acceptodds
Under review as a conference paper at ICLR 2027

Let CSP Be Your ANCHOR: Adaptive Crystal Search over Frozen Structure Priors

Abstract

De novo crystal generation (DNG) models decide where to search in composition space and how to generate structures with one set of weights. We argue that discovery is better served by separating the two. A crystal structure prediction (CSP) model is a physical prior that should be improved by likelihood training, while rewards, including novelty measured against the search's own history, should act on a search over compositions. We introduce ANCHOR, a GRPO composition policy trained with multi-objective rewards around a frozen CSP model, and continuous adaptive novelty (CAN), a graded novelty score against known structures and a growing discovery history. Using the frozen CSP model as a fixed ruler under one evaluator, we test where adaptation should act. Replacing DNG compositions with ANCHOR's policy on the same CSP backbone raises MSUN from 11.4% to 47.6% and SUN from 1.1% to 22.1% at 99.9% formula uniqueness. Fine-tuning DNG models directly on the same rewards instead moves their composition marginal without raising their on-hull fraction. We show that KL-regularized fine-tuning of a DNG model can only reweight chemistry the pretrained model already supports by a bounded factor, while unregularized DNG fine-tunes move toward known or less stable chemistry. Even a stability-only reward routed into ANCHOR's CSP backbone roughly halves SUN relative to the frozen backbone, whereas likelihood training on structures found during search can improve a CSP backbone. Under MatterGen's evaluation pipeline, ANCHOR raises state-of-the-art MSUN from 29.2% to 41.3%, transfers without retraining to two further CSP backbones, and reaches 47.1% after distillation into Crystalite-CSP. As with any model optimised against a potential, its on-hull rate depends on that potential.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
reinforcement learning, generative models, materials discovery, crystal structure prediction, de novo generation, crystal generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_D4qDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors addressed an important problem in materials science called crystal material generation and proposed a new method, ANCHOR. They have two components:

  1. A small composition policy, trained with GRPO, proposes formulas or atom types
  2. A frozen CSP diffusion model generates a structure for each formula.

Also, to improve the diversity and novelty of the generated materials, they have incorporated continuous adaptive novelty (CAN), which scores a structure against both MP-20 and the search’s own history. During training, only the composition policy is updated and CSP diffusion model is pretrained and remains frozen.

Critical strengths
  1. The idea of the paper is simple yet interesting. They propose a small composition policy, trained with GRPO, which first samples the atomic composition or formula. Then, given that atom type, a frozen CSP model generates a structure x for that composition. The composition policy is trained using a continuous adaptive novelty and stability reward, which improves its SUN/MSUN values.
  2. Another strength of the paper is its experiments section. The authors have done extensive experiments to show effectiveness of their proposed model.
Critical weaknesses
  1. Firstly, the paper is a bit difficult for me to go through. The authors could work on making it easier to read.

  2. One of the central promises of the work is that it assumes a frozen CSP model remains effective on compositions rarely seen during training, so that discovery is limited mainly by composition choice. But there is no evidence or citation for that. I am not sure, for any unseen or less-observed composition during training, how well the CSP model will learn to generate a stable structure. If the compositional prior fails there, promising compositions may be paired with poor structures and wrongly rejected, giving the policy misleading feedback. More concrete evidence or justification for this would be good.

  3. The method trains the composition policy network using stability, validity, and novelty rewards, which raises two issues.

  • Firstly, computing actual stability needs DFT, and the authors use MatterSim as a proxy to do that. Since they did not evaluate their model using DFT, the results are good. But I wonder, if we evaluate this with DFT, how much improvement we will observe
  • Next, even with MatterSim, relaxing the sampled structure, computing SUN reward, and training the policy network will be a computationally costly process. Could the authors (a) report the total computational cost of a discovery campaign (e.g., GPU hours, wall-clock time etc.), broken down by component; (b) clarify whether all methods are compared under an equal budget of evaluated candidates or oracle calls; and (c) show how the performance gap with baselines changes as a function of this budget?
  1. One of the important contribution of the work is continuous adaptive novelty (CAN) formualtion. A simpler alternative for existing DNG models is to sample sequentially and discard candidates that match previously generated or known structures. Could the authors compare ANCHOR against such a “DNG + history-based filtering” baseline and report SUN? This would clarify how much of the gain comes from learning where to search versus simply removing duplicates.

  2. Lack of citations and baseline comparisons of many recent works on DNG: FlowMM, MCFlow, SymmCD, ADiT, Crystal-Text-LLM, TGDMat, FlowLLM, CrysLLMGen. Many of them, specially the LLM ones and hybrid models like FlowLLM and CrysLLMGen, have good SUN and MSUN.

  3. Also code base is not provided by the author, would love to check the code during rebuttal, if provided by the author.

  4. Finally, a few minor questions which I might have missed:

  • How is the number of atoms in materials to be sampled?
  • How many materials you sampled for evaluation?
  • Which Diffusion Model you used as frozen CSP model?
Questions for authors

Check the Weaknesses

4 comments
AuthorsOpenReview ↗

Follow-up to Reviewer D4qD

Follow-up to W6: code

The anonymised repository is available here: https://anonymous.4open.science/r/anchor-code-78C8/. It contains the training loop, the reward and evaluation code, and the configuration of the reported run. External components (KLDM, MatterSim, MatterGen's reference hull) are listed in the README.

Follow-up to W5: Crystal-Text-LLM under the MatterGen protocol.

For completeness, SUN / MSUN:

70B:

  • Their compositions through our KLDM CSP 0.3 / 15.6%
  • Released structures scored directly 0.5 / 4.0%

7B:

  • Their compositions through our KLDM CSP 1.0 / 13.1%
  • Released structures scored directly 1.0 / 12.9% (17 of 935 structures did not converge and are excluded)

For reference: ANCHOR scores 16.0% SUN and 41.3% MSUN.

So what we reported in the previous reply, their compositions through our CSP under our internal evaluation (0.3 / 31.6% for 70B and 1.1 / 14.3% for 7B), is the most favourable setting for 70B and about equivalent for 7B.

If these results address your concerns, we would be grateful if you would consider updating your assessment. If something is still open, please tell us what, and we will answer before the discussion ends on 18 November.

Thank you for the detailed responses and the additional experiments. Some of my concerns have been addressed, but the following remain unresolved:

W1 (Readability). Beyond the proposed changes, I would encourage the authors to rethink the presentation of the motivation and methodology to make the paper easier to follow.

W3 (Stability evaluation). Matbench Discovery accuracy reflects typical materials, whereas ANCHOR is optimized against MatterSim and may exploit its errors, as the drop under LeMat-GenBench (13.6% to 0.4% SUN) suggests. Structures far below the hull, i.e., more stable than all known phases, are a typical sign of potential exploitation rather than reassurance. Since the SOTA claim under MatterGen’s pipeline uses the training potential, I would encourage DFT validation of even a small random subset (e.g., 50–100 on-hull structures), which is modest relative to the paper’s total compute.

W3 (Compute). The comparison is not compute-matched. Each ANCHOR run uses 168 GPU-hours (~123k evaluated structures) before the final N = 1000 draw, whereas standard diffusion baselines are comparatively cheap to train and sample. I would suggest reporting SUN/MSUN per GPU-hour for all methods, counting total cost (training, sampling and evaluation), to show whether ANCHOR’s gains also hold in terms of efficiency.

Filtered baseline. The filtering baseline is applied only to Crystalite-DNG and KLDM-DNG. Stronger models such as MatterGen+REPA, FlowLLM and CrysLLMGen, which report higher S.U.N., should also be included under the same pipeline.

W5 (Baselines). Since published numbers are not comparable across protocols, I would encourage the authors to evaluate all baselines, including FlowLLM, CrysLLMGen and other SOTA models, under one common pipeline, reporting Stable, SUN, Metastable and MSUN. This would provide a clear and fair benchmark comparison.

Finally, as ICLR allows, I would ask the authors to reflect these changes in the revised manuscript during the discussion period.

Thank you for the follow-up. A revised manuscript is uploaded, and a second upload follows before 18 Nov.

W1 (readability). The revision adds an overview paragraph at the start of Sec. 3, states in the caption of Table 1 what each ablation cell isolates, and gives the sampling settings at the start of Sec. 4. A reworked presentation of the motivation and of Sec. 3 comes with the second upload.

W3 (stability). The two numbers are different steps: the validity gate (23.2 → 13.6% SUN) and the hull construction (21.5 → 0.4% on the 716 structures that pass both gates). LeMat-GenBench is also scored by potentials and is not a DFT reference (App. E.3.2). With ORB-v3, which the policy never saw, 632 of the 665 structures on the MatterSim hull stay on the hull and 664 stay within 0.1 eV/atom (App. D.3).

Regarding DFT validation, we would like to clarify some points:

  • The claim is methodological. The paper asks where reward optimisation should act, and answers it under one fixed evaluator. We do not claim DFT-verified stability. The abstract states the state-of-the-art result as holding under MatterGen's evaluation pipeline, and states that the on-hull rate of a model optimised against a potential depends on that potential.

  • Every reward-optimised method had the same evaluator and the same reward. This holds for the fine-tuned DNG models, the genetic algorithm, and ANCHOR (Table 2). Each has the same opportunity to exploit errors of MatterSim. Under these equal conditions the DNG fine-tunes do not raise their on-hull fraction, the genetic algorithm reaches 5.9% SUN, and ANCHOR reaches 22.1%. Our conclusion about where reward should act rests on this comparison.

  • Evaluator error is a property of the evaluator, not of the algorithm. Any reward-optimised generator inherits the errors of the potential it is trained against, and the paper says so in Sec. 4 and Sec. 5. In ANCHOR the evaluator is a plug-in (App. E.3.2), so a better potential can replace MatterSim without changing the method. This is why we studied evaluator dependence directly, including the ORB-v3 check above.

  • We plan a DFT check for the final version. We plan to check a random subset of on-hull structures with DFT, with settings matched to the Materials Project reference hull, and to report the result whatever it shows. This cannot be done reliably before 18 November. Until then, the paper states the limitation in Sec. 5 and App. D.3.

Does this resolve the concern? If a specific statement in the paper reads as a claim of DFT stability, we will also reword it in the second upload.

W3 (compute). We agree, and report the comparison in GPU-hours (App. F.8). Both methods start from a released checkpoint (KLDM-CSP for ANCHOR, Crystalite-DNG for the baseline), so we count what comes after. At small budgets the DNG is ahead: after 24 GPU-hours it found 380 on-hull and 4,369 metastable formulas, against 133 and 404. Its on-hull rate then falls (19 to 13 per GPU-hour) while ANCHOR’s rises (5 to 38). Even if its rate stopped falling, the DNG would reach about 2,200 on-hull formulas by 168 GPU-hours, against 2,885 for ANCHOR (extrapolated past 50k structures). For metastable formulas the DNG stays more efficient. So the gain is in on-hull formulas at larger budgets. As in our W4 reply, more samples do not change what a DNG proposes (App. F.4).

Filtered baseline and W5. A common pipeline is in place for every model we can evaluate: Table 3 scores MatterGen, KLDM, Crystalite, MatterGen+REPA and ANCHOR under MatterGen's pipeline. It reports Stable, SUN, Metastable and MSUN for Crystalite and KLDM (our own sampling) and for ANCHOR, and now also the values reported for MatterGen+REPA (8.2% Stable and 2.6% SUN, against 16.0% SUN for ANCHOR). For the second upload we are sampling the released MatterGen+REPA checkpoint and scoring it under our pipeline, with and without the history filter.

FlowLLM and CrysLLMGen release training code but neither trained weights nor generated samples. Evaluating them would mean retraining both from scratch, including fine-tuning a large language model, which is not possible within the discussion period. The FlowLLM repository includes language-model samples, but these are training input for its flow model, not FlowLLM output. Both models keep the atom types proposed by their language model and only refine coordinates and lattice, so their compositions come from a model of the Crystal-Text-LLM type. Tables 27 and 28 evaluate Crystal-Text-LLM compositions through our frozen CSP: 0.3% and 1.1% SUN under our evaluator, against 22.1% for ANCHOR, and 0.3% and 1.0% under MatterGen's pipeline, against 16.0%.

We have now answered each of your remaining points. The revised manuscript contains the changes.

We hope this resolves your concerns and we would be grateful if you would raise your rating

The 2nd upload adds MatterGen+REPA results and the reworked Sec. 3. We will post a note when it is in.

Dear Reviewer D4qD,

We have uploaded the second revision of the paper, which contains the following for each of your open points:

W1 (readability).

We rethought how the motivation and the method are presented. We have now rewritten both such that:

  • The introduction now explains the idea in three steps, in plain language, before any method is introduced.
  • In Sec. 3, the reasoning for separating search from structure generation used to be two long paragraphs. It is now three numbered reasons with a heading each, so the argument can be followed from the headings alone.
  • Terms that were used without explanation are now defined where they first appear: MP-20, E_hull and "on the hull" (Sec. 2), and GRPO, the policy's state and the validity gate (Sec. 3). The overview also points to Algorithm 1 for the full training loop.

W2

We have clarified our stance on whether a frozen CSP model is effective on rare compositions.

The paper no longer states this as an assumption. The first reason in Sec. 3 now clarifies that this is possible in principle/theory but with the natural caveat that a CSP model (like any model) is usually weaker away from its training data. We therefore also agree that how often it works is an empirical question, and we state this together with the numbers, i.e. that 92% of the compositions the search proposes are in element systems absent from the training data, while the frozen CSP model still produces a stable, unique and novel structure for 19.7% of them.

Sec. 3 also keeps the point that a CSP failure can only make a good composition look bad, never a bad one look good.

Filtered baseline with a stronger model.

We sampled MatterGen+REPA from its released checkpoint and scored it under the same pipeline as Table 2 (Table 27, App. F.7). MSUN / SUN:

  • MatterGen+REPA, unfiltered: 14.3% / 1.5%
  • MatterGen+REPA with the history filter (1,779 draws for 1,000 accepted): 22.5% / 2.4%
  • ANCHOR: 47.6% / 22.1% It is the strongest of the three filtered models, and the conclusion is the same as for the other two: a filter removes repeats, but it does not change which compositions the model proposes.

W3 (stability).

Sec. 5 now has a separate Limitations paragraph. It states that:

  • the gains hold under the reward the policy was trained against
  • that our stability estimates rely on MatterSim potentials
  • that the LeMat-GenBench check is only partial.

For FlowLLM and CrysLLMGen, our earlier answer still holds, since neither of these works release weights or samples.

We hope the paper is now much easier to read, and we would be glad if you would look at it again and reconsider your assessment and rating.

Reviewer_WDp6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the coupling between chemical composition exploration and crystal structure prediction in unconditional crystal generation. The authors argue that existing de novo generation models typically learn the composition distribution and conditional structure distribution jointly, while directly optimizing crystal-level rewards with reinforcement learning may undesirably alter both components and degrade the structural prior. To address this issue, the paper proposes ANCHOR, which decomposes crystal discovery into composition search and crystal structure prediction. Specifically, ANCHOR trains an independent composition policy with GRPO to explore promising chemical compositions, while keeping a pretrained CSP model frozen to generate corresponding crystal structures. In addition, the authors propose Continuous Adaptive Novelty (CAN), which combines continuous structural similarity and historical search information to provide dynamic novelty rewards for reinforcement learning.

Critical strengths
  1. Decoupling composition search from structure generation provides an interesting perspective for crystal discovery. The paper separates chemical composition selection from conditional crystal structure generation, allowing reinforcement learning rewards to directly optimize the composition search policy without modifying the pretrained CSP model. This design provides a clear experimental framework for studying the relative roles of composition exploration and structure generation in crystal discovery.
Critical weaknesses
  1. The generality of the learned composition policy is insufficiently validated. ANCHOR claims that optimizing a dedicated composition policy enables more effective crystal discovery, but it remains unclear whether the learned composition preferences generalize beyond the reward environment used during training. Although ANCHOR achieves strong SUN/MSUN under the MatterGen protocol, its performance drops substantially under LeMat-GenBench (Table 15), where KLDM-CSP achieves only 7.8% MSUN and 0.4% SUN. Even retraining with LeMat-specific validity criteria improves MSUN to 32.5%, but SUN remains only 1.0%. This suggests that the learned composition distribution may be sensitive to the underlying evaluator. More analysis of the learned composition space and cross-MLIP consistency is needed to determine whether ANCHOR discovers generally promising chemical regions or mainly exploits specific reward biases.
  2. The independent contribution of CAN's structural novelty modeling remains unclear. CAN introduces RDF-based structural similarity, kNN-RBF scoring, and historical search information as a structure-level novelty mechanism. However, its actual contribution beyond composition-level exploration is not fully established. CAN only compares structures with the same reduced formula, while the count bonus already penalizes repeated formula visits. Moreover, because ANCHOR only updates the composition policy with a frozen CSP, CAN cannot directly improve the conditional structure distribution. Appendix E.1 further shows that, in one evaluation setting, 5,731 out of 5,737 valid structures receive default novelty scores, with only 6 structures using RDF-based kNN-RBF comparison. Therefore, current experiments do not clearly distinguish whether CAN improves structural exploration or mainly encourages exploration of unseen compositions.
Questions for authors
  1. Comparison with Heuristic and Existing Composition Generation Methods. Table 2 compares ANCHOR primarily with randomly generated formulas and uniformly sampled charge-neutral formulas. However, these baselines do not exploit existing chemical knowledge or learned composition priors. Could the authors include stronger heuristic composition search strategies, such as generating novel charge-balanced formulas through chemically informed substitutions from existing MP-20 compositions while filtering out known formulas? Furthermore, several existing studies have explored composition generation methods [1,2,3,4]. Could the authors compare ANCHOR with these methods under the same frozen CSP backbone, sampling budget, and evaluation protocol? Such comparisons would clarify whether ANCHOR provides additional benefits over existing composition search strategies and whether the GRPO-based policy learns a more effective composition distribution for crystal discovery.
  2. Are the Learned Composition Preferences Chemically Meaningful? ANCHOR optimizes a composition policy using GRPO, but the evaluation focuses primarily on the resulting SUN and MSUN, with limited analysis of the learned composition distribution itself. Could the authors compare the composition distributions of the MP-20 training set, DNG-generated samples, and ANCHOR-generated samples in terms of element frequencies, element co-occurrence patterns, stoichiometric ratios, atom-count distributions, and major material categories? Furthermore, which regions of chemical space are favored by ANCHOR, and how do these regions contribute to its SUN and MSUN improvements? Such analyses would help establish whether ANCHOR learns chemically meaningful composition preferences rather than merely shifting the statistical distribution of generated formulas.
  3. Does the Composition Policy Exploit MLIP-Specific Evaluation Biases? Tables 14 and 15 show substantial degradation in ANCHOR's stability-related metrics under LeMat-GenBench. Since ANCHOR directly optimizes its composition policy using MLIP-based evaluations of CSP-generated structures, it remains unclear whether this degradation results from differences in evaluation protocols or from the policy favoring chemical regions that receive overly favorable scores from MatterSim. Could the authors analyze stability differences across composition categories, such as element combinations, stoichiometries, and material families, under MatterSim and the MLIP ensemble used by LeMat-GenBench? In particular, could the authors further disentangle the effects of validity criteria, MLIP energy predictions, structural relaxation, and reference convex hulls while controlling for the generated candidates? This would help determine whether the observed performance reversal reflects overoptimization toward particular regions of chemical space.
  4. Does CAN Primarily Benefit from Composition-Level Repetition Suppression? CAN evaluates RDF-based structural similarity only among crystals sharing the same reduced formula, whereas the count bonus already explicitly penalizes repeated formula visits. Could the authors replace CAN with a reward based solely on formula visitation history and compare SUN and MSUN while keeping all other settings unchanged? Additionally, could the authors report how frequently CAN actually performs history-based RDF comparisons during training and quantify their contribution to the GRPO advantage estimates? These analyses would help determine whether structural novelty information provides meaningful learning signals beyond composition-level repetition suppression.
  5. Is Continuous Structural Novelty Necessary for CAN? Table 1 shows that CAN substantially outperforms the binary novelty baseline. However, Appendix F.3 indicates that the binary baseline does not use the same persistent history index, making it difficult to attribute the improvement specifically to continuous RDF-based similarity. Could the authors compare CAN with a history-aware binary structural novelty reward under identical training settings? This would help determine whether continuous structural similarity provides additional benefits beyond history-based duplicate detection.

[1] Dong, Rongzhi, et al. "Generative design of inorganic compounds using deep diffusion language models." The Journal of Physical Chemistry A 128.29 (2024): 5980-5989.

[2] Yang, Sherry, et al. "Generative hierarchical materials search." Advances in Neural Information Processing Systems 37 (2024): 38799-38819.

[3] Karpovich, Christopher, Elton Pan, and Elsa A. Olivetti. "Deep reinforcement learning for inverse inorganic materials design." npj Computational Materials 10.1 (2024): 287.

[4] Khastagir, Subhojyoti, et al. "LLM meets diffusion: a hybrid framework for crystal material generation." Advances in Neural Information Processing Systems 38 (2025): 169190-169215.

5 comments
AuthorsOpenReview ↗

Follow-up to Reviewer WDp6

Follow-up to Q5: binary flag + Adaptive History

Thank you again for suggesting this control. Below you will find the results of training with the binary cell of Table 1 (count bonus and sub-group mask on) with the adaptive history CAN implements, everything else unchanged (121.6k structures, three sampling seeds). MSUN / SUN:

  • Binary flag (Table 1): 25.0 / 8.9%
  • Binary flag with history (new): 43.1 ± 1.9 / 18.4 ± 0.5%
  • Full method (CAN) (Table 1): 47.6 ± 2.1 / 22.1 ± 1.2%

Under the MatterGen protocol the new cell gives 37.2 / 13.5% against 41.3 / 16.0% for the full method (CAN).

So the history is what mattered most: it closes about 80% of the gap. That is slightly more than we expected. It also means that a much sparser signal than we argued for in our previous reply to W2 carries most of the gain: the history-aware flag varies in only 5% of groups (18% in the first fifth), against 51% for CAN, and that is largely enough.

The continuous score adds the remaining ~4 points on top of the adaptive history, and it broadens the search: at the same budget CAN visits 14% more element systems (51.2k against 44.8k), with the same number of distinct metastable formulas within 3%.

We will add the cell to Table 1 and attribute the main gain to the adaptive history in Sec. 3 and 4.

If these results address your concerns, we would be grateful if you would consider updating your assessment. If something is still open, please tell us what, and we will answer before the discussion ends on 18 November.

The authors have addressed most of my concerns in their response, but the corresponding revisions have not been incorporated into the manuscript. I will therefore maintain my current score.

AuthorsOpenReview ↗

We will upload the updated manuscript ASAP.

Thank you. We are glad our responses addressed most of your concerns.

We note that all new results necessary to evaluate the changes are already reported in full in this thread, but of course the revision will move them into the manuscript.

We will upload the revised PDF before the discussion period ends on 18 November, and the pdfdiff will show all changes. The revision will include:

  • Table 1: the history-aware binary cell, and captions that state what each cell isolates.
  • Table 2 or appendix: the genetic algorithm, substitution, Crystal-Text-LLM, Diffusion-LM and DNG + history filter baselines.
  • Sec. 3 and 4: the main gain from CAN attributed to the adaptive history, and the continuous score to the remaining gain and broader search.
  • Main text: the LeMat-GenBench caveat, with the factor-by-factor decomposition in the appendix.
  • App. F: composition distributions for MP-20, Crystalite-DNG and ANCHOR, the per-family hull shifts, and the frequency of history-based comparisons.

We will post a short note here when the revision is up.

Since you note that our responses addressed most of your concerns and that the remaining issue is incorporating them into the manuscript, we would be grateful if you would update your score once the revision is uploaded.

If any of your concerns are still open, please tell us which, and we will address them in the same revision.

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

The joint distribution over crystal compositions and structures can be factorized into a composition distribution and a conditional structural distribution. Most diffusion-based de novo crystal generation (DNG) models learn these two components jointly, such that reward-guided optimization may simultaneously alter both distributions, potentially disrupting the learned structural prior and limiting effective materials discovery. Although existing methods attempt to mitigate this issue through techniques such as KL regularization, these constraints may also restrict the exploration of novel chemical compositions. Furthermore, jointly optimizing the novelty, stability, and diversity of generated materials remains a major challenge for practical materials discovery. To address these challenges, this paper proposes ANCHOR, a reinforcement learning-based framework that decouples chemical composition search from crystal structure generation. Specifically, ANCHOR employs GRPO to optimize a composition search policy while keeping the pretrained crystal structure prediction (CSP) model frozen to preserve its structural prior. It further introduces Continuous Adaptive Novelty (CAN), a history-aware novelty reward that, together with stability and diversity objectives, encourages the continued exploration of novel and potentially stable crystal materials. The proposed method achieves state-of-the-art MSUN performance under the MatterGen evaluation pipeline and demonstrates transferability across different CSP backbones.

Critical strengths
  1. Beyond proposing the ANCHOR framework, this work employs a series of controlled experiments to investigate a fundamental question in generative materials discovery: To what extent is discovery performance limited by chemical composition search versus crystal structure generation, and where should reinforcement learning optimization be applied? By disentangling these two factors under controlled experimental settings, the study provides valuable empirical insights into the performance bottlenecks of generative materials discovery and the appropriate targets for reward-driven optimization. This represents a meaningful methodological contribution.
  2. The paper introduces Continuous Adaptive Novelty (CAN), a reward mechanism that incorporates both known crystal structures and the evolving search history to discourage redundant exploration and improve the discovery rate of valid, novel, and stable crystal candidates. The effectiveness of its key design components is further supported by ablation studies.
Critical weaknesses
  1. Limited robustness across evaluation pipelines. The reported performance of ANCHOR appears highly sensitive to the evaluation pipeline. Specifically, ANCHOR achieves an on-hull SUN of 16.0% under the MatterGen evaluation pipeline, but only 0.4% under LeMat-GenBench with the MP-20 reference hull, compared with 9.8% for Crystalite-DNG under the latter evaluation. Although the absolute SUN values are not directly comparable across evaluation pipelines due to differences in interatomic potentials, structural relaxation procedures, validity criteria, and reference hulls, the reversal in relative model rankings raises concerns about the robustness of ANCHOR's performance advantages. While the authors acknowledge and discuss these factors, the results still suggest that the reported improvements are strongly dependent on the evaluation setting, potentially limiting the reliability and generalizability of the framework for practical materials discovery.
  2. Insufficient comparison with alternative composition search methods. The paper compares ANCHOR against random composition sampling, charge-neutral composition sampling, DNG baselines, and reward-fine-tuned DNG models. However, it lacks direct comparisons with alternative composition search or optimization algorithms under the same CSP backbone, reward function, and computational budget. Consequently, although the experiments demonstrate the effectiveness of ANCHOR's composition search strategy relative to the evaluated baselines, they do not sufficiently establish whether GRPO offers advantages over other established search and optimization approaches.
Questions for authors

The authors mention in the appendix that 5,731 out of 5,737 valid structures in the base-model selection experiments received the default novelty score because their reduced chemical formulas were absent from MP-20. This observation suggests that, for most of these samples, the internal novelty metric primarily reflects compositional novelty rather than explicitly measuring the novelty of underlying crystal structures or structural prototypes. Could the authors further distinguish and evaluate compositional novelty and structural novelty, for example, by assessing whether the generated candidates exhibit genuinely novel structural prototypes beyond previously known chemical compositions?

2 comments
AuthorsOpenReview ↗

Follow-up to Reviewer vnDe

Two short updates. If you would like to have a look at our code, we have anonymised it and made it available here: https://anonymous.4open.science/r/anchor-code-78C8/. Furthermore the history-aware binary control requested by Reviewer WDp6 (Q5) has finished: 43.1% MSUN / 18.4% SUN against 47.6 / 22.1% for the full method (CAN) and 25.0 / 8.9% for the plain binary flag, so most of CAN's gain comes from the run's adaptive history, which refines point 2 of our answer to your question.

Dear Reviewer vnDe,

  • The revised manuscript is online and covers your three points: Sec. 4 now attributes the LeMat-GenBench gap to the validity gate and the hull construction (decomposition in App. E.3.2).
  • The genetic algorithm and substitution baselines are in Table 2 (App. F.7)
  • Sec. 3 states that CAN compares structures within a reduced formula and that we do not claim new prototypes.

If these changes resolve your concerns, we would be grateful if you would consider updating your assessment. If anything is still open, please tell us and we will answer before the deadline.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Note to all reviewers: revised manuscript uploaded

We have uploaded a revised manuscript. The changes are additions and a few rewordings; existing sections and Tables 1–20 keep their numbers (the former Tables 21–23 are now 22, 25 and 26).

Reviewer vnDe

  • W1 (evaluation pipeline): Sec. 4 now attributes the LeMat-GenBench drop to the validity gate and the hull construction, with the factor-by-factor decomposition in App. E.3.2. The same appendix shows that the gap is not inherent to separating composition search from structure generation.
  • W2 (other composition-search methods): ionic substitution and a genetic algorithm with the same reward and budget are in Table 2, with details in App. F.7.
  • Q (CAN): Sec. 3 states that CAN compares structures within a reduced formula and that we do not claim new structure prototypes.

Reviewer WDp6

  • Q5 / W2: Table 1 has the new row "+ history" (binary flag with the adaptive history), and its caption says what each cell isolates. Sec. 4 attributes the main gain of CAN to the adaptive history. App. F.3 adds the analysis and Table 21 (novelty signal and frequency of history comparisons during training).
  • Q1: substitution, genetic algorithm and DNG + history filter are in Table 2; Crystal-Text-LLM and diffusion language-model compositions are in App. F.7 (Tables 27 and 28). The requested citations are in App. B and App. F.7.
  • Q2: composition distributions of MP-20, Crystalite-DNG and ANCHOR (Table 23), where the policy searches, and steering with the element mask are in App. F.4.
  • Q3: the decomposition into gate, relaxer, potential and hull construction is in App. E.3.2, and the per-family table is Table 24 in App. F.4.
  • W1: the joint validity gate is in App. E.3.1.

Reviewer D4qD

  • W1: overview paragraph at the start of Sec. 3; Table 1 caption.
  • W2: Sec. 3 states that failures of the frozen CSP mislead the policy in one direction only; App. F.4 quantifies the frozen CSP out of distribution; Sec. 5 notes it as a limitation.
  • W3: App. D.3 (MatterSim as a proxy, including a per-structure check with ORB-v3), App. D.8 (compute breakdown), App. F.8 (yield against budget, Fig. 3, also in GPU-hours).
  • W4: DNG + history filter rows in Table 2, details in App. F.7.
  • W5: the additional generators are discussed with their evaluation protocols in App. B (for space, there and not in Sec. 2); Table 13 caption; Tables 27 and 28; Table 3 now lists the Stable and SUN values reported for MatterGen+REPA.
  • W6: the anonymised code link is in the Reproducibility Statement.
  • W7: sampling settings at the start of Sec. 4.

Second upload, before 18 Nov: MatterGen+REPA sampled from its released checkpoint and scored under our pipeline, with and without the history filter (requested by Reviewer D4qD); a revised presentation of the motivation and of Sec. 3.

AuthorsOpenReview ↗

Second revision uploaded

As promised, the second revision is now online. It mainly changes how the paper is written and adds one baseline.

New result

  • MatterGen+REPA, sampled by us from its released checkpoint and scored under our pipeline, with and without the history filter (Table 27, App. F.7). Requested by Reviewer D4qD.

Aside from that, we have mainly worked on presentation, which all three reviewers asked for:

  • Sec. 1: the introduction now builds the motivation in three steps: 1. why the compositions a generative model proposes are not the ones discovery should search. 2. why fine-tuning the whole model with a reward does not fix this. 3. why the problem is worst for novelty rewards.
  • Sec. 2 is made shorter and less repetitive. It also now defines MP-20, E_hull and what "on the hull" means, and leaves the argument to Sec. 3.
  • Sec. 3: the motivation is now three numbered reasons, each with a one-line heading in bold, which hopefully makes it easier to read at a first glance. There are also some minor changes, like GRPO being spelled out at first mention, an explanation of the policy's state, and the validity gate being described where it first appears. The overview points to the full training algorithm (Algorithm 1).
  • Sec. 4: overall readability improvements in both phrasing and paragraph splitting
  • Sec. 5: the limitations now have their own paragraph.

Thank you all for the time you have put into reviewing this paper. We hope the comments made it clearer and better.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.