs of DRAFT compared with Transolver and SpiderSolver?
3 comments

We thank the reviewer for the positive assessment and for the concrete suggestions. We wanted to use different dataset in terms of geometry (2D vs 3D) but also in terms of equations so that we could test the proposed methodology on various field, e.g. navier stockes, semiconductor simulations and biology.

We think that all the comments of the reviewer are relevant so we decided to run additional experiments and test each component of the proposed architecture, showing the gain of using everything together. All ablations below use AirFRANS as main testbed.

Q1. Component-wise ablation (coarse, fine, learned selection, FiLM).

VariantVol ↓Surf ↓CL ↓ρL ↑
Transolver0.00250.00800.10260.9977
SpiderSolver0.00170.00430.07410.9988
DRAFT, coarse only (no fine pathway)0.00020.00190.01960.9998
DRAFT, random fine selection0.00020.00130.02130.9998
DRAFT, geometric fine selection0.00050.00210.03310.9994
DRAFT, without FiLM0.00040.00360.03020.9996
DRAFT (full)0.00020.00120.01690.9998
  • Fine pathway. Removing it (coarse only) increases the surface error by 58% (0.0019 vs 0.0012) and the CL error by 16%. The coarse token pathway captures the global field well, but node-level attention is needed for the surface quantities.
  • Learned selection. With the same budget, random selection reaches a similar Surf error but a 26% higher CL error (0.0213 vs 0.0169). Lift is driven by a few critical surface regions, in particular the leading edge and the forebody. Learned selection consistently refines these regions, whereas random selection spends most of the budget on the far field, where the flow is nearly uniform.
  • FiLM conditioning. It is the most important component. Without it, the Surf error triples (0.0036) and the CL error increases by 79%. The global flow variables (inlet velocity and angle of attack) determine the whole flow regime and in particular the lift. FiLM conditions every block on these global variables.
  • Overall. Every ablated variant still outperforms Transolver and SpiderSolver on all metrics. Each component contributes, and the full model is best or tied on every metric.

Q2. Learned top-k vs. random and geometry-based selection at a fixed budget.

The rows above compare three selection rules at the same budget (b = 0.09, about 9% of the nodes):

  • Random: a new uniformly random subset of nodes at every training step.
  • Geometric: the nodes closest to the airfoil (smallest signed distance). This is a hand-crafted prior that places the fine computation near the surface.
  • Learned: the scoring network of DRAFT.

Learned selection gives the best results. The geometric prior is clearly the worst of the three (CL 0.0331, about 2× the learned variant), even though focusing on the near-wall region looks like the natural choice for aerodynamics. As shown in Fig. 4, the learned allocation also covers the wake and the regions of strong pressure gradients, which a fixed distance criterion misses. This result directly supports the motivation of DRAFT: deciding where to refine from data is better than prescribing it.

Q3. Training time, inference time and peak memory vs. Transolver and SpiderSolver.

Setup: single NVIDIA H200, AirFRANS sample subsampled to 32,000 nodes (the training protocol), batch size 1, 20 warm-up and 100 timed iterations. Training time is one forward + backward + optimizer step.

ModelParamsTrain time (ms) ↓Train peak mem (GB) ↓Inference time (ms) ↓Inference peak mem (GB) ↓
Transolver3.89M57.13.6220.40.37
SpiderSolver3.99M133.211.0347.71.23
DRAFT4.01M47.14.1315.60.38
  • Model size. DRAFT has a comparable number of parameters.
  • Against Transolver. DRAFT is 1.2× faster in training and 1.3× faster in inference. Peak memory is similar: identical at inference, +0.5 GB in training.
  • Against SpiderSolver. DRAFT is 2.8× faster in training and 3.1× faster in inference, and uses 2.7× less training memory.

The fine pathway itself is cheap: at b = 0.09 it adds only 12.7 ms per training step over the coarse-only model (34.4 ms).

We also varied the fine budget b on AirFRANS:

bVol ↓Surf ↓CL ↓ρL ↑Train (ms)Inference (ms)
0.050.00030.00110.02710.999844.514.7
0.090.00020.00120.01690.999847.115.6
0.200.00030.00150.02220.999759.720.5
0.400.00030.00220.01490.9998107.537.6

Every budget outperforms SpiderSolver by 2–4× on Surf and 3–5× on CL.

We hope these additions address the reviewer's concerns and hope they can increase their score. Thanks again for the valuable comments and we would be happy to provide further details.

AuthorsOpenReview ↗

Upload revised pdf

FYI we did upload a revised pdf with all the requested analysis highlighted in blue. You can also see the general comment that summarizes it, thanks in advance!

Thank you again for your review, since the deadline is approaching could you please give us a feedback on the experiments addressed! I think they really improved the readibility and the performance of the model and its piece thanks to the ablations proposed!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revised manuscript

We thank all reviewers for their constructive feedback. We have uploaded a revised PDF. All new or modified text is highlighted in blue. The main changes are:

  1. Component ablation (new Sec. 5.1, Table 5; details in Appendix C.1). On AirFRANS, we compare DRAFT with a coarse-only model, random fine selection, geometric (near-wall) fine selection, and a variant without FiLM, all at the same budget. Learned selection outperforms both random and geometry-based selection. Every component contributes, and all variants still outperform Transolver and SpiderSolver. (Reviewers u6F6, XKZi, cDmc)

  2. Measured computational cost (new Sec. 5.1, Table 6; details in Appendix C.3). We report training-step time, inference time, peak memory and parameter count against Transolver and SpiderSolver. DRAFT is the fastest of the three models in both training and inference. (Reviewers u6F6, XKZi, cDmc)

  3. Scalability with mesh size (Appendix C.3, Table 10). We measure cost from 16k nodes up to the full 179k-node AirFRANS mesh. DRAFT remains the fastest model at every size, its memory grows linearly with N, and SpiderSolver runs out of memory on the full mesh. Sec. 5.1 now also discusses the O(M²) term of the fine pathway and the option of an absolute budget, which makes the cost linear in N. (Reviewer XKZi)

  4. Sensitivity to the fine-attention budget (Sec. 5.1; Appendix C.2, Tables 8–9).

    • A budget sweep on AirFRANS (b = 0.02–0.40) shows that performance is robust to b.
    • A sweep on SPACIER-M (b = 0.10, 0.35, 0.75) shows that b = 0.75 is the most accurate setting, which motivates its choice (Sec. 5, SPACIER-M paragraph). (Reviewers XKZi, u6F6)
  5. Contributions and novelty (Sec. 1).

    • We clarify that DRAFT's contribution is learning where to spend node-level computation, as opposed to uniform token processing (Transolver) or geometry-prescribed refinement (SpiderSolver).
    • The contribution list now states the results of the default model (no ranking loss) explicitly and presents ranking supervision as an analysed extension.
    • A new contribution summarizes the ablations and measured efficiency. (Reviewer cDmc)

The title, the method and all previously reported results are unchanged. All new experiments use the configurations and training protocol of the original submission.

Thanks again for your comments and feedback, I hope the revised analysis address your concerns, if it does it would be nice if you can take into account into the final score revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.