acceptodds
Under review as a conference paper at ICLR 2027

MeanFlux: Few-Step Crystal Generation for Accelerated Materials Discovery

Abstract

Current generative models for materials discovery produce high-quality candidates but suffer from massive sampling bottlenecks, often requiring dozens to thousands of network function evaluations (NFEs) per structure. We introduce MeanFlux, a diffusion transformer that accelerates generation by learning the average velocity across intervals of a generative trajectory. Operating in the per-atom latent space of a frozen autoencoder, MeanFlux generates viable crystals in just 1 to 16 steps. On the MP-20 benchmark, a 130M-parameter model matches the state-of-the-art yield of metastable, unique, and novel candidates while using one-sixth the inference compute. This efficiency extends to crystal structure prediction, where MeanFlux requires substantially less sampling compute than prior methods and scales predictably with model size. On LeMat-GenBench, MeanFlux produces an order of magnitude more viable crystals per compute unit than comparably pretrained models. With negative-aware fine-tuning (NFT) on a crystal reward, the MP-20 model reaches 22.5% relaxed LeMat-GenBench mSUN, twice the conditional base model, at the same sampling cost. These results demonstrate that few-step generation preserves structural quality while unlocking the extreme throughput required for large-scale materials screening.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative modeling, flow matching, crystal generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_7g3fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed MeanFlux, which is used for crystal generation and crystal structure prediction. In this work the author uses a frozen VAE to encode the crystal into a fixed latent representation and then applies a DiT transformer to denoise the latent representation using mean flow, and then finally it applies RL fine-tuning. Experimental results on MP20 are promising, and the author explores an interesting idea for crystal generation.

Critical strengths

In this work, few-step sampling helps RL fine-tuning cheaply. The main reason is each rollout is cheap, and NFT fine-tuning is practical. It roughly doubles relaxed LeMat-GenBench mSUN at the same sampling cost. Experimental results are promising on MP-20 from Table 1 for de-novo material generation. Also this paper shows that using their model crystals can be generated faster than existing baseline models like Chemeleon2.

Critical weaknesses

This paper compares only against diffusion and flow-based models in Table 1 and Figure 1 and omits recent LLM-based generative models; these are Crystal-text-LLM [1], FlowLLM [2], CrysLLMGen [3], and PLaID++ [4]. These methods obtain strong stability and novelty on MP-20, and PLaID++ is the direct reward-based fine-tuning, which makes it directly relevant to this paper. Since this paper claims the best mSUN at a given sampling cost, for fair comparison, these models should be placed on the compute-quality frontier (Figures 1 and 2) under the same relaxation protocol. Without them, it is unclear whether MeanFlux's advantage holds against the strongest current generators or only against diffusion-based baselines.

The main weakness of this work is that it lacks methodological novelty. This work uses existing components from prior work. It uses ADiT and Chemeleon2 for latent space design; the generative objective is MeanFlow, and reward fine-tuning uses NFT taken from MeanFlowNFT. Thus, this paper's contribution primarily applies existing techniques to the generation of crystals. The paper should clarify which design choices are specific to crystals (e.g. the reward design) and why they are non-trivial.

[1] https://arxiv.org/abs/2402.04379 [2] https://arxiv.org/abs/2410.23405 [3] https://arxiv.org/abs/2510.23040 [4] https://arxiv.org/abs/2509.07150

Questions for authors

Could the authors compare MeanFlux with more recent LLM-based crystal generators like PLaID++, FlowLLM, Crystal-text-LLM, and CrysLLMGen? Could the authors include the stability metric in the Table 1 for all baselines and MeanFlux? Could the authors include a methodology diagram or any pseudocode for the better understanding of their proposed framework? Otherwise, it is hard to follow the workflow.

4 comments

Thank you for your review, for the four additional references, and for noting that few-step sampling makes our reward fine-tuning cheap. We have added the comparison where weights were available, a Stable column and a pipeline figure.

W1/Q1. LLM-based generators. Of the four models, only PLaID++ has released weights, so we ran it with the authors' code (Qwen2.5-7B with the DPO adapter, temperature 1.3). Since PLaID++ is preference-aligned for stability, we compare it with MF-NFT. We scored it on relaxed LeMat-GenBench and added it to Table 11 (Appendix B.13), and to the new end-to-end comparison in Figure 12 (Appendix B.14). Values are % of 1,024 draws.

ModelSUNmSUNNovelUniqueCost per structure
Crystal-text-LLM 70B0.53.921.461.6~1,600× (est.)
PLaID++2.811.027.186.4152×
MF-NFT11.441.284.076.31×

MF-NFT reaches four times the SUN rate of PLaID++ at about 1/152 of its cost per structure. The gap remains once relaxation is included, where MF-NFT needs 25 TFLOPs per SUN crystal compared with 178 for PLaID++ (Table 4, Figure 12). Most PLaID++ samples are not novel against LeMat-Bulk (27% novel, compared with 84% for MF-NFT).

We also scored 1,024 of the LLaMA-2 70B samples that the Crystal-text-LLM authors published, under the same protocol. It reaches 0.5% SUN and 3.9% mSUN, far lower than our method.

W2. Novelty. Please see the general response. In short, we made three design choices specific to crystals. First, crystals have different numbers of atoms, and training will diverge unless we cap the JVP correction for each crystal (Appendix B.9). Second, our reward multiplies the stability terms by validity, uniqueness and novelty, so a crystal is rewarded only if it is new and stable at once, whereas using an additive counterpart rewards a known low-energy crystal. With this reward, NFT raises SUN 7.8 times over its starting model, compared with 1.7 times for GRPO on Chemeleon2 (Table 11). Third, for structure prediction we retrain on a geometry-only latent (from a new VAE) with the known species attached to each atom, which beats Crystalite on MP-20 at about 90 times less compute (Section 3.5).

Q2. Stability. We added a Stable column to Table 1. MeanFlux at eight steps has the highest SUN and the second-highest stable rate (0.160). Only ADiT has a higher stable rate (0.170), but it is also the least novel model.

Q3. Pipeline. We added Figure 6 (Appendix B.1), which now thoroughly shows the training, sampling and fine-tuning stages, each with its pseudocode. We hope this adds more clarity.

Please let us know if you have further questions. We are happy to address them!

Hey Reviewer 7g3f,

As the rebuttal period is drawing to an end, would you please let us know if we have addressed your concerns in our previous response, where we have benchmarked additional baselines, clarified our contribution, and crafted a methodology chart for easier understanding?

Best, Paper Authors

Reviewer_7g3fOpenReview ↗

Comment on Revision

Thank you for the revisions. While the revision has improved clarity, my overall assessment remains unchanged due to the following concerns:

Comparing against pre-released Crystal-text-LLM samples does not constitute a controlled evaluation under identical budgets.

The crystal-specific design choices like multiplicative reward in my opinion is a practical engineering adaptations and not something fundamental advancement.

It is still not clear whether the performance gains are reward, objective or interaction.

Thank you for reading our revision so closely and for the follow-up. We address your three points below.

A controlled comparison with Crystal-text-LLM. Since the model weights are not public, we now directly ran the authors' own stability protocol from their released code (M3GNet relaxation against the Materials Project hull) on both models, through one pipeline with the same validity filter and relaxation settings, on 1,024 random samples each.

ModelMetastable (M3GNet)Below hull (M3GNet)
Crystal-text-LLM 70B45.9%8.4%
MF-NFT78.5%32.3%

Under the very same protocol, MF-NFT reaches 78.5% metastable, about 33 points higher. A third of its samples fall below the hull, compared with 8% of theirs, and each one costs roughly 1/1,600 as much to generate. Their protocol does not require novelty or uniqueness. Once both are required, as on LeMat-GenBench, the gap widens to 11.4% SUN against 0.5%. Their DFT column needs VASP, which we could not run within this short period, so we report the M3GNet column here.

Reward, objective, or both. Thanks for asking this directly. Both really mattered in our runs, which we demonstrated on their own.

  • Objective. Before NFT, we fine-tuned the same model with GRPO (the objective Chemeleon2 uses) in eleven variants. The best of them gained only 4 points of mSUN on the MP-20 suite, while NFT gained 24. GRPO needs a stochastic sampling trajectory, and adding noise to an 8-step sampler pushes samples off the data manifold. NFT has no such requirement, so that it suits few-step models naturally.
ObjectivemSUN beforemSUN afterSUN beforeSUN after
GRPO (11 attempted variants)0.3510.327–0.3910.0870.070–0.100
NFT (MF-NFT)0.3350.5750.0720.141
  • Reward. With NFT fixed, our improved recipe raised relaxed mSUN from 22.5% to 41.2%, and the diversity weight alone adds 9 percentage (Table 2). The gain also holds up well. Every setting in our sensitivity analysis raises mSUN at least 2.9 times over the starting model (Appendix B.3, Table 2), and even with no LeMat-GenBench metrics in the reward (i.e. only mp20), MF-NFT still reaches 6.2% SUN.

We tried and finally adopted the effective objective with careful reward design to achieve the frontier performance. Similar rewards gave little under GRPO and paid off only once the objective fit a few-step sampler.

On the design choices. We understand the view that some of these are engineering adaptations, and we are not claiming a new learning principle. Still, each of them came out of extensive experiments and turned out to be necessary. Training does not even converge without the per-crystal cap (Appendix B.9), and the reward and objective above account for most of the gain. Beyond the method itself, the paper also changes how generators are compared. Rather than ranking models by hit rate alone, we measure how many SUN crystals each one yields for a fixed compute budget, which is what a screening pipeline actually is paying for. We re-score every baseline under one protocol, measure its cost on the same hardware, and place all of them on this discovery-efficiency frontier, both for generation alone and with relaxation included (Figures 1 and 12). On this frontier, MF-NFT yields the most SUN crystals per unit of compute among all models we evaluated, 9 times as many as Chemeleon2 with RL from generation alone and 2.5 times as many with relaxation. We believe this result, together with a frontier that future models can be placed on directly, is a useful contribution to the community.

Thank you again for your time and for helping us strengthen the paper.

Reviewer_qgfiOpenReview ↗
Rating
3(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

MeanFlux introduces a few-step crystal generation framework that adapts the average-velocity objective of MeanFlow to a frozen per-atom VAE latent space, enabling crystal generation in only 1–16 network function evaluations. It substantially improves the quality–compute trade-off on MP-20, extends the approach to composition-conditioned crystal structure prediction, studies scaling with model and dataset size, and demonstrates that negative-aware reward fine-tuning can further increase the yield of stable, unique, and novel candidates under a fixed sampling budget.

Critical strengths

Outstanding sampling efficiency. MeanFlux generates competitive crystal structures in only 1–16 network evaluations and matches Chemeleon2’s MP-20 mSUN performance at roughly one-sixth of the sampling cost. Broad and systematic evaluation. The paper covers de novo generation, crystal structure prediction, out-of-distribution evaluation, scaling laws, reward fine-tuning, and several datasets. Careful experimental methodology. Most baselines are re-evaluated under a common protocol, confidence intervals are reported, and the central average-velocity objective is supported by a meaningful ablation.

Critical weaknesses

Limited first-principles validation. The main stability results rely on machine-learning interatomic potentials rather than large-scale DFT calculations. Potential benchmark adaptation. Reward fine-tuning uses the LeMat-GenBench potentials and novelty reference, weakening the independence of the corresponding evaluation. Partially estimated compute comparisons. The inference costs of several baselines are calculated from architectural assumptions instead of direct wall-clock measurements. Some conclusions rely on extrapolation. In particular, the projected 2–5B-parameter model needed to match Crystalite on MPTS-52 is not empirically trained or validated.

Questions for authors

Same as weakness

3 comments

Thank you for your careful review, and for recognizing both the sampling efficiency and our faithful evaluation. We address each of your four concerns below.

W1. DFT. We agree that machine-learned potentials cannot fully replace first-principles validation. Our stability results follow the relaxed LeMat-GenBench protocol that launched recently and is accepted as one of the standard generative crystal evaluation protocols, in which we relax each structure with ORB, MACE and UMA and score it against the benchmark's ensemble hull (Appendix B.13). For the final version, we will add an MP-compatible VASP check (PBE/PBE+U with MP2020 corrections) on 50 to 100 randomly selected SUN candidates per model, with Chemeleon2 as a reference.

W2. Benchmark adaptation. To test this, we trained a variant whose reward uses no benchmark component. It uses MACE-MP-0 energies, the Materials Project hull, novelty against MP-20 and our own validity check, both for training and for checkpoint selection. This variant still reaches 24.9% relaxed mSUN and 6.2% SUN (Table 11), above the submitted MF-NFT in mSUN (22.5%) and above Chemeleon2 + RL in SUN (4.6%). A second sampling seed gives 26.2% mSUN and 6.2% SUN (Appendix B.3).

W3. Compute. Thanks for raising this. We now measure the cost of every model with public weights on one A100 (Appendix B.4, Table 3), and Table 1 and Figure 1 use these updated measurements (TFLOPs per 1,024 structures).

ModelMeasuredSubmittedms per structure
MeanFlux15.515.52.1
Chemeleon294.394.313.3
Crystalite33087055
DiffCSP4,2111,200567
MatterGen56,42917,0006,387

W4. The 2–5B projection. This is an empirical study including some projections for informing compute allocation on model and data, as is common in the scaling-law literature due to the prohibitive costs of actually training one. With the stated sizes that we had the capacity to train, our models performed very well while saving ≥ 90x compute. In a leave-one-out test, a fit to the smaller models predicts our largest model to within 1pp (Appendix B.5). Varying the ceiling broadly from 0.5 to 1.0 gives the 2–5B range as shaded in Figure 4b.

Please let us know if you have further questions. We are happy to address them!

Hey Reviewer qgfi,

As the rebuttal period is drawing to an end, it would be great if you could let us know if we have addressed your concerns in our previous response, where we have measured FLOPs more carefully, devised a reward function with only MACE energies, MP hull and novelty reference, and an explanation of the extrapolated performances of larger models on CSP task.

Best, Paper Authors

Thank you I keep the positie reviews.

Reviewer_YTK9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes MeanFlux, a few-step generator for crystals. It applies MeanFlow to a DiT that works in the per-atom latent space of a frozen VAE, following the ADiT/Chemeleon2 setup. It also uses crystal-system conditioning with CFG distilled into the training target, a norm cap on the JVP correction, and a flow-ratio curriculum. The curriculum starts with pure flow matching and moves to the MeanFlow target.

The main claim is about efficiency. On MP-20, a 130M model at 8 NFE matches Chemeleon2's mSUN (0.384 vs 0.397) at about one-sixth of the inference FLOPs.

Other results:

  • CSP: a composition-conditioned variant beats Crystalite on MP-20 (69.1% vs 66.1%) at about 90× less compute. It is behind Crystalite on MPTS-52 (25.0% vs 31.7%).
  • Scaling: model-size curves and an isoFLOP sweep.
  • Reward fine-tuning: an off-the-shelf MeanFlowNFT step on a hand-built crystal reward. It doubles relaxed LeMat-GenBench mSUN, to 22.5%.
Critical strengths
  • Clear motivation. Large screening runs use 10⁵–10⁹ candidates, so cheaper sampling can mean more discoveries per budget.
  • Strong few-step results. At 8 NFE, MP-20 quality matches a 50-step baseline (Table 1). The CSP result on MP-20 is good at a much lower cost.
  • Fair evaluation effort. The authors re-score every baseline under one protocol and report 95% bootstrap intervals.
  • Interesting scaling study. The model-size ladder (1.3M–307M) and the isoFLOP sweep.
  • The flow-ratio ablation (Fig. 5) shows that the average-velocity term matters at 8 steps.
Critical weaknesses
  1. Limited technical novelty. Every main component comes from prior work:
  • the latent VAE + DiT setup (ADiT, Chemeleon2)
  • MeanFlow
  • CFG distilled into the target (from MeanFlow itself)
  • interval guidance (Kynkäänniemi et al.)
  • logit-normal time sampling (Esser et al.)
  • NFT/MeanFlowNFT, used "without modification".

The paper reads as "MeanFlow applied to crystals". It does not name a problem specific to crystals that needs a new method.

  1. Prior work is presented as method. Sec. 2.1–2.2 describe the VAE, the MeanFlow DiT, guidance distillation, time sampling and sampling as if they were new. They should move to a background section. The VAE follows ADiT/Chemeleon2, and one variant reuses Chemeleon2's released VAE. The paper should say clearly what was built here and what was reused.

  2. The few parts that may be new get too little attention.

  • The JVP norm cap. The text says that without it training "diverges in the first few hundred steps". No table or ablation supports this.
  • The flow-ratio curriculum. Only its end value is ablated, not the schedule itself.
  1. The "fixed budget" metric favors cheap samplers by design. Figure 1 and Sec. 3.4 count SUN crystals per 100 TFLOPs of generation compute only. Results at several budgets, or total pipeline cost per SUN crystal, would make the claim convincing.

  2. The significance of the reward fine-tuning is unclear. NFT/MeanFlowNFT is used without modification. MF-NFT reaches 22.5% relaxed mSUN, against 44.0% for Chemeleon2+RL (Table 4), and the paper itself says this comparison is not controlled (App. B.6).

  3. Presentation.

  • The abstract and contributions repeat the efficiency claim several times.
  • Figure 1 is dense. It relies on "relaxed LeMat-GenBench", a protocol that is only explained in the appendix.
  • The caption "under one sampling budget" is confusing, because the models run at very different costs.
  1. Baselines run only at their default step counts (e.g. Chemeleon2 at 50, MatterGen at 2000), while MeanFlux is swept over 1–16 NFE. Without a step sweep for baselines (e.g. Chemeleon2 with DDIM at 4/8/16 steps), it is unclear whether MeanFlux wins at equal cost or the baselines are simply run at expensive settings.
Questions for authors
  1. Apart from applying MeanFlow to crystals, what is the technical contribution? Which crystal-specific problem does your method solve that plain MeanFlow does not?
  2. What is the VAE's role? Did you design a new autoencoder, or reuse ADiT/Chemeleon2's? Which VAE do Table 1 and Figure 2 use?
  3. Have you tried MeanFlow directly in coordinate space, as JiT and pixel-space MeanFlow do for images? This would show whether the VAE is needed for few-step quality.
  4. Can you compare against another few-step method at equal NFE?
  5. Please add an ablation of the JVP cap κ: no cap, and several values of κ. Also add a schedule ablation for the flow-ratio curriculum.
  6. mSUN rises a lot from 1 to 4 NFE (Fig. 3). Could a loss term or a crystal-specific change raise 1-2 step quality?
  7. What is new in the crystal reward (Eq. 6), compared with LeMat-GenBench criteria and Chemeleon2's reward? How sensitive are the results to its hand-set weights?
  8. If NFT is applied to other baselines, does the gain hold?
5 comments

Thank you for the detailed reviews and for acknowledging the strengths and faithfulness of our results! Below are our responses to most questions you have raised.

W1/Q1. Our contribution. Please see the general response. Like ADiT and Chemeleon2, MeanFlux builds on established components, and our contribution is making them work well for few-step crystal generation. We made some specific adaptation to fit the crystal generation task. Because crystals have different numbers of atoms, we cap the norm of the JVP correction per crystal (Section 2.2). Without the cap, training diverges (training error 55 compared with 1.3 at step 1,500, Appendix B.9). The second is the reward (Eq. 6). It scores stability with two smooth terms, one at the metastability threshold and one just below the hull, and multiplies both by validity, uniqueness and novelty, so a crystal receives the main reward only if it is new and stable at once. With this reward, NFT raises novelty from 29.0% to 84.0% and SUN from 1.5% to 11.4% together (Table 11), and the gain holds across seven recipe settings, three seeds and a variant with no benchmark component (Q7). The resulting few-step sampler matches Chemeleon2's mSUN on MP-20 at one sixth of its compute (Section 3.2), which in turn makes reward fine-tuning very cheap and tunable (W6).

W2/Q2. VAE. All models except the binary-code variant (bcb) reuse Chemeleon2's released VAE without changes, as Section 2 now clarifies. Table 1 and Figure 2 use this VAE, apart from the curves labelled bcb. The bcb autoencoder is our original design (Section 2.1) and a meaningful alternative. It halves the decoder cost (0.265 compared with 0.530 TFLOPs per 1,024 structures) at similar quality (Figure 2, Table 11).

W3/Q5. Cap and schedule. Thank you for pointing out these gaps. We have now ablated both on DiT-M at eight steps (Appendix B.9, Table 7).

SettingmSUNSUN
No capdivergesdiverges
κ = 0.10.3010.056
κ = 1 (default)0.3190.069
κ = 100.2850.044
No anneal0.3380.065
30k-step anneal0.3180.067

The cap is shown to be necessary, and the resulting values from 0.1 to 1 are within error. The anneal schedule had no measurable effect up to the scale we have tested, so we no longer claim it as necessary in our updated manuscript.

Q3. Coordinate space. Chemeleon1 learns to generate directly in coordinate space. So we ran it at reduced step counts as a reference for this question. Also, Chemeleon1 directly precedes Chemeleon2, whose VAE we reuse, which makes it the closest coordinate-space counterpart of our latent.

ModelEvaluationsTFLOPsSUNmSUN
Chemeleon1480.0130.069
Chemeleon18170.0110.087
Chemeleon116340.0210.173
Chemeleon132670.0330.273
Chemeleon1 (released)5121,0780.0360.334
MeanFlux815.50.0890.384

At 8 evaluations Chemeleon1 reaches 0.087 mSUN, while MeanFlux reaches 0.384 at a similar cost.

Q4. Equal NFE. We retrained the 130M model with flow matching (ρ = 1) using the same recipe and evaluated both models in the same run (mSUN, Appendix B.9, Table 7).

NFE24816
MeanFlow0.1840.3080.3460.374
Flow matching0.1760.2880.3190.337

MeanFlow is better at every NFE, and the difference is significant when pooled over the four step counts (p ≈ 0.03). We also compare MeanFlux with other baselines like Chemeleon2 and Crystalite at matched cost in W8 below.

W5. Budget metric. We agree that generation compute alone leaves out part of the pipeline, so we now also report the total cost per SUN crystal, as you suggested (Appendix B.4, Table 4). Including relaxation, MF-NFT needs 25 TFLOPs per SUN crystal, compared with 61 for Chemeleon2 + RL.

Continuing on the next thread.

Contining from the previous thread.

W6. Fine-tuning. With an improved NFT recipe, MF-NFT now reaches 11.4% relaxed SUN, compared with 4.6% for Chemeleon2 + RL, at one sixth of the sampling cost (Table 11). Its mSUN has nearly doubled, from 22.5% to 41.2% (39.7 - 41.2% over three seeds). Regarding the comparison between the two RL-ed models, since the two differ in sampler, training budget, and reward (Appendix B.7), we report the relative gain over each base model as the main comparison: NFT raises SUN 7.8 times over its starting model, while GRPO raises Chemeleon2's SUN 1.7 times only.

W7. Presentation. Thank you for these suggestions. We added a pipeline diagram with pseudocode (Appendix B.1, Figure 6), and we will deduplicate the claims from the abstract and contributions in the final manuscript.

W8. Fewer-step baselines. With Chemeleon2 at 4 to 25 DDIM steps and Crystalite at 7 to 31 steps (Figure 2, Appendix B.8, Table 6). MeanFlux has the highest mSUN in every cost band, and it still wins at equal cost settings.

TFLOPsMeanFluxChemeleon2Crystalite
~80.3450.1860.009
~160.3840.2940.226
~310.3850.3370.319

Q6. One and two steps. Most one- and two-step empirical failures are geometric. At two steps 76% of samples fail the interatomic-distance check, compared with 26% at sixteen steps, which accounts for 0.491 of the 0.496 gap in validity (Table 9). A distance-aware loss (which we applied in NFT) addresses this directly. At two steps MF-NFT reaches 0.426 mSUN, twice the starting point (0.213).

Q7. Reward. We build our reward from the public tools LeMat-GenBench scores with, which are hull energies from three potentials (ORB, MACE and UMA) and novelty against LeMat-Bulk, similar to what Chemeleon2's RL did with its MP-20 evaluation metrics. What differs is how the terms are combined. We multiply the stability terms by validity, uniqueness and novelty, whereas Chemeleon2 adds separate creativity, energy and diversity terms (Section 2.3).

Our results are robust to the hand-set weights. In our sensitivity analysis we varied the diversity weight, the KL weight, the learning rate and the number of iterations (Appendix B.3, Table 2). Across the seven settings, relaxed mSUN ranges from 293 to 422 and SUN from 61 to 157 out of 1,024, and all settings improve on the starting model by at least 2.9 times in mSUN and 4 times in SUN. The result also does not depend on the benchmark's tools. When we replace every LeMat-GenBench component with Materials Project ones (MACE-MP-0 energies, the MP hull and novelty against MP-20), MF-NFT still reaches 6.2% SUN, above Chemeleon2 + RL (Table 11).

Q8. NFT on other models. We have not applied NFT to other generators but are happy to do so given time. The NFT objective places no constraint on the sampler, so it can be applied to them. It is particularly inexpensive for MeanFlux, as one MeanFlux rollout costs one sixth of a Chemeleon2 rollout.

Please let us know if you have further questions. We are happy to address them!

Reviewer_YTK9OpenReview ↗

Response to rebuttal: soundness addressed, positioning concerns remain

Thank you for the detailed response and the substantial new experiments. The JVP-cap ablation, the equal-NFE comparison with flow matching, the step sweeps for Chemeleon2 and Crystalite, and the end-to-end cost with relaxation address my soundness concerns. I have raised my soundness score accordingly.

Several points remain open:

  1. Positioning against few-step methods. The paper's central claim concerns few-step generation. The new comparisons against flow matching and against multi-step baselines at fewer steps are useful, but neither is a few-step method (W8 and Q4). It remains unclear why MeanFlow was chosen over other few-step approaches, such as consistency models, shortcut models, distillation of Chemeleon2, or drifting. Even one such baseline on the same backbone and latent space would show whether the gains come from MeanFlow specifically or from few-step training in general.

  2. Crystal-specific contribution. The rebuttal attributes the need for the JVP cap to the varying number of atoms, but this cause is not tested. Does training also diverge at a fixed atom count? If it does, the cap is a generic stabilization rather than a crystal-specific adaptation.

  3. Prior work still presented as method. The introduction and Section 2.2 still present guidance distillation, logit-normal time sampling and the flow-ratio anneal as contributions. The anneal is also still motivated in the method, although the new ablation shows no benefit (0.338 without the anneal vs 0.319 with it). I would ask the authors to move these components to a background section in the revision, and to remove or justify the anneal.

Further suggestion (optional): since the main contribution is system and application level, it could be well presented with a figure in the main body.

I keep my rating for now. I would raise it if the authors add one few-step baseline on the same backbone and latent space, and clearly mark the adopted components as prior work in the revised PDF.

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

The paper provides a new means to increase the throughput of proposing crystal structures compared to prior methods. It achieves this by using Chemeleon2’s frozen VAE and training a MeanFlow model that operates in that latent space.

Critical strengths
  • It achieves MP-20 mSUN performance comparable to Chemeleon2 using approximately 6× less generation compute.

  • Although the individual components are established in prior works, they successfully show that they can adapt them for few-step crystal generation, and this could be a quite meaningful contribution.

Critical weaknesses

I do have a few places where the authors could clarify so that the contribution can become a little bit more precise:

  • The discovery yield numbers for the fixed budget simply multiply the SUN/mSUN rate measured on a smaller sample size of 1,024 by a larger count. However, the uniqueness fraction depends on the sample count and may decrease as more samples are generated, so this linear extrapolation needs validation.

  • The cost that is reported mostly covers generation and decoding and excludes the cost of relaxation and scoring.

  • In the fine-tuning comparison, NFT starts from an unconditional model, but the highlighted LeMat improvement uses a conditional baseline.

Questions for authors
  1. Could the discovery yield be validated on a larger generation count? Doing that shouldn’t be too expensive given the efficiency of the method.
  2. Could the authors discuss how the energy relaxation cost compares with the generation cost? This could help the reader see where the bottleneck is in the overall discovery process and what future studies should aim to optimize in order for the whole pipeline to be faster.
  3. Could the authors report the actual unconditional parent under the same evaluation protocol in the fine-tuning comparison? This would isolate the contribution of fine-tuning.
2 comments

Thank you for your kind feedback, and for recognizing our meaningful contribution!

Q1. Yield at larger sample counts. We agree on your point that uniqueness can drop as more samples are drawn. We have now drawn 10,000 structures from each of MeanFlux, MF-NFT, Chemeleon2 and Chemeleon2 + RL (Appendix B.10), and we report counts at smaller n as averages over random subsets of these draws. The base models stay close to linear (Table 8).

ModelUnique, n=1kUnique, n=10kSUN vs linear, n=10k
MeanFlux0.9860.9020.947
Chemeleon20.9960.9610.979

At the 6,600 draws that a 100-TFLOP budget allows, MeanFlux still keeps 96.5% of its linear SUN yield. Both fine-tuned models lose uniqueness, MF-NFT a bit more than Chemeleon2 + RL. The new Figure 10 plots their yield against the number of draws and against generation compute, and the table below presents the counts.

SUN crystals found, by draws2565121,0242,0484,0968,19210,000
MF-NFT4276133224370607701
Chemeleon2 + RL3059114216396699818

Neither yield grew quite linearly. At 10,000 draws MF-NFT has found 701 SUN crystals and Chemeleon2 + RL 818, which is 54% and 73% of the linear extrapolation from 1,024 draws. Per draw the two models are close in performances, and the difference lies in the cost of a draw. A 100-TFLOP budget allows about 6,600 MF-NFT draws but only about 1,100 Chemeleon2 + RL draws. So MF-NFT yields about 460 SUN crystals per 100 TFLOPs on relaxed LeMat-GenBench, compared with 50 for Chemeleon2 + RL.

Q2. Relaxation cost. Thanks for raising this question. Relaxing costs are indeed non-negligible. Relaxing one structure with ORB, MACE and UMA costs ~2.88 TFLOPs, about 190 times the cost of generating it with MeanFlux (8 steps). The cost per SUN crystal therefore depends mainly on the “hit” rate (Appendix B.4, Table 4).

This demonstrates the importance of generating structures that are already close to their relaxed state, which LeMat-GenBench measures as the RMSE between generated and relaxed positions (last column). MF-NFT's samples move 0.36 Å on relaxation, among the lowest of the models we scored, and 22.6% of its draws are already mSUN before any relaxation (Table 11).

ModelTFLOPs per relaxed SUN crystalShare of Generation FlopsRMSE (Å)
MF-NFT250.5%0.36
Chemeleon2 + RL613.3%0.39
MeanFlux880.8%0.58
Chemeleon2953.7%0.63
MatterGen6,58695.2%0.34

With relaxation included, MF-NFT needs 2.4 times less compute per SUN crystal than Chemeleon2 + RL. In deployment, we can relax only valid, unique and novel candidates so that fewer candidates reach relaxation to save more flops.

Q3. The unconditional starting model. Table 11 (Appendix B.13) now also reports the metrics of the unconditional model that NFT starts from, under the same eval protocol.

ModelSUNmSUN
MeanFlux, unconditional (NFT start)1.59.6
MF-NFT11.441.2
Chemeleon22.619.1
Chemeleon2 + RL4.644.0

Our conclusion does not change while we see an even larger gain. NFT raises SUN 7.8 times and mSUN 4.3 times over its true starting model (5.6-fold SUN over the conditional model), while GRPO raises Chemeleon2's SUN 1.7-fold and mSUN 2.3-fold. On the MP-20 suite, MF-NFT reaches 0.583 mSUN, compared with 0.352 for its starting model (Appendix B.7).

Hey Reviewer Ucvb,

As the rebuttal period is drawing to an end, it would be great if you could let us know if we have addressed your concerns in our previous response, where we yielded larger sample counts for understanding the behavior of SUN rate, counted relaxation costs as part of discovery efficiency, and clarification on the comparison between NFT and base MF model.

Best, Paper Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Authors' Response to All Reviewers

We thank all reviewers for their thoughtful feedback. We are encouraged that the reviews recognize our work for "successfully adapting prior methods for crystal generation" (Ucvb), its "strong few-step results", "fair evaluation protocol" and "interesting scaling efforts" (YTK9), its "careful experimental methodology", "outstanding sampling efficiency" and "broad and systematic evaluation" (qgfi), and its "promising, cheap, and practical sampling" (7g3f).

We have updated our submission and highlighted changes in the revised manuscript in red. Here are a few common queries we addressed:

New results on reward fine-tuning. With an updated NFT recipe (cf. formula 6), MF-NFT now reaches 11.4% SUN on relaxed LeMat-GenBench, the highest of all models we evaluated, while with the lowest compute (Appendix B.13, Table 11). Relative to the unconditional model it starts from, NFT raises the SUN rate 7.8 times, while GRPO raises Chemeleon2's SUN rate at a lower 1.7 times (Section 3.4). Even when we train with a reward that uses no LeMat-GenBench metrics (i.e. only mp20), MF-NFT still reaches 6.2% SUN (Appendix B.3, Table 11). We also updated the discovery efficiency chart accounting for relaxation costs (Figure 12), with MF-NFT placed unarguably on the frontier of quality and efficiency, enabling the most SUN crystals with a fixed FLOPs budget.

ModelTFLOPs per 1,024SUN (%)SUN per 100 TFLOPsSUN per 1,000 TFLOPs, with relaxation
MF-NFT15.511.445640.3
Chemeleon2 + RL94.34.65016.3
PLaID++2,3552.81.25.6
MeanFlux15.52.113511.3
Chemeleon294.32.62910.5
ADiT376*3.28.810.8
Crystalite3301.95.86.2
MatterGen56,4290.90.020.15

SUN is the relaxed LeMat-GenBench rate over 1,024 draws. Reward-finetuned models are listed first. *Estimated, as ADiT has no public weights.

Novelty of our method. Many of them are rooted in our methodologies. MeanFlux builds on prior methods (MeanFlow, latent-space modeling and NFT), while our contributions are the effective adaptations that make them work extremely well for crystals. Our work gives a deeper empirical understanding of few-step crystal generation with the underlying methods across scales, and of how far efficiency and quality can be pushed together.

  • Discovery-efficiency frontier. We compare crystal generators by the number of SUN crystals they yield per unit of compute, which combines quality and sampling cost in one measure, with and without relaxation (Figures 1 and 12). Every baseline is evaluated under the same protocol and costed on the same hardware; future models can be placed on the same chart.
  • JVP cap per crystal. We cap the norm of the JVP correction for each crystal, without which training simply does not converge (Section 2.2, Appendix B.9).
  • Reward design. Our new reward design scores stability with two smooth terms and multiplies them by validity, uniqueness and novelty; a crystal is rewarded only if it is new and stable at once (Section 2.3, Eq. 6).
  • Structure prediction. For CSP, we retrain the model on a geometry-only latent (from a new VAE) with the known species attached to each atom (Section 2.4).
  • Few-step efficiency. The sampler matches or surpasses best-performing models at 6 to 90 times less sampling compute (Sections 3.2 and 3.5), which makes reward feedback looping much cheaper. At equal NFE, MeanFlow also outperforms flow matching and every other comparable baseline running at each step count (Figure 2, Table 7).
  • Empirical scaling. mSUN follows a power law in model and data size, and the best model size grows with the training budget as C^0.80 (Section 3.3). This is a useful reference for investing compute budgets on worthwhile areas (data or model sizes).
  • Fair evaluation for future reference. We re-score every baseline under a unified protocol and measure costs on the same accelerator setup including both generation and relaxation (Tables 1 and 3, Figure 12).

New baselines and ablations. We added PLaID++ run from its released weights (Table 11), CrystalFlow and FlowMM (Tables 4 and 11), Chemeleon2 and Crystalite at equivalent step counts (Figure 2, Appendix B.8, Table 6), measured actual FLOPs for all models with public weights (Tables 1 and 3, Appendix B.4), ablations of the JVP cap, the schedule, decoding variants, and the objective (Appendix B.9, Table 7), yields at more draws (Appendix B.10, Table 8), end-to-end cost including relaxation (Figure 12, Appendix B.4, Table 4), and a pipeline figure with pseudocode (Appendix B.1, Figure 6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.