Do LLMs Truly Generalize in the Molecular Domain? A Perturbation-Based Analysis

2026-07-02

AI Summary

This study investigates the generalization capabilities of molecular large language models (LLMs) under structural perturbations, revealing their excessive reliance on local sequence neighborhoods and a narrow trust region. To address this limitation, the authors introduce the first controllable structure perturbation generation and evaluation framework based on graph edit distance, along with an In-Context Tuning strategy to enhance model robustness. Experimental results demonstrate that even a single structural edit can significantly degrade model performance, whereas the proposed fine-tuning approach effectively expands the trust region and improves adaptability to structural variations. This work establishes a novel evaluation paradigm and provides actionable pathways for advancing structure-aware generalization in molecular LLMs.

Abstract

Large Language Models (LLMs) have recently shown promise in molecular discovery, yet a gap remains between their probabilistic nature over discrete sequential tokens and the rigid topological constraints of chemical space. This raises the question of whether molecular LLMs can generalize beyond the local neighborhoods induced by their sequence-based representations. To systematically investigate this question, we introduce a Molecular Perturbation framework that generates syntax-valid structural variants of training molecules under controlled Graph Edit Distance (GED) to probe the manifold regularity of molecular LLMs. Our analysis shows that even a single edit can cause substantial performance drops on common molecular tasks, revealing a narrow local trust region and fragile sensitivity to structural changes. Since similar molecules tend to exhibit similar properties, In-Context Tuning (ICT), which anchors predictions on structurally similar molecules, offers a natural way to mitigate such fragility. Our experiments also examine whether ICT confers robustness under controlled structural perturbations, and the results suggest that it can partially expand the local trust region and offer a promising direction for stabilizing molecular LLMs against structural variation.

Problem

• generalization
• molecular LLMs
• structural perturbation
• trust region
• molecular representation

Innovation

• Molecular Perturbation
• Graph Edit Distance
• In-Context Tuning
• Local Trust Region
• Molecular LLMs