acceptodds
Under review as a conference paper at ICLR 2027

Atomistic Language Models Understand and Generate Materials

Abstract

Atomistic structure and natural language have long been modeled separately, with language models either calling atomistic models as tools or being fine-tuned on lossy textual encodings that discard atomistic information. We introduce **Atomistic Language Models (ALMs)** to pursue native multimodality, in which a single language backbone understands atomistic structures, generates materials from natural language, and optimizes crystal structures as instructed by text. By unifying a pretrained atomistic encoder, large language model, and denoising diffusion model through purely continuous projectors and staged training, ALMs achieve state-of-the-art results on crystal structure prediction and *de novo* generation. ALMs are enabled by a *continuous bridge that maps language model embeddings directly into the steering space of atomistic diffusion*, and are assisted by **Text-to-Crystal Feynman–Kac (T2C-FK)**, a particle-based sampler that scores partial denoising trajectories to enforce stoichiometric targets at inference time. We also introduce **ALM Bench**, the first benchmark for language-instructed crystal generation and materials optimization. ALMs beat frontier LLMs, including GPT-5.6 Sol, on inorganic crystalline doping, inverse design, and generation for desired applications. We release all code, training data, and model weights.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multimodal machine learning, materials inverse design, large language models, inorganic crystals

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

This paper proposes a materials model based on large language models (LLMs) that can be used to design materials with specific properties or to predict material performance. The authors aim to construct a unified model for structures and text; the "producer-consumer" bridging mechanism they introduce is of significant practical value.

Critical strengths
  1. Unify sequence- and structure-based representations in materials science.
  2. Introduce an additional prompt-based benchmark for evaluation.
Critical weaknesses

see my question

Questions for authors
  1. I wonder whether T2C-FK performs better than a simple Best-of-N sampling strategy. As shown in Table 2. As shown in Table 4, ALM does not consistently outperform GPT-5.6-sol on OOD tasks. What might explain this relatively limited generalization performance?3. what the effect fo simple prompt for mattergen? 3. How would MatterGen perform with a simpler prompt-based conditioning approach? I believe this comparison is important to demonstrate whether the proposed LLM-based framework offers meaningful advantages over simpler conditional generation methods.
4 comments
AuthorsOpenReview ↗

Response to Reviewer hMvF's comments (#1)

We thank the reviewer for recognizing the value of unifying language and structural representations for materials science and of a benchmark for prompt-based materials design. We agree that the benefits of T2C-FK and of learned language-to-structure conditioning should be tested against simpler alternatives, and that ALM's OOD performance needs a clearer explanation. To address these questions, we compared T2C-FK with best-of-64 sampling, analyzed OOD errors and tested an external composition planner, and evaluated language models coupled to the released conditional MatterGen checkpoints. We present the setup and results of each experiment below.

T2C-FK versus Best-of-N sampling

[Q1] I wonder whether T2C-FK performs better than a simple Best-of-N sampling strategy.

We compared T2C-FK with best-of-64 sampling using the same ALM Gen checkpoint and the same budget of 64 generation trajectories. T2C-FK steers 64 particles during denoising, while best-of-64 generates 64 unsteered samples and ranks them with the same stoichiometry reward. As shown in Table R9, T2C-FK raises MR@20 from 8.7% to 41.0% on MP-20 and from 2.8% to 10.0% on MPTS-52.

Table R9. ALM Gen CSP with and without T2C-FK (reproduced from our response to Reviewer xtzu; N=1000, K=64). Match rates in %.

MethodMP-20 MR@1 ↑MP-20 MR@20 ↑MP-20 MR@64 ↑MPTS-52 MR@1 ↑MPTS-52 MR@20 ↑MPTS-52 MR@64 ↑
ALM Gen, T2C-FK off0.33.98.70.00.82.8
ALM Gen, best-of-646.18.78.72.02.82.8
ALM Gen + T2C-FK22.341.044.26.010.010.7

T2C-FK is 3.7 times (MR@1) and 4.7 times (MR@20) better than best-of-64 on MP-20, and 3.0 times and 3.6 times better on MPTS-52, whose structures are up to 2.6 times larger. Best-of-N can only select among samples the model already produces. Without steering, just 1.4% of samples have the target stoichiometry, and 72% of compositions have none among 64. T2C-FK instead resamples particles toward the target during denoising, so it reaches compositions that the unsteered sampler almost never produces.

AuthorsOpenReview ↗

Continuation of response to Reviewer hMvF's comments (#2)

Generalization relative to GPT-5.6 Sol

[Q2] As shown in Table 4, ALM does not consistently outperform GPT-5.6-sol on OOD tasks. What might explain this relatively limited generalization performance?

We first corrected an evaluation error. A bug had run Sol at the API's default (medium) reasoning effort instead of the stated low effort. Table 4 now reports Sol at correctly applied low effort (Table R13; both effort levels are in Table R12 of our response to Reviewer xtzu).

Table R13. Updated Table 4 (N=5×200, 95% CI over seeds).

ModelApplication ↑Describe (Comp.) ↑Describe (Struct.) ↑OOD (Comp.) ↑OOD (Struct.) ↑
GPT-4.10.224±0.0320.254±0.0190.090±0.0180.228±0.0290.050±0.005
GPT-5.20.252±0.0510.356±0.0140.162±0.0130.366±0.0260.108±0.008
GPT-5.6 Sol (low)0.302±0.0330.748±0.0290.468±0.0350.675±0.0210.305±0.019
ALM Edit0.323±0.0200.730±0.0310.412±0.0410.641±0.0320.316±0.022

To explain the remaining gap, we sorted every OOD prompt into one of five outcomes.

Table R14. Where OOD prompts fail (fraction of prompts, each row sums to 1).

ModelCorrect ↑Right composition, wrong structureWrong ratio of elements ↓Wrong elements ↓No structure ↓
GPT-5.20.1080.2580.3380.0090.287
GPT-5.6 Sol (low)0.3050.3700.2550.0030.067
GPT-5.6 Sol (medium)0.3570.3390.2640.0030.037
ALM Edit0.3160.3250.2950.0100.054

No model chooses the wrong elements more than about 1% of the time. ALM Edit's gap to Sol comes almost entirely from choosing the wrong stoichiometric ratio (0.295 vs 0.255). Because ALM Edit's language model writes the target composition before the diffusion model generates the structure, this reflects the underlying Qwen3-8B failing to recall the right formula rather than a failure of the bridge or the decoder. When ALM Edit gets the composition right, it finds the correct structure about as often as Sol.

To test this, we replaced only the composition-planning step with a newer, more knowledgeable language model (Qwen3.5-9B), leaving ALM Edit unchanged (Table R15).

Table R15. Replacing ALM Edit's composition planner (N=5×200, 95% CI over seeds).

ModelApplication ↑Describe (Comp.) ↑Describe (Struct.) ↑OOD (Comp.) ↑OOD (Struct.) ↑
GPT-5.6 Sol (low)0.302±0.0330.748±0.0290.468±0.0350.675±0.0210.305±0.019
GPT-5.6 Sol (medium)0.263±0.0670.728±0.0220.479±0.0040.696±0.0300.357±0.026
ALM Edit0.323±0.0200.730±0.0310.412±0.0410.641±0.0320.316±0.022
ALM Edit + Qwen3.5-9B planner0.512±0.0670.864±0.0190.329±0.0160.780±0.0260.371±0.030

With the stronger planner, ALM Edit beats both Sol settings on four of the five tasks, including OOD composition (0.780 vs 0.696) and structure (0.371 vs 0.357). The generalization gap therefore tracks the knowledge of the language model ALM is built on, and it shrinks as that model improves, without retraining the bridge or decoder.

AuthorsOpenReview ↗

Continuation of response to Reviewer hMvF's comments (#3)

MatterGen with prompt-based conditioning

[Q3] How would MatterGen perform with a simpler prompt-based conditioning approach? I believe this comparison is important to demonstrate whether the proposed LLM-based framework offers meaningful advantages over simpler conditional generation methods.

We agree that this comparison is central. We have added several additional baselines accordingly: a language model reads each prompt and writes a chemical system (e.g., "Li-Fe-P-O"), which conditions the released MatterGen [1] checkpoints, either with chemical_system alone or with chemical_system plus an energy-above-hull target of zero. MatterGen cannot take an exact count of each element or a structure to edit, so only the generation tasks are possible with this approach. We use two language models: Qwen3-8B, the same model ALM is built on, which isolates the effect of ALM's learned conditioning interface from the language model itself, and GPT-5.2, which has considerably broader materials knowledge.

Table R16. Prompt-conditioned MatterGen vs ALM Edit (N=5×200, 95% CI over seeds).

ModelApplication ↑Describe (Comp.) ↑Describe (Struct.) ↑OOD (Comp.) ↑OOD (Struct.) ↑
Qwen3-8B → MatterGen (chemical system)0.344±0.0450.080±0.0150.029±0.0060.058±0.0100.020±0.003
Qwen3-8B → MatterGen (chemical system + Ehull=0)0.360±0.0280.092±0.0200.028±0.0130.056±0.0140.019±0.007
GPT-5.2 → MatterGen (chemical system)0.398±0.0460.083±0.0070.020±0.0100.052±0.0210.017±0.013
GPT-5.2 → MatterGen (chemical system + Ehull=0)0.446±0.0200.091±0.0100.031±0.0110.063±0.0170.026±0.008
ALM Edit0.323±0.0200.730±0.0310.412±0.0410.641±0.0320.316±0.022

With either composition planner (Qwen3-8B or GPT-5.2), the language model chooses the correct element set for at least 99.9% of Describe prompts, yet composition and structure recovery fall to 0.02 to 0.09, 7 to 14 times below ALM Edit. The language model is therefore not the bottleneck; MatterGen's published conditioning interface is. MatterGen's fields can name the elements but not their ratio, the structure type, or a requested change, so a request for SrTiO₃ becomes "some Sr-Ti-O compound". Prompt-conditioned MatterGen scores above ALM Edit only on Application consistency, where the judge mainly rewards a plausible choice of elements (0.360 vs 0.323 with Qwen3-8B, within CIs, and 0.446 with GPT-5.2). It also cannot perform any of the nine editing tasks, because MatterGen cannot take an input structure. ALM's producer-consumer bridge closes this gap by passing the language model's hidden states, which encode composition, structure, and the requested change, into every block of the diffusion model instead of reducing the request to a few condition labels.

We will add these sampling controls, the OOD error analysis, and the prompt-conditioned MatterGen comparison to the revised manuscript, and we hope they address the reviewer's questions and inspire a reconsideration of our score.

[1] Zeni, C. et al. A generative model for inorganic materials design. Nature 639, 624-632 (2025).

Thanks for the detailed responses and additional experiments. I appreciate the effort you've put into addressing my concerns, and I'll raise my score accordingly.

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

This work introduces Atomistic Language Models (ALMs) which combines continuous space encoders like OrbV3 with discrete token based Language Models like Qwen3 and a diffusion-based decoder (MatterGen) for generating outputs in the crystal space. They also introduce a new benchmark for language-instructed editing and generation of materials called ALMBench.

Critical strengths
  1. The paper is well written and easy to follow
  2. The problem is well motivated and relevant to current literature. No existing model can perform consistently across the range of tasks as demonstrated by ALM in table 1.
  3. Extensive ablation studies justifies encoder choice, LoRA fine tuning, etc.
Critical weaknesses
  1. ALM claims to be a single backbone model, while there are two separate variants of the model ALM Edit and ALM Gen with separate checkpoints.
  2. Table 2 claims to achieve SoTA performance on MP-20 and MPTS-52 datasets at the CSP task, but the MR@1 does not achieve the best results for either dataset.
  3. The authors have not shown a comparison between ALM Gen and ALM Gen without T2C-FK. It would help in understanding the effectiveness of T2C-FK.
  4. Section 2.2 states that ALM Edit was trained on all tasks in ALM Bench, but the other LLMs were zero-shot prompted. This might not be a fair comparison. Yet, in table 4, GPT-5.6 Sol is able to perform better on three tasks.
Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer xtzu's comments (#1)

We thank the reviewer for recognizing the paper's motivation, clear presentation, and ablations supporting the encoder and fine-tuning choices. We agree that the relationship between ALM Edit and Gen, the scope of our crystal structure prediction claims, and the fairness of the baseline comparisons need to be clearer. Below, we clarify that Edit and Gen are separately fine-tuned variants derived from ALM Core, but whose capabilities are unified into a single checkpoint (as per our headline claim) by T2C-FK; describe our state-of-the-art claim of MR@20 and re-scope it in the Results section; and report ALM Gen results without T2C-FK and against best-of-64 sampling. We also expand the comparisons with few-shot, fine-tuned, and frontier language-model baselines, as well as correct the reasoning-effort settings used in the evaluation. We will reflect these clarifications and comparisons in the revised manuscript.

A shared language backbone with two generation heads

[W1] ALM claims to be a single backbone model, while there are two separate variants of the model ALM Edit and ALM Gen with separate checkpoints.

We thank the reviewer for raising this, and agree that our wording should be more precise about what "single backbone" refers to. Crystal structure prediction (CSP), which predicts the structure of a given composition, and de novo generation (DNG), which proposes new, stable materials, are fundamentally different use cases. Following the conventional practice of training separate models for the two (e.g., FlowMM [1], DiffCSP [2], MatterGen [3]), we release ALM Edit for composition-controlled generation and editing, and ALM Gen for de novo generation. As discussed in Appendix B.2.2, this split is because of a deficiency of our denoising diffusion model, rather than in the ALM architecture. At g=0, the de novo MatterGen [3] decoder reaches 0.750 metastability but only 0.370 CSP Match@128, while the CSP decoder reaches 0.777 Match@128 but only 0.167 metastability.

Our central point in calling ALM a single-backbone model is architectural. In both variants, one model reads the input (text and crystal structure through the OrbV3 encoder) and steers denoising: the language model's hidden states condition the diffusion decoder through the learned bridge. Both are trained end-to-end, with the diffusion loss backpropagating through the bridge into the language model. Both variants are fine-tuned from the same ALM Core; neither stitches together separately trained components. We discuss when to use each variant, and how to tune conditioning strength at inference, in our response to Reviewer UtRc.

T2C-FK shows that the two capabilities can also be served by a single checkpoint and a single decoder, establishing the headline claim. Applied to ALM Gen at inference time, T2C-FK enforces the requested composition during denoising. This turns the de novo model into a CSP model without retraining: MR@20 rises from 3.9% to 41.0% on MP-20 (Table R9). The unified model currently trades some CSP accuracy for this generality, relative to ALM Edit's dedicated decoder (MR@20 of 80.4% for the full model).

In response to the reviewer's concern, we have revised the manuscript to clarify that ALM shares a single language backbone, with two task-specialized generation heads that T2C-FK can unify at inference time.

Effect of T2C-FK

[W3] The authors have not shown a comparison between ALM Gen and ALM Gen without T2C-FK.

We thank the reviewer for this suggestion. We ran ALM Gen on CSP with T2C-FK turned off, keeping the same checkpoint, element mask, guidance, and K=64 samples (Table R9). We also re-ranked those same 64 unsteered samples by the stoichiometry reward T2C-FK uses (best-of-64). This gives a compute-matched alternative, since T2C-FK also uses 64 particles.

Table R9. ALM Gen CSP with and without T2C-FK (N=1000, K=64). Match rates in %.

MethodMP-20 MR@1 ↑MP-20 MR@20 ↑MP-20 MR@64 ↑MPTS-52 MR@1 ↑MPTS-52 MR@20 ↑MPTS-52 MR@64 ↑
ALM Gen, T2C-FK off0.33.98.70.00.82.8
ALM Gen, best-of-64 (stoichiometry rerank)6.18.78.72.02.82.8
ALM Gen + T2C-FK22.341.044.26.010.010.7

Without T2C-FK, only 1.4% of MP-20 samples have the target stoichiometry, and only 27.9% of compositions have even one such sample among 64. Re-ranking therefore cannot recover what the unsteered sampler rarely produces. At an equal sample budget, T2C-FK improves MR@1 by 3.7× and MR@20 by 4.7× over best-of-64 on MP-20, and by 3.0× and 3.6× on MPTS-52.

[1] Miller, B. K. et al. FlowMM: generating materials with Riemannian flow matching. ICML (2024).

[2] Jiao, R. et al. Crystal structure prediction by joint equivariant diffusion. NeurIPS (2023).

[3] Zeni, C. et al. A generative model for inorganic materials design. Nature 639, 624–632 (2025).

AuthorsOpenReview ↗

Continuation of response to Reviewer xtzu's comments (#2)

Crystal structure prediction should reward covering plausible polymorphs

[W2] Table 2 claims to achieve SoTA performance on MP-20 and MPTS-52 datasets at the CSP task, but the MR@1 does not achieve the best results for either dataset.

We thank the reviewer for this observation; it is correct that ALM Edit does not achieve the best MR@1 on either benchmark, and we have revised our claim accordingly in the Results section. However, we believe, as recent literature has supported, that MR@1 alone mis-measures what an experimentally useful generative model should do on this task.

The input to CSP is a composition, and many valid crystal structures can exist for a given composition at finite temperature. MP-20 and MPTS-52 nevertheless designate a single structure as the label. As Martirossyan et al. [4] show, this label is often an arbitrary choice: it is not necessarily the most stable, most common, or experimentally preferred polymorph. A realistic generative model should therefore not memorize the exact label for each composition (MR@1). Instead, it should learn a distribution of realistic structures that covers the likely choices of labels. When sampled once, such a model will often return a different valid polymorph (potentially far from 100% MR@1). When sampled several times, it should include the reference (100% MR@20). MR@20 measures exactly this, and it is where ALM Edit leads on both benchmarks (83.2% on MP-20 and 45.7% on MPTS-52, against 78.3% and 41.5% for the strongest baselines).

To make this rigorous rather than rely on one match-rate column, we also report the metrics Martirossyan et al. [4] propose:

  • METRe: the fraction of reference structures matched by any generated structure.
  • cRMSE: counts each unmatched reference at the matching tolerance (stol = 0.5), so a model cannot score well by matching only a few references very closely.

We evaluate both at k=1 and k=20 (Table R8).

Table R8. METRe and cRMSE at k=1 and k=20 (per-composition estimates from Table 2 values; cRMSE = METRe × (RMSE − 0.5) + 0.5) [4]. Bold and italic denote best and second best.

ModelMP-20 METRe@1 (%) ↑MP-20 cRMSE@1 ↓MP-20 METRe@20 (%) ↑MP-20 cRMSE@20 ↓MPTS-52 METRe@1 (%) ↑MPTS-52 cRMSE@1 ↓MPTS-52 METRe@20 (%) ↑MPTS-52 cRMSE@20 ↓
CDVAE33.90.36667.00.2345.30.48520.80.439
DiffCSP51.50.27577.90.14912.20.46134.00.389
FlowMM61.40.228——17.50.443——
CrystaLLM-large58.70.23074.00.15619.20.42533.80.367
CrystalFlow62.00.23478.30.15421.00.42937.80.371
OMatG63.80.227——25.20.423——
MCFlow-L64.10.21676.10.14927.20.40241.50.346
ALM Edit43.70.28980.40.11022.70.39145.70.289

At k=1, ALM Edit trails the strongest baselines on MP-20, as the reviewer notes. At k=20, it achieves the best METRe and cRMSE on both benchmarks. When it is allowed several samples, it both matches the most references and matches them most closely.

The distribution-learning interpretation is also supported by a stricter test. If ALM Edit's MR@20 came from memorizing labels, it should collapse when test compositions are removed from training. We retrained on composition-split versions of MP-20 and MPTS-52, in which no test composition appears in any training stage. MR@20 instead increased, to 85.1% on MP-20 and 49.8% on MPTS-52 (Table R1 in our response to Reviewer UtRc). The model generalizes polymorph coverage to compositions it has never seen.

We have revised the accompanying text for Table 2 to state this claim precisely. We also report the data-matched control from our response to Reviewer UtRc (Table R2), which separates the contribution of our broader training data from that of the architecture.

[4] Martirossyan, M. M. et al. All that structure matches does not glitter. NeurIPS Datasets and Benchmarks Track (2025).

AuthorsOpenReview ↗

Continuation of response to Reviewer xtzu's comments (#3)

Stronger baselines for ALM Bench

[W4] ALM Edit was trained on all tasks in ALM Bench, but the other LLMs were zero-shot prompted. This might not be a fair comparison. Yet, in table 4, GPT-5.6 Sol is able to perform better on three tasks.

We agree that zero-shot frontier models alone do not establish a fair comparison and added three controls:

  1. Few-shot frontier model: GPT-5.2 given three solved examples from the same task's training split.
  2. Fine-tuned language model: Qwen3-8B (ALM's base model) fine-tuned to read and write CIFs on the same task pairs ALM Edit was trained on.
  3. Correct reasoning effort: a bug ran GPT-5.6 Sol at API-default (medium) instead of low reasoning effort; we fixed it and re-ran both levels (Table R12).

For the first two, ALM Edit is retrained on our composition-split data (Appendix C.3), and all models are scored on the same held-out test prompts, unseen in training.

Table R10. ALM Bench editing tasks against few-shot and fine-tuned baselines (held-out by composition test pool, N=5×200).

ModelEf↑Ef↓ρ↑ρ↓V↑V↓Poly. ↑Doping ↑Strain ↑
GPT-5.2 (zero-shot)0.3890.3680.0570.2230.0140.0620.0450.0000.000
GPT-5.2 (3-shot)0.3800.5780.0870.0920.0000.0480.1010.0050.000
Qwen3-8B (base)0.2290.2610.0000.0280.0560.0250.0060.0040.002
Qwen3-8B (fine-tuned)0.6500.6480.0240.1040.0110.0680.0830.6010.105
ALM Edit0.4020.7770.4880.5450.5320.3800.1700.9270.150

Table R11. ALM Bench generation tasks against few-shot and fine-tuned baselines (held-out by composition test pool, N=5×200).

ModelApp. ↑Desc. comp ↑Desc. struct ↑OOD comp ↑OOD struct ↑
GPT-5.2 (zero-shot)0.2780.3270.1410.4140.124
GPT-5.2 (3-shot)0.2860.4300.1920.3460.123
Qwen3-8B (base)0.0240.0210.0040.1140.009
Qwen3-8B (fine-tuned)0.0340.6480.4250.7930.452
ALM Edit0.0600.6590.2700.8110.352

Fine-tuned Qwen3-8B is a strong baseline, beating every GPT row on most columns and recovering structures from text better than ALM Edit (Describe and OOD structure). However, it does not learn the editing tasks, where ALM Edit leads on Ef↓ and polymorph by 0.13 and 2×, on doping by 0.33 (0.927 vs 0.601), and on strain. Three examples help GPT-5.2 on Ef↓, polymorph, and Describe, but it still relabels the input atoms instead of doping. ALM Edit's large lead on density and volume partly reflects producing valid, distinct structures, which LLM-written CIFs often fail to do.

Table R12. Effect of reasoning effort on GPT-5.6 Sol (held-out test pool as in the manuscript, N=5×200).

ModelEf↑Ef↓ρ↑ρ↓V↑V↓Poly. ↑Doping ↑Strain ↑App. ↑Desc. comp ↑Desc. struct ↑OOD comp ↑OOD struct ↑
GPT-5.6 Sol (low)0.5160.5600.1930.1990.0040.0060.3200.0020.0010.3020.7480.4680.6750.305
GPT-5.6 Sol (medium)0.5130.6360.2200.2380.0500.0280.4440.0040.0000.2630.7280.4790.6960.357
ALM Edit0.5860.6260.3800.3700.4510.3550.2240.8790.1510.3230.7300.4120.6410.316

Correctly applied low effort makes Sol slightly weaker on editing and slightly stronger on Application; ALM Edit's margins are unchanged or larger, and Table 4 now uses low-effort Sol. Where Sol still leads (Describe and OOD), our failure-mode analysis (Table R14, response to Reviewer hMvF) shows that neither Sol nor ALM Edit gets the element set wrong (≤1%), but ALM Edit more often chooses the wrong stoichiometric ratio (0.295 vs 0.255). That choice is made by the underlying Qwen3-8B rather than by the ALM architecture. Replacing only that composition-planning step with a newer Qwen3.5-9B lets ALM Edit beat Sol on four of the five generation tasks (Table R15, response to Reviewer hMvF).

We hope these controls and corrections address the reviewer's concerns about fairness and inspire a reconsideration of our score.

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

This paper introduces Atomistic Language Models (ALMs), a multimodal framework that connects an atomistic encoder, an LLM, and a diffusion decoder through continuous representations to jointly model natural language and three-dimensional inorganic crystal structures. The framework supports property prediction, language-guided crystal editing, and new crystal generation, and introduces both a sampling method for steering generation toward desired chemical compositions and ALM Bench for evaluating language-guided materials design. Experiments report competitive property prediction, higher best-of-20 crystal structure matching rates on MP-20 and MPTS-52, and improvements over evaluated baselines on several editing tasks.

Critical strengths
  1. The paper presents a unified framework that connects a pretrained atomistic encoder, an LLM, and a diffusion-based crystal generator through continuous representations. Unlike prior approaches focused primarily on property prediction or standalone crystal generation, ALM enables language-guided structural editing and generation within a shared architecture. This is a meaningful step toward more flexible, language-driven materials design, even though the individual architectural components are not entirely novel.

  2. The ALM Bench provides a useful evaluation setting for language-guided crystal editing and inverse design, covering tasks such as property optimization, polymorph generation, and element substitution. The paper also presents extensive experiments across property prediction, crystal structure prediction, and de novo generation, supported by a considerable number of architectural and training ablations.

Critical weaknesses
  1. The performance improvements are not disentangled from the effects of additional training data and pretrained components. In particular, Appendix B.2.2 reports that expanding the CSP training data alone improves Match@20 by approximately 4%, which is comparable to ALM Edit's improvement over the strongest baseline. Moreover, the paper does not establish crystal-level separation across its multiple training stages and evaluation datasets. Without stronger data-matched comparisons and clearer guarantees against structural overlap, it is difficult to determine how much of the reported improvement can be attributed to the proposed continuous cross-modal architecture.

  2. While the paper compares against specialized crystal generators on CSP and de novo generation, ALM Bench primarily evaluates language-guided editing against general-purpose LLMs that directly generate CIF files. The evaluation would be more convincing with modular baselines that combine language understanding with established materials tools, such as MatSciAgent [1] and AtomisticSkills [2]. More targeted comparisons could include an LLM coupled with conditional MatterGen [3] for text-guided generation, or MCFlow [4] combined with structural relaxation for substitution tasks. Without such comparisons, it remains unclear whether the proposed joint architecture offers substantial advantages over modular approaches.

  3. The paper does not establish structure-level isolation across its training stages and evaluation benchmarks. There is still a possibility that reported generalization results are affected by overlapping crystal structures.

[1] Chaudhari, Akshat, Janghoon Ock, and Amir Barati Farimani. "Modular large language model agents for multi-task computational materials science." Communications Materials 7.1 (2026): 131.

[2] Deng, Bowen, et al. "Harnessing AtomisticSkills for Agentic Atomistic Research." arXiv preprint arXiv:2605.24002 (2026).

[3] Zeni, Claudio, et al. "A generative model for inorganic materials design." Nature 639.8055 (2025): 624-632.

[4] Seong, Kiyoung, et al. "Multimodal crystal flow: Any-to-any modality generation for unified crystal modeling." arXiv preprint arXiv:2602.20210 (2026).

Questions for authors
  1. ALM Edit and ALM Gen use different diffusion backbones and conditioning mechanisms, resulting in different trade-offs between instruction following and generation quality. In practical materials design scenarios where both exact compositional constraints and high structural stability are required, how should the two variants be used? Is there a way to select between them or adjust the conditioning strength based on the task?

  2. Table 7 reports a learning rate of (2e-4) for ALM Gen's LoRA adaptation, whereas Appendix B.2.4 and Figure 12 identify (1e-5) as the selected learning rate and indicate that (2e-4) causes representation collapse. Could you clarify which configuration was correct?

  3. The input adapter preserves one soft token per atom, making the input sequence length dependent on the number of atoms. What is the practical limit on the number of atoms that ALM can process, and how does its computational cost scale with crystal size? Does the model require any special handling for structures approaching or exceeding the LLM's context limit?

5 comments
AuthorsOpenReview ↗

Response to Reviewer UtRc's comments (#1)

We thank the reviewer for recognizing the value of ALM's unified architecture and ALM Bench for language-guided materials design. We agree that the learned interface must be distinguished from the benefits of additional training data and pretrained components, and evaluated against stronger modular baselines. We retrained ALM Core and Edit with compositions held out across training and evaluation, report controls on training data and language-model adaptation, and compare with agent and tool-based materials workflows, including an agent given access to ALM. We also clarify the intended use of ALM Edit and Gen, the training configuration, and the computational cost of larger structures.

Data separation and training data

[W1, W3] The paper does not establish crystal-level separation across its multiple training stages and evaluation datasets...

All test sets were held out from training by structure, which we now state in the first results paragraph. ALM Bench follows the MP-20 and MPTS-52 convention of splitting by structure. To fully address their concern, we re-partitioned all training data for ALM Core and Edit (including MP-20, MPTS-52, and ALM Bench) by composition, retrained, and evaluated on fully held-out compositions (Appendix C.3). MR@20 remained SoTA, while RMSE to the label structures slightly increased (Table R1).

Table R1. Crystal structure prediction on composition-split MP-20 and MPTS-52 (App. C.3; MR in %; ALM Edit retrained).

ModelMP-20 MR@1 ↑MP-20 RMSE@1 ↓MP-20 MR@20 ↑MP-20 RMSE@20 ↓MPTS-52 MR@1 ↑MPTS-52 RMSE@1 ↓MPTS-52 MR@20 ↑MPTS-52 RMSE@20 ↓
CrystaLLM-large58.700.040873.970.034919.210.111033.750.1059
MCFlow-L64.080.056176.080.038327.160.140141.450.1296
ALM Edit45.20.05585.10.04521.90.07049.80.075

Contribution of data and pretrained components

[W1] Performance improvements are not disentangled from the effects of additional training data and pretrained components.

We separate the training data, the diffusion decoder, and the pretrained language model.

Training data and the diffusion decoder

For CSP, ALM Edit's decoder receives the target composition. We trained it alone (without the language model) on different training sets and evaluated the full ALM Edit on the same benchmark (Table R2).

Table R2. Data-matched CSP control on MP-20 ("decoder" is ALM Edit's CSP diffusion model without the language model; MR in %; weaker baselines omitted).

ModelTraining dataMR@1 ↑MR@20 ↑RMSE@1 ↓
CrystalFlowMP-20 train62.078.30.071
OMatGMP-20 train63.8—0.072
MCFlow-LMP-20 train64.176.10.056
ALM Edit decoderMP-20 train45.077.80.061
ALM Edit decoderBroader corpus (App. C.2)45.683.20.021
ALM Edit decoderBroader corpus, composition-split (App. C.3)45.285.10.055
ALM EditBroader corpus (App. C.2)43.780.40.017

Trained only on MP-20, the decoder matches the strongest baselines (MR@20 77.8 vs 78.3 for CrystalFlow). Our broader corpus adds 5.4 points, which persist under a composition split. Adding the language model and bridge slightly reduces MR@20 (80.4), which still exceeds every baseline. Our data mixture thus drives much of the CSP gain, and our text-conditioning architecture preserves this strength while enabling conditioning on natural language and atomistic structures.

Pretrained language model

We retrained ALM Edit with the language model frozen and fully fine-tuned, on the same composition-split data and held-out ALM Bench prompts (Table R3).

Table R3. Frozen vs fine-tuned language model in ALM Edit (N=5×200; 95% CIs ±0.006–0.054).

ModelEf↑Ef↓ρ↑ρ↓V↑V↓Poly.DopingStrainApp.Desc. compDesc. structOOD compOOD struct
Frozen LLM0.4420.7680.5150.4680.4850.4100.1610.9180.1330.1580.5500.1950.6300.244
Fine-tuned LLM0.4020.7770.4880.5450.5320.3800.1700.9270.1500.0600.6590.2700.8110.352

Editing is mostly within CIs, and both perform substitution (doping ~0.92) that frontier LLMs cannot, so this capability does not depend on adapting the language model. Fine-tuning helps when the model must write the right composition from text (OOD composition 0.630 vs 0.811), while the frozen model retains more general knowledge for application prompts (0.158 vs 0.060). The decoder and its data drive CSP, adapting the language model drives text-to-structure recovery, and language-guided substitution emerges from training the bridge (Table R4).

AuthorsOpenReview ↗

Continuation of response to Reviewer UtRc's comments (#2)

ALM as a learned interface between language and structure generation

[W2] The evaluation would be more convincing with modular baselines... such as MatSciAgent [1] and AtomisticSkills [2]... an LLM coupled with conditional MatterGen [3]... or MCFlow [4] combined with structural relaxation for substitution tasks.

We agree, and present all four comparisons below. However, we also would like to present ALM not as an alternative to agentic tool use, but as a way to learn the interface between a language model and the tools it uses. Instead of passing a request through a tool's fixed input fields, ALM's language-model hidden states condition the structure generator directly, through a bridge trained end-to-end. We therefore compare the standalone systems, then give the same agent access to ALM Edit as a tool.

Agents with and without ALM

Agents had no Materials Project or web access, which could return reference structures. MCFlow has no public code or weights, so for substitution we evaluated the closest open pipelines: CrystaLLM-large followed by MatterSim relaxation, and a direct element swap followed by relaxation.

Table R4. Agent and tool baselines (95% CIs ±0.07–0.25, and n/a indicates that the pipeline cannot take this input).

ModelEf↑Ef↓ρ↑ρ↓V↑V↓Poly.DopingStrainApp.Desc. compDesc. structOOD compOOD struct
GPT-5.20.3680.3230.1470.1880.0000.0360.0600.0000.0000.1600.3800.2400.5200.140
MatSciAgent0.2110.3870.3240.3750.0910.1430.1200.3910.0430.0800.5000.2400.5400.300
CrystaLLM → relax0.3680.6130.5880.4380.4550.3570.4000.7170.196n/an/an/an/an/a
Element swap → relax0.2630.2580.0000.0000.0000.0000.1000.1300.022n/an/an/an/an/a
AtomisticSkills0.7370.8710.2940.1880.1360.1790.8000.0000.0000.5600.9200.6400.8800.520
AtomisticSkills + ALM Edit0.7370.8060.5000.4380.5000.2500.5400.4130.0220.4400.8000.4800.8800.440
ALM Edit0.6320.5480.3820.4380.4550.2860.2800.8910.1960.1200.5800.3200.6200.340

AtomisticSkills, driven by a frontier model, is strongest on formation-energy edits, polymorphs, and text-to-structure recovery, where it can recall known crystals, but 86% of its doping outputs relabel the input without changing its geometry. ALM Edit performs substitution reliably (doping 0.891), ahead of CrystaLLM + relaxation, and ties it on strain. Giving the agent ALM Edit as a tool transfers this capability: doping rises from 0.000 to 0.413 (paired p = 3.8×10−63.8×10^{-6}). The benefit is task-specific: polymorph falls from 0.800 to 0.540 (p = 0.0024) where the agent defers to ALM, and text-to-structure is unchanged within noise. ALM thus supplies a capability the agent lacks, as a complementary component rather than a replacement.

The learned interface

Bridge ablations explain where this capability comes from (CSP, MP-20; Appendix B):

Table R5. Bridge architecture ablation.

BridgeMatch@64 ↑
Element-count conditioning vector0.470
Pooled embeddings (MLP over K=8 atomistic tokens)0.513
Producer–consumer bridge (ALM)0.886

A fixed tool interface shows the converse. When a language model conditions MatterGen through its chemical_system field (optionally with E_hull=0), it picks the correct elements for ≥99.9% of Describe prompts, yet composition and structure recovery collapse to 0.02–0.09, 8-15× below ALM Edit, because the field cannot express stoichiometry, structure, or a requested change (Table R16, response to Reviewer hMvF).

Combining ALM with an external planner

Because ALM separates generating the composition from guiding denoising, an external Qwen3.5-9B can choose the composition, which ALM Edit then conditions on together with the prompt.

Table R6. External composition planner (paper test pool, OOD-v2, N=5×200).

ModelApp. ↑Desc. comp ↑Desc. struct ↑OOD comp ↑OOD struct ↑
ALM Edit0.323±0.0200.730±0.0310.412±0.0410.641±0.0320.316±0.022
+ Qwen3.5-9B planner0.512±0.0670.864±0.0190.329±0.0160.780±0.0260.371±0.030

The planner improves four of five scores. Together, these show two ways a learned, latent interface between language and atomistics fits a materials-design workflow: as a tool an agent calls, and as a generator that consumes an external planner's proposal.

AuthorsOpenReview ↗

Continuation of response to Reviewer UtRc's comments (#3)

Choosing between ALM Edit and ALM Gen

[Q1] ALM Edit and ALM Gen use different diffusion backbones and conditioning mechanisms... how should the two variants be used? Is there a way to select between them or adjust the conditioning strength based on the task?

ALM Edit is useful for a known composition or input crystal (crystal structure prediction, polymorph search, doping, and directional edits): its CSP decoder fixes the composition and searches over structures. ALM Gen is for de novo generation (DNG), where stability and novelty matter most, like for large-scale screening campaigns. Our reporting of two checkpoints reflects the fundamentally different use cases for CSP vs. DNG, and is the conventional method for reporting generative materials architectures (e.g., MatterGen was released with separate CSP and DNG architectures and checkpoints). Despite this convention, T2C-FK directly addresses the reviewer’s concern, as when an exact composition is also required, T2C-FK enforces it at inference time, and supplies almost all of Gen's compositional control (MP-20 MR@1 0.3% to 22.3%; Table R9 in our response to Reviewer xtzu). A practical recipe is to generate candidates with Gen + T2C-FK and refine a chosen candidate with Edit, as with the planner in Table R6 or the agent in Table R4 selecting the variant from the request. Conditioning strength can be adjusted at inference without retraining: both variants expose the classifier-free guidance scale g, and Gen additionally exposes the T2C-FK reward weight and particle count.

Learning rate for ALM Gen

[Q2] Table 7 reports a learning rate of (2e-4) for ALM Gen's LoRA adaptation, whereas Appendix B.2.4 and Figure 12 identify (1e-5)... Could you clarify which configuration was correct?

Thank you for catching this; 1e-5 is correct. ALM Gen's fresh rank-8 LoRA was trained at 1e-5, as described in Appendix B.2.4 and Figure 12. The value 2e-4 is the learning rate of ALM Core's rank-128 understanding LoRA, and is now corrected.

Scaling with the number of atoms

[Q3] What is the practical limit on the number of atoms that ALM can process, and how does its computational cost scale with crystal size? Does the model require any special handling for structures approaching or exceeding the LLM's context limit?

We benchmarked ALM Core on NaCl supercells from 8 to 21,952 atoms.

Table R7. ALM Core inference cost vs number of atoms (one H200 GPU, bf16, NaCl supercells). The * indicates that the first call includes one-time warm-up.

AtomsSequence length (tokens)Atom encoding (s)LLM prefill (s) ↓64-token generation (s) ↓Peak memory (GB) ↓
8280.90*0.032.1516.6
64840.040.031.9616.6
2162360.040.031.9216.7
5125320.010.031.9116.9
1,0001,0200.010.041.9217.2
1,7281,7480.010.071.9317.7
4,0964,1160.030.162.0320.3
8,0008,0200.060.332.1824.0
13,82413,8440.140.632.4638.5
21,95221,9720.331.122.9771.5

Generating 64 tokens takes 1.9-3.0 s at every size, and encoding the atomistic input takes at most 0.33 s, so decoding dominates cost up to roughly 10,000 atoms. Prefill grows roughly linearly with atom count over this range because of FlashAttention. The practical limit is the 32k-token context window minus the prompt, about 30,000 atoms. Inputs below this run without special handling, although ALM Core was trained on structures of up to 1,792 atoms. Structures beyond the context limit could be handled by pooling per-atom tokens into fixed-size groups, at the cost of per-atom resolution.

We hope these experiments and further analyses address the reviewer's concerns and inspire a reconsideration of our score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Final Recap for Area Chairs

Dear Area Chairs,

Many thanks to all reviewers and to the ACs for chairing the upcoming review. We wanted to provide a brief final recap. After the rebuttal, Reviewer hMvF raised their rating to 3, writing that they "appreciate the effort [we've] put into addressing my concerns." We were unfortunately not able to have a discussion with Reviewer xtzu, and our conversation with Reviewer UtRc is ongoing, so we summarize here how we addressed their concerns. Their main concerns were the separation of training data from architectural contributions, comparisons against modular tool-based pipelines, the scope of our crystal structure prediction (CSP) claim, and the fairness of the frontier LLM baselines. We ran new experiments for each of these and summarize them below.


Paper Recap

Scientists understand materials through both atomic structure and language, yet prior models handle only one of them natively, as language models that write crystals as text lose their geometry, agents that chain separate models through text never share what each model learns, and multimodal models that read structures cannot generate them. We propose Atomistic Language Models (ALMs), a new multimodal paradigm for materials in which a single model natively reads, writes, and edits crystals by connecting an atomistic encoder, a language model, and a diffusion decoder through learned continuous interfaces. Trained from pretrained components with standard instruction tuning, ALM Edit outperforms frontier language models like GPT-5.6 Sol on most tasks for editing materials as instructed. Because these capabilities had no existing evaluation, we built ALM Bench to measure language-instructed crystal editing and generation.

Reviewers recognized our motivation, construction, and clarity of presentation. UtRc called the framework "a meaningful step toward more flexible, language-driven materials design," xtzu noted that "no existing model can perform consistently across the range of tasks as demonstrated by ALM," and hMvF called the bridge "of significant practical value."


Revision Highlights

Reviewer feedback motivated substantial additional experiments and analyses:

  • Composition-split retraining: We clarified how our data was already held-out by structure, re-partitioned MP-20, MPTS-52, and ALM Bench by composition, retrained ALM Core and Edit, and evaluated CSP and ALM Bench on fully held-out compositions (Table R1, Appendix C.3).
  • Data and language model controls: We elucidated which components drive the performance of ALMs. We trained the CSP decoder on MP-20 alone as a data-matched control, ablated a frozen against a fine-tuned language model, and clarified each CSP model configuration (Tables R2, R2b, R3). T2C-FK can unify these capabilities (Table R9).
  • Agent and tool baselines: We added several more, strong baselines to ALM Bench evaluations, including MatSciAgent, AtomisticSkills with and without ALM Edit as a tool, CrystaLLM and element swap with relaxation (Table R4), and Qwen3-8B and GPT-5.2 conditioning MatterGen (Table R16).
  • LLM baselines: We added few-shot GPT-5.2 and a Qwen3-8B fine-tuned on ALM's task pairs (Tables R10, R11), and reported GPT-5.6 Sol at both reasoning efforts with an updated Table 4 (Tables R12, R13).
  • CSP scope: We thoroughly discuss which CSP metrics we claim are SoTA, and show that even with corrected CSP metrics from literature (METRe and cRMSE, Table R8), ALMs still maintain SoTA with Best-of-N=20 settings.
  • OOD generalization: We analyzed every OOD failure by element set, stoichiometry, and structure (Table R14), and paired ALM Edit with an external composition planner to illustrate how generalization can be improved (Tables R6, R15).

Overall, the discussion led to a clearer and more rigorously supported paper. The CSP claims are scoped to what the evidence shows, the contribution of the learned interface is separated from data and pretrained components, and ALM is placed in context against modular and agentic alternatives.

Thank you again 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.