acceptodds
Under review as a conference paper at ICLR 2027

FieldForge: A Pretraining and Post-Training Framework for Flow-Based Crystal Generation

Abstract

Generating thermodynamically competitive crystals while retaining novelty remains a central challenge in materials discovery. To address this challenge, we present FieldForge, a flow-based training framework that unifies pretraining (PT), supervised fine-tuning (SFT), and reinforcement learning (RL) within a single crystal generator. Pretraining combines a Transformer supporting variable atom counts with flow matching, while respecting crystal geometry throughout the generative process. SFT then adapts the pretrained generative distribution through controlled data selection, preparing the generator for RL without introducing additional training structures. RL uses a GRPO-inspired critic-free policy update to refine the same velocity field using stability, novelty, and diversity rewards, with stability estimates provided by an interatomic potential. On the MP-20 track of LeMat-GenBench, the pretrained FieldForge model achieves state-of-the-art stable-unique-novel (SUN) and metastable-unique-novel (MSUN) rates, combining thermodynamic quality with uniqueness and novelty. SFT followed by RL further improves generation performance on both MP-20 and Alex-MP-20, demonstrating the effectiveness of the complete training framework.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal Generation, Flow Matching, Pretraining, Post-Training, Reinforcement Learning, Materials Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sp3qOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FieldForge, a flow-based training framework that unifies pretaining, quality-based supervised fine-tuning, and reinforcement learning in order to train a transformer-based backbone to generate inorganic crystalline materials with variable atom counts. The authors carry out a detailed study of the effect of every training stage on the final performance and report state-of-the-art performance in the de novo generation and crystal structure prediction tasks.

Critical strengths
  1. The de novo generation benchmark results from LeMat-GenBench are strong, and the study of the influence of the different training stages on the final performance is interesting.
  2. The quality-based supervised fine-tuning training stage is novel and appears to have a favorable effect on FieldForge's competitive performance.
Critical weaknesses

From a technical perspective, the paper mainly combines pre-existing ideas from generative models for inorganic crystalline materials:

  • The transformer-based architecture is very similar to Crystalite.
  • The stochastic flow policy in the reinforcement-learning setup in Eq. (3) seems very similar to the stochastic perturbations in OMatG-IRL.
  • The rewards in the reinforcement-learning setup are taken from Chemeleon2.
  • The anti-annealed sampling is taken from FlowMM.
  • The trade-off between stability and novelty was already noted in the Crystalite paper and in the paper on the continuous SUN metric.

While I do think that such a combination of pre-existing knowledge in one framework is worthwhile because it can lead to state-of-the-art performance as in FieldForge, I do not think that this is enough for this ICLR conference. Besides the quality-based supervised fine-tuning step, I do not see any new technical insights that are informative for the generative modeling community for inorganic crystalline materials. With respect to the quality-based supervised fine-tuning, it is unclear whether similar performance could be achieved by using classifier-free guidance on the energy above hull as in MatterGen.

Questions for authors
  • How does the performance of classifier-free guidance on the energy above hull as in MatterGen compare to the performance of quality-based supervised fine tuning?
  • The CDVAE splits of MP-20 and MPTS-52 are flawed because different polymorphs of the same composition are separated across training, validation, and test sets. What is the performance of FieldForge on the polymorph-split datasets from Martirossyan et al., "All that structure matches does not glitter," NeurIPS 2025, including the METRe and cRMSE metrics?
5 comments
AuthorsOpenReview ↗

Technical Contributions, CFG Comparison, and Polymorph-Aware Generalization

We thank the reviewer for the assessment and address the novelty concerns and requested experiments below.

Technical contributions

FieldForge introduces a unified PT–SFT–RL framework for flow-based crystal generation, in which a single joint velocity field over elemental features, coordinates, and lattice variables supports pretraining, quality-based supervised adaptation, and reward-driven optimization. The framework connects explicit control of the stable–metastable balance through data selection with joint RL over composition and geometry. Specifically, QB-SFT adjusts the proportions of on-hull and near-hull structures selected from the original training set to reshape generation preferences without changing the architecture or flow-matching objective, providing an effective initialization for subsequent RL.

  1. Flow Transformer (Crystalite). FieldForge learns a joint velocity field over elemental features, coordinates, and lattice variables, whereas Crystalite uses a diffusion Transformer. Beyond GEM, crystal-periodic 3D-RoPE and lattice-metric flow paths encode periodicity and non-orthogonal geometry. A learnable null-species prototype enables atom counts to be generated jointly with composition (Appendix B).

  2. Joint RL and periodic likelihoods (OMatG-IRL). Our policy jointly reinforces all three crystal variables, unlike Chemeleon2’s latent-space RL and OMatG-IRL’s DNG extension, which fixes discrete species generation (Appendix K). Channel-wise transition likelihoods use a nearest-image approximation for periodic coordinates, active-atom masks, and reference-policy regularization (Eqs. 3–7; Appendix E).

  3. Reward design (Chemeleon2). Shared objectives do not imply identical rewards. We combine dense energy, near-hull and stability bonuses, and SMACT validity; diversity uses element-fraction and AMD descriptors instead of VAE-derived features. Chemeleon2’s default DNG reward lacks the SMACT term. Removing our near-hull, stability, or validity terms reduces SUN+MSUN by 2.24, 4.80, or 3.68 percentage points in the 625-structure internal diagnostic (Appendix F.2).

  4. Anti-annealing (FlowMM). We adopt FlowMM’s linear velocity scaling and show that anti-annealing extends the observed PT metastability–novelty frontier relative to the unannealed checkpoint trajectory. It also increases MSUN across matched training checkpoints without retraining (Figures 2 and 5).

  5. Stability–novelty trade-off. The PT–SFT–RL mean lies beyond the observed non-RL frontier across PT checkpoints (0.9M–2.5M training steps) and PT/PT–SFT anti-annealing sweeps (Appendix F.3/Figure 5). With the sampler fixed, RL after QB-SFT raises MP-20 novelty from 63.56% to 75.48% and MSUN from 33.96% to 41.12%, while maintaining similar metastability (57.00% → 57.32%).

The full pipeline raises MSUN over PT from 29.88% to 41.12% on MP-20 and 42.04% to 53.20% on Alex-MP-20, and achieves higher SUN+MSUN than direct PT–RL on both datasets.

Energy-conditioned CFG versus QB-SFT

We tested CFG with guidance 2 and target hull energy 0 eV/atom. FieldForge-CFG and QB-SFT use the same pretrained checkpoint, matched post-training budgets, and identical LeMat-GenBench evaluation. MP-20 results (%) are:

MethodNovelMetastableSUNMSUNSUN+MSUN
MatterGen-CFG45.9259.122.9223.5226.44
FieldForge-CFG49.2857.964.6425.8430.48
FieldForge-QB-SFT63.5657.001.8833.9635.84

At similar metastability, QB-SFT exceeds FieldForge-CFG by 14.28 points in novelty and 5.36 points in SUN+MSUN; CFG has higher SUN. Thus, tested CFG settings do not reproduce QB-SFT’s higher novelty and combined discovery yield. The equal-sized random-SFT control on MP-20 further supports the selection rule (Appendix D.1). And QB-SFT’s selector is model-agnostic by design.

Polymorph-aware generalization

We trained FieldForge from scratch on MP-20-polymorph-split, with no SFT or RL. Revised Table 3 includes the baselines from Martirossyan et al., Table 2:

ModelMETRe (%) ↑RMSE ↓cRMSE ↓
DiffCSP53.10.0840.279
FlowMM65.20.0790.226
OMatG-LinearODE70.50.0560.187
OMatG-IRL70.68—0.183
FieldForge75.720.03010.1442

FieldForge achieves the highest METRe and lowest reported RMSE/cRMSE among the compared models. Following the benchmark protocol of Martirossyan et al. cited by the reviewer, we evaluate all 9,032 test structures, generating one candidate per reference, using OMatG’s official all/unfiltered evaluation (Appendix G.4). Martirossyan et al. do not provide a polymorph-aware split for MPTS-52, so we retain its standard chronological split.

I thank the authors for their reply. However, I am still not convinced that FieldForge introduces any foundational new approach to generative modeling for inorganic materials. Like I said before, it does combine lots of scattered ideas from various papers in a single framework (my previous list did not mention MiAD for a model with variable atom counts). This is a practical thing but, in my opinion, not enough methodological novelty for ICLR.

I do agree with the authors that their QB-SFT is actually novel and I appreciate the new comparison with energy-conditioned CFG. Two follow-up questions about this that might nudge my score:

  1. The realized on-hull fraction in QB-SFT appears to be quite important. In fact, for an on-hull fraction of 1, SUN+MSUN performance drops to around 30% on MP-20 which curiously enough is quite close to FieldForge-CFG with a target hull energy of 0 eV/atom. Does FieldForge-CFG improve if one increases the target hull energy (which seems similar to reducing the on-hull fraction in QB-SFT)?
  2. If the authors would like to claim that QB-SFT is helpful independently of the model, I would be curious to see MatterGen's performance under QB-SFT.
AuthorsOpenReview ↗

Energy-Guided Sampling in FieldForge and Transfer of QB-SFT to MatterGen

Thank you for suggesting these experiments. We extended the CFG target sweep on FieldForge and tested whether QB-SFT transfers to MatterGen on MP-20.

1. Higher CFG targets improve novelty and bring FieldForge’s MSUN to a level comparable to QB-SFT.

The sweep supports the reviewer’s intuition. Our previous reply considered only a target energy above hull of 0 eV/atom. Here we vary the target at a fixed guidance scale of 2.

FieldForge settingNoveltyMetastableSUNMSUNSUN+MSUN
CFG, target 0 eV/atom49.2857.964.6425.8430.48
CFG, target 0.025 eV/atom64.0857.881.9233.4035.32
CFG, target 0.050 eV/atom68.4857.000.8834.2835.16
CFG, target 0.075 eV/atom71.1256.441.0034.7235.72
────────────────
QB-SFT63.5657.001.8833.9635.84

At 0.025 eV/atom, the metastable fraction is almost unchanged from the zero-target run, yet novelty and MSUN both rise. CFG at this target and QB-SFT give similar results across all five metrics. Increasing the target further continues to improve novelty, with little change in combined yield. Our earlier zero-target comparison therefore understated what CFG can achieve on FieldForge.

QB-SFT also provides an effective starting point for RL. With the sampler fixed, subsequent RL raises MSUN from 33.96% to 41.12% while maintaining a similar metastable fraction. In our initialization ablation, RL starting from QB-SFT gives higher SUN+MSUN than RL starting directly from pretraining on both MP-20 and Alex-MP-20.

2. QB-SFT transfers from flow-based FieldForge to diffusion-based MatterGen.

Our previous reply used MatterGen’s released energy-conditioned checkpoint, trained on Alex-MP-20. For this comparison, we followed MatterGen’s setup to train an energy-conditioned model on MP-20. We also applied QB-SFT to MatterGen on MP-20, varied the on-hull fraction in the fine-tuning set, and included random-subset SFT as a control. The table reports results after MatterSim relaxation. Released checkpoint configuration

MatterGen setting on MP-20NoveltyMetastableSUNMSUNSUN+MSUN
Base model69.4437.880.2019.8820.08
────────────────
CFG, target 0 eV/atom67.9639.080.5220.0820.60
CFG, target 0.025 eV/atom70.8837.200.2819.5219.80
CFG, target 0.050 eV/atom73.4035.280.3619.1619.52
CFG, target 0.075 eV/atom74.4835.880.5620.1620.72
────────────────
Random-subset SFT, on-hull fraction 0.43868.4440.040.4821.6822.16
QB-SFT, on-hull fraction 0.3071.0038.000.4021.6822.08
QB-SFT, on-hull fraction 0.5072.7636.280.4821.6422.12
QB-SFT, on-hull fraction 0.7071.5234.680.5620.2820.84
QB-SFT, main setting69.3642.120.9223.3624.28

On MatterGen, novelty also rises with the CFG target, but the metastable fraction generally falls and MSUN changes little. Our main QB-SFT setting retains the base model’s novelty while generating a larger fraction of metastable structures. It gives the highest SUN and MSUN among the settings tested.

The random-subset control helps separate the benefit of data selection from that of additional training. With all SFT runs trained for 1,800 steps, the 30% and 50% on-hull mixtures give combined yields close to random-subset SFT. At 70%, SUN improves slightly but MSUN falls. The main QB-SFT setting improves both metrics over random-subset SFT, showing that the gain on MatterGen depends on the fine-tuning mixture.

MatterGen before relaxation. We evaluated both sweeps on the structures as generated. Row labels follow the table above.

Unrelaxed MatterGenNoveltyMetastableSUNMSUNSUN+MSUN
Base model69.8033.440.1214.4814.60
────────────────
CFG, 067.8035.800.4416.0816.52
CFG, 0.02570.8832.600.1614.9215.08
CFG, 0.05073.4430.840.3214.3214.64
CFG, 0.07574.4830.240.3614.5214.88
────────────────
Random-subset SFT, 0.43868.7235.320.2416.1616.40
QB-SFT, 0.3071.0433.720.1616.2416.40
QB-SFT, 0.5072.7631.200.2015.8416.04
QB-SFT, 0.7071.4030.160.1614.4414.60
QB-SFT, main setting70.2037.600.2818.6018.88

Before relaxation, zero-target CFG gives the highest SUN. The main QB-SFT setting already generates the largest fraction of metastable, unique, novel structures and gives the highest combined yield.

Reviewer_tpFGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FieldForge is a flow-matching crystal generator trained in three unified stages on a single Transformer velocity field: rectified-flow pretraining (PT) with geometry-aware inductive biases (null-species token for variable atom count, crystal-periodic RoPE + GEM geometric attention, lattice-aware periodic coordinate path), a quality-balanced supervised fine-tuning stage (QB-SFT, a controlled data-selection reweighting), and GRPO-style critic-free RL with a composite terminal reward (MLIP stability E_hull + near-hull/stable bonuses + SMACT validity + novelty + diversity). Inference uses deterministic anti-annealed sampling. The authors report state-of-the-art SUN/MSUN on the MP-20 track of LeMat-GenBench for the pretrained model, claim post-training further improves generation on both MP-20 and Alex-MP-20, and report competitive CSP@1. The central concerns are that (a) the full pipeline's headline "improvement" rests on MSUN while SUN and the stable rate regress on MP-20, (b) the RL reward and the evaluation metric share both a novelty/uniqueness factor and the same MLIP (NequIP) relaxation back-end, creating a reward–evaluation coupling, and (c) no uncertainty is reported on any headline table despite a multi-seed protocol, and no DFT verification is performed in a paper whose premise is materials discovery.

Critical strengths
  • the pretrained mp-20 result is real and fairly measured. PT MSUN 29.88 vs next-best 22.60, like-for-like, same benchmark, pre-relaxation. this is the solid, correctly-scoped empirical core and it's a sizeable gap.
  • clean unified pipeline. PT -> SFT -> RL on one velocity field with one eval protocol, reproducible appendices. plus QB-SFT is a practically useful controllability lever that reframes the stable-vs-metastable tradeoff as a tunable knob.
Critical weaknesses
  1. Missing citations
  • CrystalGRPO (2608.06582), OMatG-IRL (ICML 2026), OMatGRPO — as RL baselines and for positioning.
  • The reward-hacking paper (2610.03880) — as direct evidence for the reward–metric coupling risk the paper should pre-empt.
  • LeMat-GenBench ensemble-MLIP evaluation (2512.04562) should be cited specifically where single-potential (NequIP) relaxation is used, to justify or problematize that choice.
  1. the headline contradicts the paper's own table. the advocated full PT-SFT-RL pipeline is worse than plain PT on SUN (2.72 < 3.12) and on stable rate (7.08 < 11.28, a ~37% drop); only MSUN and novelty go up. "improves generation on both datasets" only holds if you silently crown MSUN the one true metric, which they never justify. on alex-mp-20 they don't even beat chemeleon2 on MSUN or novelty.

  2. reward-evaluation coupling with no DFT and no cis. the RL novelty/uniqueness reward is the same U-factor the eval multiplies in, and both reward and eval relax with nequip, so the reported MSUN/novelty gain is partly circular. with zero DFT verification and zero error bars, there's nothing independent to adjudicate whether the gain is real or reward-hacking (see 2610.03880, which documents exactly this failure mode and isn't cited).

Questions for authors

can you address weaknesses from above?

also:

  • (clarification) if you re-score the RL outputs with a different potential (orb-v3/uma, which your appendix says you have) and the lemat-genbench mlip ensemble, how much of the MSUN/novelty gain survives?
  • (clarification) DFT-relax a random 50-100 of your "stable/metastable novel" structures: what is the MLIP→DFT agreement rate?
  • (clarification) with per-seed std or bootstrap cis on tables 1-3, do the CSP@1 leads (0.06-0.26 MR) still count as "highest"?
  • (clarification) how much MSUN remains if you promote the "without creativity" ablation row (−6.40 pp) into the main comparison?
  • (suggestion, not scoring) head-to-head vs crystalgrpo / omatg-irl for the RL stage.
2 comments
AuthorsOpenReview ↗

Critical Weaknesses

Citations and positioning

We thank the reviewer for these suggestions. CrystalGRPO and OMatG-IRL were already cited and discussed in Section 2, Reward-based post-training, of our original submission. Their relationship to our RL setting is addressed in Question 5.

OMatGRPO (arXiv:2610.03880) was first submitted to arXiv on October 2, 2026, after our September 25 submission, and was therefore absent from the original manuscript. We acknowledge its analysis of reward exploitation and benchmark limitations. Appendix H.5 now includes composition-level checks addressing these concerns. We will add the formal citation and explain this connection in the revision.

Section 4.1 clarifies and cites the LeMat-GenBench protocol: NequIP pre-relaxes generated structures; final stability is evaluated using MACE-MP, UMA, and Orb-v3, not NequIP.

Weakness 1: Metrics and scope of our claims

We acknowledge the metric differences. Our goal is to evaluate the framework and its training stages. RL should be assessed through PT–RL versus PT and PT–SFT–RL versus PT–SFT. Both comparisons improve MSUN, novelty, and SUN+MSUN on both datasets:

DatasetComparisonMSUN (%)Novelty (%)SUN+MSUN (%)
MP-20PT → PT–RL29.88 → 33.8859.12 → 66.8833.00 → 37.60
MP-20PT–SFT → PT–SFT–RL33.96 → 41.1263.56 → 75.4835.84 → 43.84
Alex-MP-20PT → PT–RL42.04 → 52.3267.80 → 74.6444.96 → 53.64
Alex-MP-20PT–SFT → PT–SFT–RL42.36 → 53.2062.84 → 77.4047.28 → 58.64

SUN and MSUN cover disjoint stable and metastable categories. Their sum measures the fraction of generated structures that are unique, novel, and within 0.1 eV/atom of the hull. We do not treat MSUN as the sole criterion.

We acknowledge that, on MP-20, the full pipeline has lower SUN (2.72% vs. 3.12%) and stable rate (7.08% vs. 11.28%) than PT. However, both PT–RL versus PT and PT–SFT–RL versus PT–SFT improve MSUN, novelty, and SUN+MSUN on both datasets. Adding RL to PT–SFT also improves SUN on both datasets. These corresponding comparisons support the effectiveness of RL. The lower SUN and stable rate of the full pipeline relative to PT reflect a trade-off across the training stages, alongside higher novelty and MSUN.

The full pipeline also exceeds direct PT–RL in SUN+MSUN: 43.84% versus 37.60% on MP-20 and 58.64% versus 53.64% on Alex-MP-20. This supports the value of SFT initialization for subsequent RL.

We acknowledge that FieldForge has lower MSUN and novelty than Chemeleon2 on Alex-MP-20, but higher SUN and stable rate. We do not claim superiority across all metrics. Our main objective is to validate the framework through corresponding training-stage comparisons.

Weakness 2: Reward–evaluation coupling

We acknowledge this risk. Training and evaluation both involve novelty and uniqueness, but evaluation does not reuse training-time scores or decisions. The creativity reward combines reduced-composition novelty, a within-batch uniqueness proxy, and continuous structural distance. LeMat-GenBench independently uses StructureMatcher and its reference corpus. SUN and MSUN are joint rates, not products of independent marginal rates.

NequIP is used for RL reward relaxation and local pre-relaxation. Main-table scores come from LeMat-GenBench using MACE-MP, UMA, and Orb-v3 with their respective reference hulls. Questions 1–4 report cross-potential evaluation, DFT validation, CSP uncertainty, and full-protocol reward ablations.

We also address OMatGRPO’s concerns about single-element structures and sparse reference hulls. A separate audit generated 2,500 new structures per training stage and recomputed MSUN under cumulative restrictions: excluding single-element crystals, requiring at least 12 near-hull reference phases, and retaining at most one candidate per formula.

Under the strictest setting, PT–SFT–RL achieves 37.88% MSUN, retaining 91.9% of its original audit yield. The equally filtered PT–SFT baseline reaches 31.92%, a 5.96-point gap. Appendix H.5 includes elemental-diversity and reference-phase distributions.

AuthorsOpenReview ↗

Reply to questions

Question 1: Cross-potential evaluation

We fixed the PT–SFT–RL checkpoint and the same 2,500 structures, changed only the pre-relaxation potential, and used the same LeMat-GenBench multi-MLIP evaluation:

PotentialNovelty (%)MSUN (%)
NequIP75.4841.12
MACE75.6040.56
Orb-v375.9240.28
UMA75.3340.24

MSUN drops by only 0.56–0.88 percentage points, retaining 97.9%–98.6% of the original score. Novelty remains at 75.33%–75.92%. These are retention rates of the RL model’s MSUN score, not recomputed RL-versus-baseline gains. Full results are in Appendix H.2.

Question 2: MLIP-to-DFT agreement

We performed DFT validation on 100 structures. Of 2,500 generated structures in the validation pool, 819 were classified by the MLIP as stable or metastable and novel. We selected a stratified random sample of 5 stable and 95 metastable candidates.

We used two-stage VASP relaxation followed by a static calculation, with consistent PBE/PBE+U settings for candidates and competing reference phases. For the 24 candidates requiring +U, the corresponding hull reference phases were treated consistently.

87/100 structures remained within 0.1 eV/atom of the hull under DFT: 18 stable and 69 metastable. The 87% agreement rate has a Wilson 95% CI of 79.0%–92.2%. The mean absolute and signed hull-energy errors were 0.032 and −0.004 eV/atom, respectively. Of the 13 structures that failed the threshold, 8 exceeded it by no more than 0.006 eV/atom.

Question 3: CSP uncertainty and conclusions

We report means and standard deviations across four sampling seeds, together with 95% confidence intervals from bootstrapping the test samples:

DatasetSettingMR (%) mean ± SDMR 95% CI (%)RMSE mean ± SDRMSE 95% CI
MP-20CSP@166.31 ± 0.04[65.38, 67.24]0.0322 ± 0.0005[0.0304, 0.0341]
MP-20CSP@2074.68 ± 0.12[73.82, 75.55]0.0434 ± 0.0002[0.0412, 0.0456]
MPTS-52CSP@131.55 ± 0.16[30.56, 32.50]0.0991 ± 0.0014[0.0952, 0.1031]
MPTS-52CSP@2039.61 ± 0.02[38.56, 40.65]0.0952 ± 0.0004[0.0910, 0.0995]

We agree that the small CSP@1 leads over Crystalite do not establish statistical significance. With the OMatG baselines included, FieldForge has the numerically highest CSP@1 MR on MPTS-52 and the lowest CSP@1 RMSE on MP-20, but not the highest MP-20 MR. Our manuscript emphasizes competitive performance across both datasets and sampling budgets, considering CSP@1 and CSP@20 together, rather than significant superiority in every setting.

We also added MP-20-polymorph-split experiments. With pretraining alone, FieldForge achieves 75.72% METRe, 0.0301 RMSE, and 0.1442 cRMSE, the best reported values among the compared methods.

Question 4: Ablation without creativity

We added new reward-component ablations under the full LeMat-GenBench multi-MLIP protocol. Each configuration uses 10 sampling seeds and 2,500 structures, with the same SFT initialization, RL training schedule, and inference settings.

ConfigurationNovelty (%)SUN (%)MSUN (%)
PT–SFT63.561.8833.96
RL without chemical validity66.602.0436.68
RL without composition diversity73.762.9237.96
RL without creativity67.401.9236.80
Full RL75.482.7241.12

Without creativity, MSUN is 36.80%, 4.32 percentage points below full RL but 2.84 points above the PT–SFT baseline of 33.96%. Novelty also rises from 63.56% to 67.40%. Explicit novelty and uniqueness incentives contribute, but do not fully account for the RL improvements. The other ablations support complementary contributions from the reward components. Full results are in Appendix F.2.

Question 5: CrystalGRPO and OMatG-IRL

Below are the DNG-style results from Table 6 of OMatG-IRL. Both evaluations use LeMat-GenBench. OMatG-IRL evaluates 1,000 structures; we evaluate 2,500. All rates are percentages.

MethodNValidUniqueNovelStableMetastableSUNMSUN
DNG-style OMatG-IRL, relaxed1,00098.2097.7061.207.6045.600.6018.20
FieldForge (PT–SFT–RL), MP-202,50097.6896.7275.487.0857.322.7241.12

FieldForge reports higher novelty, metastable rate, SUN, and MSUN.

CrystalGRPO primarily addresses CSP with a given composition, reinforcing coordinates and lattices. Our RL stage addresses DNG and jointly optimizes composition and geometry. Our CSP results use pretraining alone, without SFT or RL. The RL task settings therefore differ.

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

FieldForge trains one Transformer velocity field for de novo crystal generation in three stages (pre-training, SFT, RL), then samples it with a tuned inference step. The results reported on LeMat-GenBench show competitiveness. The paper argues that RL from the SFT checkpoint "moves beyond" the pretraining trade-off, in which metastability rises while novelty falls.

Critical strengths
  • The pretrained generator is a clear step up on MP-20, with clear ablation. This is a solid empirical result for the community.
  • Steps are clearly reported, helping practitioners.
Critical weaknesses
  • The headline post-training claim, that RL "moves beyond the pretraining trade-off", is not separated from two simpler explanations: 1/ the evaluator is not independent of training and 2/ The "frontier" is compared against the pretraining checkpoint trajectory but not against the envelope the non-RL model traces under anti-annealing and sampler noise, which the paper shows also trade novelty against metastability.
  • QB-SFT is the main methodological novelty, because pretraining components (Crystalite's GEM, MiAD's null species) and critic-free RL for crystals (Chemeleon2) all have precedents. Without evidence that the selection rule matters, the framework contribution reduces to combining known stages.
Questions for authors
  • For the LeMat-GenBench numbers, which potential computes the hull energy, and is any part of it shared with the NequIP model used in the RL reward and in your local relaxation?
  • Were the SFT configuration (T, K, on-hull fraction) and the SFT checkpoint chosen using the same benchmark metrics reported in Tables 1 and 2?
  • How large is the pretrained DNG model (~67M parameters) compared with Crystalite and MCFlow on the leaderboard? Is the MP-20 gain partly a capacity effect?
1 comment
AuthorsOpenReview ↗

Response to Reviewer kJQb: RL Gains, QB-SFT, Evaluation, and Model Capacity

We thank the reviewer for the careful assessment. We address the two main concerns first, followed by the three questions.

RL beyond the pretraining trade-off

We have added Appendix F.3 and Figure 5. The comparison covers PT checkpoints from 0.9M to 2.5M steps at a=0 and 4, PT anti-annealing settings a=0,4,8,12, and PT-SFT settings a=4,8,12.

All four main-table variants use the same deterministic inference sampler with a=4. On MP-20, RL after SFT raises novelty from 63.56% to 75.48% and MSUN from 33.96% to 41.12%, while metastability changes from 57.00% to 57.32%. These gains occur without changing inference settings or adding inference-time stochasticity.

PT-SFT-RL lies outside the envelope of measured non-RL means: no evaluated non-RL configuration matches or exceeds it in both novelty and metastability. This supports extending the observed frontier beyond a single pretraining trajectory. Our claim concerns the evaluated configurations; broader anti-annealing and stochastic-sampler coverage requires further sweeps.

All main-table metrics, including MSUN, are computed by the official LeMat-GenBench pipeline, which assesses stability using MACE-MP, UMA, and Orb-v3 rather than NequIP. NequIP provides local pre-relaxation and RL rewards, not the final benchmark scores.

QB-SFT and the framework's methodological contribution

We acknowledge the precedents for GEM, null species, and critic-free RL. FieldForge differs in their adaptation to crystal geometry and training requirements. Lattice-metric flow paths account for non-orthogonal periodic cells, while periodic 3D-RoPE consistently encodes relative positions across cell boundaries. A learnable null-species prototype accommodates variable atom counts within the continuous elemental representation. RL directly reuses the learned flow over crystal variables, with a minimum-image approximation for coordinate transition likelihoods. These choices connect geometric modeling, variable-composition generation, and reward optimization within a shared formulation.

QB-SFT has a distinct role: broad pretraining learns the data distribution, whereas subsequent optimization should favor promising structures while retaining chemical and structural coverage. QB-SFT combines thermodynamic quality, distribution balancing, and structural diversity selection to shape an RL initialization that preserves exploration opportunities.

Appendix D.1 provides direct evidence. With 5,000 training structures on MP-20, QB-SFT improves MSUN over random SFT from 30.12% to 33.96% and SUN+MSUN from 33.12% to 35.84%. Thus, the gain is not solely a redistribution from stable to metastable successes. The on-hull-fraction scan at fixed K=5,000 and T=0.025 further shows that selection controls the SUN–MSUN balance. These results support the overall selection rule on MP-20, without isolating every component's contribution. The geometric and cross-stage adaptations also contribute to the framework's novelty.

Which potentials compute the benchmark hull energies?

The main-table LeMat-GenBench hull energies are computed using MACE-MP, UMA, and Orb-v3, each with its corresponding MLIP-specific hull constructed from LeMat-Bulk. NequIP is used for rewards and local pre-relaxation. The benchmark recomputes energies without reusing NequIP parameters, predicted energies, or reward-side hulls. Section 4.1 and Appendix G.1–G.2 now distinguish these roles explicitly; the evaluation protocol and results remain unchanged.

How were SFT configurations and checkpoints selected?

On both datasets, SFT configurations and checkpoints were selected using internal SUN+MSUN, with novelty calculations and NequIP-based stability assessment differing from LeMat-GenBench. Higher internal scores did not always yield better external results, so this score served as a selection proxy; the final scores in Tables 1–2 were not the direct selection criterion.

Could model capacity explain the MP-20 gains?

We have added Appendix B.3 and Table 8. FieldForge-67M (PT) and Crystalite each have approximately 67M parameters, versus 34M/134M/464M for MCFlow-S/B/L.

At comparable capacity, FieldForge-67M (PT) improves MSUN over Crystalite from 22.60% to 29.88%. FieldForge-9M (PT), FieldForge-17M (PT), and FieldForge-27M (PT) achieve 27.72%, 30.72%, and 30.52%, respectively; even FieldForge-9M (PT) exceeds every listed baseline in MSUN. All FieldForge entries use pretraining only . Baseline numbers come from the respective papers and the LeMat-GenBench leaderboard.

These results show that a larger parameter count is not necessary for the observed MSUN advantage and provide empirical support for FieldForge’s parameter efficiency.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Additional Public Benchmark Results and Clarification of Contributions

Dear Area Chair and Reviewers,

Before the discussion period closes, we would like to share an additional public benchmark result relevant to the assessment of FieldForge.

As of November 19, FieldForge ranks second on the public LeMat-GenBench leaderboard. The publicly listed FieldForge-Pretrain+SFT+RL entry achieves a SUN + MSUN rate of 44.6%. LeMat-GenBench was updated around October 2026, explaining the numerical differences from our manuscript, which reports pre-update results. These differences do not affect our performance conclusions. This provides additional evidence under a community-defined evaluation protocol and complements the experiments reported in our submission.

Regarding the originality concern, our contribution centers on the coordinated design of geometry-aware pretraining, stable–metastable data selection for fine-tuning, and reinforcement learning with shared flow dynamics for discovery-oriented distribution refinement. Our responses explain the specific design choices, their relationship to prior work, and the supporting experimental evidence; see Appendices B, D.1, and E of our manuscript, as well as our detailed response to Reviewer sp3q in the OpenReview discussion. The public benchmark result provides further evidence of the framework’s effectiveness.

We would appreciate consideration of this update together with our point-by-point responses during the subsequent discussion and meta-review, particularly when assessing the contribution and empirical significance of the coordinated training framework.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.