acceptodds
Under review as a conference paper at ICLR 2027

Adaptive LLM-Diffusion Integration for Robust Crystal Generation

Abstract

Text-guided generative models have shown strong promise for crystal generation by incorporating natural-language metadata into the generative process to produce valid and stable materials conditioned on textual descriptions. However, they have some inherent limitations: their LLM backbones are sensitive to the ordering of conditions within the prompt, the LLM and generative model are coupled through a fixed, sample-agnostic diffusion timestep , and their effectiveness has largely been demonstrated using only a limited set of LLM architectures. In this work, we introduce , a robust and adaptive LLM-diffusion framework for crystal generation. First, we improve robustness to prompt ordering through order-invariant training augmentation and auxiliary condition-specific objectives that encourage consistent generation across different permutations of the conditioning information. Second, we replace the fixed diffusion timestep with a lightweight per-sample policy network that adaptively determines the amount of diffusion refinement required for each LLM-generated structure. Finally, we benchmark across a diverse collection of open-source LLMs spanning multiple model families and scales, and observe that , across all backbones, consistently outperforms existing baselines in generating valid, novel, and stable materials, providing a broad assessment of the generality and robustness of LLM-diffusion integration for crystal generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ML4Materials, LLMs, Diffusion Models, Crystal Materials, Material Generation, AI4Science

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_re45OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose adding prompt-order robust training and a model that decides how much diffusion is needed for the 2 stage LLM + Diffusion model conditional crystal generation pipeline. They show ablations suggesting the method outperforms relevant baselines, but exact implications if the two methods contributions remain unclear.

Critical strengths

Critical strengths include:

  • Broad evaluation across different model backbones
  • Consistent improvements over matched backbone baseline
Critical weaknesses

Critical weaknesses include:

  • Prompt-robust training is simple data augmentation, with unclear gains (no clean ablation)
  • Noise prediction network lacks simple baselines, e.g. geometry-based heuristics, which would make the contribution of author’s method much clearer
  • Empirical gains are modest and don’t outweigh the other weaknesses.
  • Conditional experiments only use LLaMA-2, which is very old.
  • The error bars are poorly explained
Questions for authors

Could we have ablations for the different components?

What are the error bars showing exactly? And I would like some discussion on why they appear to be smaller than the baselines. There may be very good reasons for that, but I would like to see it discussed.

Reviewer_uU2POpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AdaCrysGen, which combines a fine-tuned LLM for initial crystal structure generation with an independently trained diffusion model for refining atomic coordinates and lattice parameters. The method introduces prompt-order augmentation, auxiliary condition-specific losses, and a policy network that selects the diffusion injection timestep for each sample. The framework is evaluated on MP-20 for both unconditional and text-conditioned crystal generation.

Critical strengths

The paper evaluates the LLM–Diffusion paradigm for crystal generation across multiple LLM families and model sizes. These experiments provide useful empirical results for future research in this direction.

Critical weaknesses
  1. The novelty of prompt-order augmentation and the formulation of auxiliary losses are insufficiently established. The paper introduces random prompt-order augmentation to improve robustness in crystal generation. However, prior works, including CrysLLMGen and CrystalTextLLM, already incorporate randomized condition construction and partial randomization of condition ordering. The paper does not sufficiently clarify the substantive difference between its proposed augmentation and these existing approaches, or establish the specific limitation of prior methods that motivates the new design. Furthermore, Equation (1) introduces auxiliary losses only as a weighted sum, without specifying their supervision targets, mathematical formulations, or computation procedures. As a result, the novelty of the proposed prompt-order augmentation over prior methods remains unclear, and the auxiliary losses cannot be fully understood or reproduced from the description provided.
  2. The motivation for adaptive timestep selection is insufficiently established. The paper introduces a policy network to select a diffusion injection timestep for each LLM-generated crystal structure, based on the assumption that different structures require different levels of refinement. However, the paper does not establish why a sample-specific selection strategy is necessary, rather than using a properly tuned global timestep. The policy selects the diffusion timestep based on a set of handcrafted features extracted from the LLM-generated crystal structure. However, it is unclear whether these features adequately capture the structural errors that determine how much diffusion refinement is needed. Consequently, the motivation for introducing a separate adaptive selection mechanism, beyond replacing a fixed hyperparameter with a learned policy, is not sufficiently established.
Questions for authors
  1. What unresolved problem does random prompt-order augmentation address? Prior works, particularly CrystalTextLLM and CrysLLMGen, already employ randomized condition construction and partial randomization of condition ordering. Which specific conditions and prompting strategies in these methods remain sensitive to condition ordering? Please clarify the substantive difference between the proposed augmentation and the existing randomization mechanisms, and provide controlled comparisons that retain the original randomization settings of these prior works. Could canonicalizing the order of input conditions at inference time already mitigate the reported problem?
  2. How are the auxiliary losses in Equation (1) defined and implemented? Please specify the supervision targets, mathematical formulations, and loss weights for chemical formula, elemental composition, space group, and crystal system. For each condition, please explain how the relevant target tokens are identified and how the corresponding auxiliary loss is computed. Please also provide ablation studies to quantify the contribution of these auxiliary losses to condition satisfaction and generation quality.
  3. How is “robust crystal generation” defined and evaluated in this work? Although robustness is emphasized in the title, Adaptive LLM-Diffusion Integration for Robust Crystal Generation, the paper primarily evaluates robustness to prompt condition ordering in Figure 2. It remains unclear whether this evaluation adequately supports the broader claim of robust crystal generation. The authors should clarify the intended scope of this claim and provide corresponding experimental evidence.
  4. Is adaptive timestep selection necessary, and what does the learned policy actually capture? Please compare the generation performance across candidate diffusion timesteps using the same set of LLM-generated initial structures to characterize the effects of shorter and longer diffusion refinement. Then, compare a properly tuned global fixed timestep, the adaptive policy, and a control that randomly reassigns the predicted timesteps across samples while preserving their overall distribution. This comparison would help determine whether per-sample adaptation provides benefits beyond selecting a favorable timestep distribution. Finally, please report the selected-timestep distributions across different LLM backbones and analyze how the selected timesteps relate to initial structure quality, structural changes after diffusion refinement, and final generation quality.
  5. Can the reported stability improvements be validated through independent evaluation and additional assessment on LeMat-GenBench? Both the reward used to train the τ\tau policy and the primary stability evaluation rely on CHGNet. Please validate the reported stability improvements using an independent interatomic potential or DFT calculations to determine whether the gains generalize beyond the surrogate model used to optimize the policy. Furthermore, please consider evaluating the adaptive-τ\tau and fixed-τ\tau variants under the unified evaluation protocol of LeMat-GenBench, reporting validity, stability, SUN, diversity, and structural quality before and after relaxation. These additional evaluations would help establish whether the reported gains persist across independent physical assessments and a more comprehensive set of crystal-generation metrics.
Reviewer_ffgROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines the problem of generating crystal structures via a two stage process. A first pass involves using an LLM to autoregressively generate a starting structure, after which a diffusion/flow model is used to refine the generation. Such methodology has been examined by prior work such as CrysLLMGen and FlowLLM. In this paper, the framework is improved via two changes, 1) data augmentation of LLM text descriptions via shuffling along with auxiliary losses, and 2) a dynamically learned diffusion timestep. The paper itself also compares such methodology across multiple LLM backbones.

Critical strengths

One strength of the paper is the breadth of its base LLM coverage, rather than focusing on only a single base LLM model, this paper considers five base LLM settings covering a range of sizes. The experiments show that the proposed AdaCrysGen method outperforms the two main prior works. This paper is generally reasonably written, and describes the implementation details along with a motivation for why those design choices were made.

Critical weaknesses

The main weakness of the paper is the ad-hoc nature of its contribution. The contribution that the paper proposed on top of existing methods are two relatively simple deltas which are do not interact with each other conceptually – that of doing data augmentation for text descriptions in the sense of simple shuffling, and using a learned time step policy rather than a fixed time step. It is a critical issue that these deltas were not tried separately. In addition, these design choices are seemingly unrelated to the specific crystal generation setting and if they were to be proposed as fundamental generative modeling ideas, should be compared in more general settings.

Questions for authors

n/a

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revised Manuscript

Summary of Revised Manuscript

We thank the reviewers for their constructive feedback. We have revised the main paper and the appendix, with all changes highlighted in blue. A summary of the major updates is as follows.

Key Updates

1. Clarifications and corrections

  • Defined "robust crystal generation" (Section 1).
  • Gave the full definition of the auxiliary condition-specific losses in Eq. (1): target tokens, loss form and loss weights (Section 3.1 and Appendix D).
  • Corrected the description of the comparison with CrysLLMGen. The LLM in AdaCrysGen is modified, and we no longer describe it as held fixed or identical (Sections 4.1 and 4.2, Appendix).
  • Stated in the Fig. 2 caption that error bars are the sample standard deviation over 5 runs, each with a different random shuffling of the condition order (Section 4.3).

2. Additional experiments

  • Ablation of the auxiliary losses on LLaMA-2 7B (Section 4.3, Table 3).
  • Ablation of the order-invariant augmentation and the auxiliary loss, separately and together, on three backbones under shuffled prompts (Appendix E.5).
  • Ablation of the LLM modifications and the adaptive timestep in four setups on three backbones (Appendix E.3).
  • Ablation on timestep selection, comparing the shortest and longest timesteps, a validation-selected timestep, random assignment of the predicted timesteps and the adaptive policy, and the distributions of the selected timesteps on three backbones (Appendix E.4).
  • Text-conditioned generation on LLaMA-3 8B and Qwen3-4B in addition to LLaMA-2 7B (Appendix E.2).

We believe these revisions address the reviewers' concerns and improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.