[Paper Review] Enhancing Spatial Reasoning in Large Language Models for Metal-Organic Frameworks Structure Prediction

Mianzhi Pan, J. Li+6 more|arXiv (Cornell University)|Jan 14, 2026

Metal-Organic Frameworks: Synthesis and Applications0 citations

TL;DR

MOF-LLM introduces a three-stage training pipeline (spatial-aware CPT, structural SFT, and matching-driven RL) to enable autoregressive MOF block assembly with improved spatial reasoning and efficiency, outperforming denoising-based and other LLM-based methods.

ABSTRACT

Metal-organic frameworks (MOFs) are porous crystalline materials with broad applications such as carbon capture and drug delivery, yet accurately predicting their 3D structures remains a significant challenge. While Large Language Models (LLMs) have shown promise in generating crystals, their application to MOFs is hindered by MOFs' high atomic complexity. Inspired by the success of block-wise paradigms in deep generative models, we pioneer the use of LLMs in this domain by introducing MOF-LLM, the first LLM framework specifically adapted for block-level MOF structure prediction. To effectively harness LLMs for this modular assembly task, our training paradigm integrates spatial-aware continual pre-training (CPT), structural supervised fine-tuning (SFT), and matching-driven reinforcement learning (RL). By incorporating explicit spatial priors and optimizing structural stability via Soft Adaptive Policy Optimization (SAPO), our approach substantially enhances the spatial reasoning capability of a Qwen-3 8B model for accurate MOF structure prediction. Comprehensive experiments demonstrate that MOF-LLM outperforms state-of-the-art denoising-based and LLM-based methods while exhibiting superior sampling efficiency.

Motivation & Objective

• Motivate accurate 3D MOF structure prediction given complex, block-level modular MOFs.
• Develop an LLM-based pipeline that handles block roto-translations and lattice parameters.
• Incorporate spatial priors and a stability-focused learning objective to ensure physically plausible MOF assemblies.
• Demonstrate superior predictive accuracy and sampling efficiency over baselines.

Proposed method

• Convert MOF structures into block-level textual representations suitable for LLMs (lattice parameters, fractional coordinates, Euler angles, and canonical SMILES blocks).
• Apply spatial-aware continual pre-training (CPT) to inject geometry and topology priors (topology code, molecular weight, PCA spans, rotated axes).
• Perform structural supervised fine-tuning (SFT) to autoregressively generate lattice parameters and block roto-translations from SMILES blocks.
• Use matching-driven reinforcement learning (SAPO) with a structure-matching reward to refine generation toward stable MOF configurations.
• Reward r(i) balances parsing success and geometric fidelity via RMSE-based matching, guiding policy updates with SAPO objective.
• Render 3D rotations via Euler angles to avoid degenerate representations and facilitate learning.

Figure 1: Illustration of MOF structures. An MOF structure and its unit cell are shown on top. The unit cell can be decomposed into building blocks, including metal nodes and organic linkers (bottom). Atom color: Zn (purple), O (red), C (gray), H (white).

Experimental results

Research questions

• RQ1 Can an LLM be trained to predict MOF lattice parameters and block-level roto-translations in a block-wise assembly framework?
• RQ2 Does spatial-aware CPT improve LLM spatial reasoning for MOF block assembly compared to standard fine-tuning?
• RQ3 Can matching-based RL further improve the structural stability and accuracy of MOF predictions beyond SFT?
• RQ4 How does MOF-LLM compare to denoising-based and other LLM-based MOF structure prediction methods in accuracy and inference speed?

Key findings

• MOF-LLM achieves higher match rates (MR) and lower RMSE than state-of-the-art denoising-based and LLM-based baselines across settings.
• On unseen MOFs, MOF-LLM attains a match rate of 35.78% with stol=0.5 and 93.25% with stol=1.0 for 1 sample, and 53.64% with stol=0.5 and 98.48% with stol=1.0 for 5 samples.
• MOF-LLM inference time is 0.04 seconds per structure, far faster than competing methods.
• Geometric property RMSEs show MOF-LLM competitive performance with MOF-BFN and better than MOFFlow on key descriptors (VSA, GSA, AV, UCV, VF, PLD, LCD, DST).
• Ablation studies confirm spatial-aware CPT and RP-based RL (SAPO) each provide substantial gains over baselines and reduced instability in generated MOFs.

Figure 2: Overview of our MOF-LLM framework. It enhances the spatial reasoning ability of LLMs for MOF structure prediction via a three-stage training: spatial-aware CPT to inject spatial and geometric priors, structural SFT enables block assembly, and matching-driven RL refines structure stability