LLMs Depend on Specialist Tools for Out-of-Distribution Generalization in MS/MS Structural Elucidation

Gabriel Asher; Cailum M. K. Stienstra; Tomo Oga; Devesh Shah Antonio H. de O. Fonseca; Niall O'Connor; Michael Widrich

cs.LGNov 2, 2026

LLMs Depend on Specialist Tools for Out-of-Distribution Generalization in MS/MS Structural Elucidation

Authors: Gabriel Asher, Cailum M. K. Stienstra, Tomo Oga, Devesh Shah Antonio H. de O. Fonseca, Niall O'Connor, Michael Widrich

Organizations: Matterworks, Inc. Somerville, MA 02143, USA

Abstract

A frontier language model was recently reported to solve 1D NMR structural elucidation without tools. We ask whether this capability transfers to tandem mass spectrometry (MS/MS) in the de novo setting, where the model must propose a structure directly from the spectrum without a candidate list or reference library. To separate reasoning from recall, we build a 100-analyte benchmark from MassSpecGym, split by each analyte's textual footprint into in-distribution (well documented in literature) and out-of-distribution (essentially undocumented) arms. We hold the model, harness, and output format fixed and vary only the LLM's access to domain-specific tools. Without tools, a frontier LLM reaches 52% Top-1 accuracy in-distribution but only 2% out-of-distribution. Only access to de novo structure generators improves out-of-distribution accuracy, raising it to as much as 38% Top-1, and the resulting agent outperforms the standalone de novo models both in and out of distribution. These results suggest that LLMs are not effective at de novo MS/MS structural elucidation on their own, and that further gains depend on the specialist tools available to the agent rather than on the LLM itself.

Figures & tables

Figure 1: Method overview. Analytes are split by documentation and matched 1:1 on acquisition; one member of a matched pair is shown per benchmark arm with its structure and spectrum. All experiments share one agent, prompt, and scorer, differing only in tool access.

Top-1

\uparrow

Top-5

\uparrow

Tanimoto@1

\uparrow

MCES@1

\downarrow

Model

OOD

All

OOD

All

OOD

All

OOD

All

Non-agentic reference (formula given)

–

Public de novo (FRIGID)

24 (*)

30 (*)

0.51

0.38

0.44

9.4 (*)

11.6

10.5

–

Public de novo (MS-GPT)

18 *

30 ***

42 *

0.69

0.50 ***

0.60 **

6.2

9.8 ***

8.0

–

In-house de novo ( LSM-Gen )

32 ***

46 **

36 ***

50 **

0.73

0.58 ***

0.66 ***

4.9

8.0 ***

6.5 **

Agentic ladder (formula given)

Table 1: The ablation ladder on

N{=}100

benchmark analytes (ID and OOD,

n{=}50

each).

Figure 2: Agent accuracy against its generator’s candidate coverage. Top-1 accuracy on the (a) ID and (b) OOD splits (

n{=}50

each). Dashed: coverage, the share of spectra whose true structure is anywhere in the generator’s pool. Brackets: exact McNemar, agent vs. generator, unadjusted.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Spectra
	Removed	Remaining
MassSpecGym v1.5 test fold		17,556
absent from the candidate pool	107	17,449
fewer than four fragment peaks	1,900	15,549
Analytes
Standardized test-fold analytes		2,992

Appendix

Table 2: Benchmark curation.

Figure 3: Agent harness. A spectrum is rendered into a query and passed to the model, which may dispatch tools through a per-spectrum dispatcher up to twenty times before it must answer. Answers, traces, and tool calls are recorded per spectrum for later scoring.

Non-agentic reference
experiment	Token USD	GPU s	GPU USD	Total USD	Top-1	USD / correct
FRIGID	0.00	37	0.043	0.04	15%	0.28
MS-GPT	0.00	32	0.036	0.04	33%	0.11
LSM-Gen	0.00	41	0.046	0.05	46%	0.10
Agentic ladder
L0	1.27	0	0.000	1.27	27%	4.70

Appendix

Table 3: Per-spectrum cost by experiment, as an upper bound: every input token is billed at the full rate. GPU seconds are queue-free service time, not wall clock, priced at the instance’s on-demand rate. The non-agentic baselines call one generator once and spend no tokens. Cost per correct is the total divided by Top-1 accuracy over all 100 analytes.

Figure 4: The in-distribution/out-of-distribution split is also a split on encoder training density. For each of the 100 benchmark molecules, the maximum 2048-bit ECFP4 Tanimoto to any structure in the MIST vFRIGID training dataset.

Figure 5: Top-1 hit rate by whether PubChem prints peak values for the analyte’s structure, on any CID of its InChIKey connectivity block. Left bar of each pair: no spectrum on PubChem (

n{=}69

); right: spectrum on PubChem (

n{=}31

). Brackets are two-sided Fisher exact tests within each method, unadjusted.

Figure 6: Each bar is the 50 spectra of one arm at one experimental level and sums to 100%: correct at Top-1, then the ways a wrong answer went wrong. The truth was never proposed in the two rightmost bands, which differ only in whether the run committed on its own or spent its token budget and answered when asked to. Availability is tested before the budget, so a run that had the truth and ranked it below first counts as a ranking failure even where it also ran out of tokens. Reachability is judged against each agent’s own generator pool, so it is absent from L0–L2, which have none.

Chemical Reasoning in LLMs

See all 20 papers

MS-GPT: Rethinking MS/MS De Novo Structure Elucidation as Spectrum-Induced Posterior Querying of a Molecule-Language Model
Jul 26, 2026Xin Zhao, Yumin Liu, Zhuo Li +5Language Modeling Molecular Generation
MoE Key Lab of Artificial Intelligence, AI Institute, School of Computer Science Shanghai Jiao Tong University Shanghai, China · Instrumental Analysis Center Shanghai Jiao Tong University Shanghai, China · ByteDance Inc. Beijing, China +1
Explainable Molecular Structure Inference from GC--MS with Diffusion Models and LLM Reranking
Oct 2, 2026Changlin Liu, Tianyu Yi, Chengchun Liu +2Molecular Graph Generation Mass Spectrometry
MARLIN: De Novo Molecular Structure Elucidation from Tandem Mass Spectra without a Ground-Truth Formula
Jul 6, 2026Xujun Che, Xiuxia Du, Depeng XuMolecular Generation Mass Spectrometry
Department of Software and Information Systems; University of North Carolina at Charlotte, Charlotte, NC 28223 · Center for Environmental Monitoring and Informatics Technologies for Public Health; University of North Carolina at Charlotte, Charlotte, NC 28223 · Department of Bioinformatics and Genomics; University of North Carolina at Charlotte, Charlotte, NC 28223

Jul 26, 2026J/K move · Enter open · S save

MS-GPT: Rethinking MS/MS De Novo Structure Elucidation as Spectrum-Induced Posterior Querying of a Molecule-Language Model

Xin Zhao, Yumin Liu, Zhuo Li, Weichu Zheng+4

MoE Key Lab of Artificial Intelligence, AI Institute, School of Computer Science Shanghai Jiao Tong University Shanghai, China · Instrumental Analysis Center Shanghai Jiao Tong University Shanghai, China · ByteDance Inc. Beijing, China+1

Language Modeling Molecular Generation Mass Spectrometry Conditional Generative Modeling

Open paper AskPDF arXiv

Molecular structure elucidation from tandem mass spectra (MS/MS) is a central inverse problem in analytical chemistry. Most existing approaches to MS/MS identification remain tied to reference libraries or predefined candidate sets, whereas de novo methods aim to generate structures directly from spectra. A common de novo route predicts a molecular fingerprint from the spectrum and then decodes structures from it, enabling decoder pretraining on large molecule-only corpora. However, this paradigm creates a training-inference mismatch: the decoder is trained on oracle fingerprints computed from molecules, but at inference it is queried with a noisy spectrum-induced fingerprint posterior that is typically collapsed to a single thresholded fingerprint. We introduce MS-GPT, which recasts fingerprint-mediated de novo elucidation as spectrum-induced posterior querying of a conditional molecule-language model. MS-GPT conditions a molecule-language model on fingerprints and formulas, then converts the spectrum-induced posterior into a band of fingerprint queries near the oracle-fingerprint manifold through active-bit density calibration. Candidates sampled across this band are pooled and ranked by generation frequency and posterior similarity. A lightweight LoRA adapter further mitigates domain-specific posterior bias while preserving the pretrained molecular prior. On NPLIB1 and MassSpecGym, MS-GPT sets a new state of the art, reaching Top-1/Top-10 exact-match accuracy of 17.1%/27.6% and 10.2%/13.2%, respectively, and candidate-pool scaling improves recovery with a near-linear increase in inference cost. The source code and model checkpoints are available at https://github.com/VIKI623/MS-GPT.

Abstract

Chemical Reasoning in LLMs

Explore similar work

MS-GPT: Rethinking MS/MS De Novo Structure Elucidation as Spectrum-Induced Posterior Querying of a Molecule-Language Model

Explainable Molecular Structure Inference from GC--MS with Diffusion Models and LLM Reranking

MARLIN: De Novo Molecular Structure Elucidation from Tandem Mass Spectra without a Ground-Truth Formula