mlsanityData to 14 Oct 2026

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

arXiv 2609.33407 ↗First posted 27 Sep 2026cond-mat.mtrl-sci cs.LG physics.comp-phcited by 0

Not on the map. Why?

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 31 of its 51 references that are in the map. First posted on arXiv 27 Sep 2026.

Builds on

31 of 51 references.

References · 31

  1. 2017Proximal Policy Optimization Algorithmscited by 15,215
  2. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  3. 2020Score-Based Generative Modeling through Stochastic Differential Equationscited by 9,122
  4. 2017Curiosity-driven Exploration by Self-supervised Predictioncited by 1,194
  5. 2022Scaling Laws for Reward Model Overoptimizationcited by 1,095
  6. 2018Exploration by Random Network Distillationcited by 993
  7. 2023Training Diffusion Models with Reinforcement Learningcited by 982
  8. 2016Unifying Count-Based Exploration and Intrinsic Motivationcited by 639
  9. 2018Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generationcited by 364
  10. 2021Crystal Diffusion Variational Autoencoder for Periodic Material Generationcited by 270
  11. 2023Crystal Structure Prediction by Joint Equivariant Diffusioncited by 148
  12. 2024Fine-Tuning of Continuous-Time Diffusion Models as Entropy-Regularized Controlcited by 143
19 more
  1. 2018Episodic Curiosity through Reachabilitycited by 109
  2. 2024FlowMM: Generating Materials with Riemannian Flow Matchingcited by 109
  3. 2020An invertible crystallographic representation for general inverse design of inorganic crystals with targeted propertiescited by 53
  4. 2020Reinforcement Learning for Molecular Design Guided by Quantum Mechanicscited by 51
  5. 2019Generative adversarial networks (GAN) based efficient sampling of chemical space for inverse design of inorganic materialscited by 46
  6. 2025Open Materials Generation with Stochastic Interpolantscited by 28
  7. 2025LeMat-GenBench: A Unified Evaluation Framework for Crystal Generative Modelscited by 23
  8. 2025Guiding Generative Models to Uncover Diverse and Novel Crystals via Reinforcement Learningcited by 21
  9. 2022Deep Reinforcement Learning for Inverse Inorganic Materials Designcited by 19
  10. 2026Reinforce Adjoint Matching: Scaling RL Post-Training of Diffusion and Flow-Matching Modelscited by 19
  11. 2020Curiosity in exploring chemical space: Intrinsic rewards for deep molecular reinforcement learningcited by 16
  12. 2026Crystalite: A Lightweight Transformer for Efficient Crystal Modelingcited by 10
  13. 2025Kinetic Langevin Diffusion for Crystalline Materials Generationcited by 9
  14. 2026Open Materials Generation with Inference-Time Reinforcement Learningcited by 8
  15. 2025Continuous SUN (Stable, Unique, and Novel) Metric for Generative Modeling of Inorganic Crystalscited by 6
  16. 2019Atomistic structure learningcited by 5
  17. 2026CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Modelscited by 2
  18. 2025Reinforcement Learning for Chemical Ordering in Alloy Nanoparticlescited by 1
  19. 2026Composable Crystals: Controllable Materials Discovery via Concept Learningcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026CrystalReasoner: Reasoning and RL for Property-Conditioned Crystal Structure Generation
  2. 2025Introducing physics-informed generative models for targeting structural novelty in the exploration of chemical space
  3. 2026Generative structure search for efficient and diverse discovery of molecular and crystal structures
  4. 2026Deep Generative Crystal Structure Prediction: A Benchmark Study and a Controlled Test of Prototype Dependence
  5. 2026SVGym (SciVerseGym): An Environment for Reinforcement Learning and Bayesian Optimization in Crystal Discovery
  6. 2026CrystalGRPO: Target-Aligned and Coverage-Preserving Reinforcement Learning for Flow-Based Crystal Structure Prediction
  7. 2026Towards Automated Discovery: A Review of Generative Models, Multimodal Learning and Closed-Loop Workflows in Inverse Materials Design
  8. 2026APO: Unsupervised Atomic Policy Optimization for 3D Structure Prediction of Atomic Systems