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

2026-07-02 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

Large Language Models (LLMs) show promise in molecular discovery. However, their generalization in chemical space is questioned due to the gap between probabilistic token sequences and rigid topological constraints. A new Molecular Perturbation framework investigates this. It generates syntax-valid structural variants of training molecules under controlled Graph Edit Distance (GED). Analysis reveals that even a single structural edit causes substantial performance drops on molecular tasks. This indicates a narrow local trust region and fragile sensitivity. In-Context Tuning (ICT), which anchors predictions on structurally similar molecules, can partially expand this region. This offers a promising direction for stabilizing molecular LLMs against structural variation.

Key takeaway

For Machine Learning Engineers developing molecular LLMs, understanding their generalization limits is crucial. Your models are highly sensitive to structural changes; even minor edits cause significant performance degradation. You should consider implementing In-Context Tuning (ICT) to expand the local trust region and stabilize predictions against structural variations. This approach can improve the reliability of molecular discovery applications in real-world chemical space.

Key insights

Molecular LLMs exhibit fragile generalization to structural changes, but In-Context Tuning can partially mitigate this sensitivity.

Principles

• Molecular LLMs have narrow local trust regions.
• Similar molecules tend to exhibit similar properties.
• Single structural edits can degrade performance.

Method

The Molecular Perturbation framework generates syntax-valid structural variants of training molecules under controlled Graph Edit Distance (GED) to probe manifold regularity.

In practice

• Employ In-Context Tuning for molecular LLM robustness.
• Anchor predictions on structurally similar molecules.

Topics

• Large Language Models
• Molecular Discovery
• Chemical Space
• Graph Edit Distance
• In-Context Tuning
• Model Generalization

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Rethinking Molecular Text Representations for LLMs: An Empirical Study — LLM molecular competence is strongly representation-dependent, with no single representation optimal across all tasks.
• Generalization in LLM Problem Solving: The Case of the Shortest Path — LLMs generalize spatially but fail at length scaling in shortest-path problems due to recursive instability.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific datasets and LoRA fine-tuning significantly enhance LLM performance in specialized tasks like classical poetry analysis.
• Shuffle the Context: RoPE-Perturbed Self-Distillation for Long-Context Adaptation — RoPE-Perturbed Self-Distillation improves LLM long-context understanding by reducing reliance on absolute positional encoding.
• Customized Amazon Nova models improve molecular-property prediction in drug discovery — Fine-tuned LLMs can unify complex molecular property prediction, outperforming or matching specialized GNNs.
• CDR-Bench: Evaluating Faithful Execution of Compositional, Order-Sensitive Data Refinement Recipes — Current LLMs struggle with the procedural faithfulness required for compositional, order-sensitive data refinement recipes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.