How well can off-the-shelf LLMs elucidate molecular structures from mass spectra using chain-of-thought reasoning?

Mass spectrometry (MS) is a powerful analytical technique for identifying small molecules, yet determining complete molecular structures directly from tandem mass spectra (MS/MS) remains a long-standing challenge due to complex fragmentation patterns and the vast diversity of chemical space. Recent progress in large language models (LLMs) has shown promise for reasoning-intensive scientific tasks, but their capability for chemical interpretation is still unclear. We present a zero-shot chain-of-thought (CoT) benchmark for evaluating how well off-the-shelf general-purpose LLMs can infer molecular structure from text-serialized MS/MS peak lists and a known molecular formula, without fine-tuning, retrieval, external chemistry tools, or constrained search. Rather than proposing a new spectrum-trained model, our goal is diagnostic: to isolate what the evaluated off-the-shelf LLMs can and cannot do in a direct MS/MS-to-SMILES generation setting when only textualized spectral evidence and prompting are available. This distinction is important because our benchmark targets a deliberately constrained regime: direct structure generation from text-formatted spectral evidence alone, rather than multimodal reasoning with spectral images, candidate reranking over structures generated by specialized chemistry systems, or knowledge and search-augmented LLM workflows. We formalize expert chemists' reasoning steps (such as double bond equivalent (DBE) analysis, neutral loss identification, and fragment assembly) into structured prompts and assess multiple state-of-the-art LLMs (Claude-3.5-Sonnet, GPT-4o-mini, and Llama-3 series) using the MassSpecGym dataset. Our evaluation across metrics of SMILES validity, formula consistency, and structural similarity reveals that while LLMs can produce syntactically valid and partially plausible structures, they fail to achieve chemical accuracy or link reasoning to correct molecular predictions. These findings highlight both the interpretive potential and the current limitations of LLM-based reasoning for molecular elucidation, providing a foundation for future work that combines domain knowledge and reinforcement learning to achieve chemically grounded AI reasoning.Scientific contributionWe introduce a zero-shot chain-of-thought benchmark for direct MS/MS-to-SMILES prediction with off-the-shelf general-purpose LLMs, providing a controlled test of text-only, no-tools molecular structure elucidation. In contrast to prior approaches based on spectrum-trained models, multimodal reasoning, or tool-augmented candidate pipelines, our study isolates the capabilities and limitations of prompting alone. We show that fluent chain-of-thought reasoning does not translate into chemically grounded structure recovery, establishing a reproducible baseline and a clear target for future method development.