Tandem mass spectrometry identifies an unknown molecule by breaking it into charged fragments and recording their masses. The hard cases are structural isomers: molecules with the same atoms wired differently. Their fragment patterns overlap almost entirely, so any evidence that separates two such isomers sits in a handful of peaks. A few fragment masses can be produced by one wiring and not the other, and a few expected fragments can be conspicuously missing. Vision-language models show a known failure pattern called perception-reasoning decoupling. They read a chart's title instead of its axes, and they fail at deduction even when they perceive the inputs correctly (Plot Twist, MATH-Vision). We suspect spectrum-to-structure models have the same failure. They may pick among isomers using formula and database-popularity priors while ignoring the few peaks that actually discriminate. Some evidence already points this way. MassSpecGym in the Wild found that knowing the formula alone lifts a random ranker's hit rate from 0.43% to 12.98%, and that a popularity-only ranker competes with neural models.
This note proposes two things. The first is a counterfactual test of whether a model's isomer choice depends on the diagnostic peaks: we delete exactly those peaks and see whether the preference collapses. The second is a pairwise verifier that makes the discriminating evidence explicit and executable. For each pair of candidate isomers, it computes from structure alone which fragment masses only one of them can produce. It then weighs the presence and the absence of those masses in the spectrum. When the spectrum does not separate the pair, it abstains. The multimodal-reasoning literature supplies the lens: separate perception from deduction, probe elementary perception directly, and replace free-form latent reasoning with executable rationales. Mass spectrometry supplies the application, where a confident wrong isomer means a forensic false positive.
Core Claim
We make two claims. First, current elucidation and retrieval systems (GEMS, FOAM, DiffMS, SpectraLLM, and the MIST retriever evaluated in MassSpecGym) are only weakly grounded in diagnostic peaks. Their preference for the true isomer over its nearest decoy should survive deletion of the peaks that distinguish them nearly as well as it survives deletion of random, equally intense peaks.
Second, an explicit pairwise likelihood-ratio verifier improves the final isomer choice on top of GEMS or FOAM candidates, and it gives a calibrated "cannot distinguish" answer. The verifier scores diagnostic masses using fragments enumerated from structure rather than a learned whole-spectrum forward simulator, and it counts missing expected fragments as evidence. Its gains should concentrate on scaffold-disjoint splits, where forward simulators are weakest; MassSpecGym reports a simulation cosine of 0.52.
The test is new; nobody has probed spectral models the way Chart101 probes chart models. What distinguishes the verifier is that it is pairwise, uses absence evidence, and can abstain.
Research Program
Isomer-pair construction. For each test spectrum of true molecule A, we collect hard decoys B that share A's formula. They come from three sources: GEMS-style 2-switch neighbours (local bond rewirings that keep valences valid), the top-ranked wrong candidates of existing systems, and same-formula PubChem entries in the MassSpecGym retrieval candidate sets. We enumerate the fragments each candidate can form by breaking up to three bonds, with hydrogen shifts of up to two, in the style of classical combinatorial fragmenters. The diagnostic set D(A,B) is the set of fragment masses, within instrument tolerance, that are reachable from one candidate but not the other. Because this uses only graph enumeration, it sidesteps the forward-model bottleneck that GEMS and FOAM inherit.
The grounding test. This is modelled on Chart101's elementary probes and on GeoSQA's perception-isolation control (GeoSQA). Each model scores A against B under three conditions: the original spectrum, the spectrum with observed peaks in D(A,B) removed, and the spectrum with random non-diagnostic peaks of matched total intensity removed. The grounding gap is the drop in pairwise preference under diagnostic removal minus the drop under control removal. A model that reasons from evidence should fall to chance when the diagnostic peaks are gone. A prior-driven model should barely move.
We add a symmetric "swap" condition. We implant the diagnostic peaks of B into A's spectrum, and the preference should flip.
For language-model systems such as SpectraLLM, we also run an oracle-perception condition. The model receives the correct fragment-formula annotations as text, which lets us split its errors into perception versus deduction, mirroring the MATH-Vision audit (31.9% vision errors, 42.2% reasoning errors for GPT-4V).
The verifier. For a candidate pair, the verifier computes a log-likelihood ratio over the diagnostic masses:
$$\Lambda(A,B)=\sum_{m\in D(A,B)}\Big[y_m\log\frac{\pi_A(m)}{\pi_B(m)}+(1-y_m)\log\frac{1-\pi_A(m)}{1-\pi_B(m)}\Big]$$
Here $y_m$ is 1 if a peak is observed at mass $m$, and $\pi_X(m)$ is the probability that candidate $X$ yields an observable fragment at $m$. That probability comes from a small graph network trained on annotated library spectra to estimate bond-cleavage propensity, conditioned on collision energy and adduct. The second term is what "explained-peak" scorers lack. A fragment that B would almost surely produce but that is absent counts against B. The verifier abstains when $|\Lambda|$ falls below a threshold set on validation data.
Over the top-k candidates from GEMS or FOAM, pairwise ratios are aggregated with a Bradley-Terry model, which fits one strength per candidate from all pairwise comparisons. Each decision comes with its rationale: the specific masses, their fragment formulas, and the substructures that produce them. This follows the question, rationale, answer decomposition that Synthesize Step-by-Step showed lifts ChartQA human-written accuracy from 38.2% to 53.9%. The difference is that our rationale is a deterministic RDKit program rather than LLM-generated code.
Relation to prior work. GEMS builds a correspondence graph between the experimental spectrum and a simulated one. It notes that cosine similarity fails among decoys but offers no pairwise discriminator. We keep its search and replace its final choice. FOAM mutates candidates blindly and scores them with ICEBERG. Our verifier could later steer its mutations, but we do not claim that here. AbductiveMLLM motivates the idea of inferring the hypothesis that best explains partial evidence. We make that inference explicit and symbolic rather than training it into a network. SpectraLLM names chain-of-thought as future work. We test whether reasoning of any kind helps once perception is fixed.
Baselines
The naive reference points come from MassSpecGym in the Wild: a formula-constrained random ranker (12.98% hit rate) and a PubChem-popularity ranker. Every system's accuracy is reported relative to both, so that prior-driven gains are visible. The same audit reported no popularity-ranker number that we could find on Lacuna, so we will measure it ourselves.
The strongest existing method is GEMS, which reports 30.48% Top-1 on the NIST unique subset, against DiffMS at 6.53% and MolSpecFlow at 8.22% (GEMS). GEMS reaches the target in 72.19% of queries yet ranks it first in only 30.48%. That gap is exactly what a better final discriminator can close, which makes GEMS the natural host. We run GEMS with its released settings and 16 trajectories. The verifier reranks only GEMS's final candidate pool, with no extra search budget, so any gain is attributable to choice rather than exploration. Beating GEMS means a Top-1 gain on the NIST unique subset that exceeds seed-to-seed variance across three runs, plus a larger gain on CASMI 2016, where GEMS reaches the target 32.02% of the time zero-shot.
On MassSpecGym, the reference numbers are FOAM's refinement of DiffMS (Top-10 exact match from 4.25% to 10.28%) and MIST retrieval Hit@1 of 14.6%. We rerank FOAM's and MIST's outputs on the official scaffold-disjoint split, following the Wild audit's corrected protocol, including the fixed MIST masking bug that had inflated Tanimoto from 0.37 to 0.52.
Three ablations isolate the key ingredients. The first removes the absence term, keeping only observed-peak evidence. The second replaces $\pi$ with ICEBERG-simulated intensities, which tests whether structure-enumerated fragments matter or whether a forward model suffices. The third replaces the pairwise ratio with the difference of independent per-candidate explained-intensity scores.
Why not just…?
Why not just score each candidate independently with a classical fragmenter like MAGMa+ or CFM-ID? If per-candidate scores were purely additive over observed peaks, the peaks two candidates share would cancel in their difference, and our pairwise form would reduce to MAGMa-style scoring. The real difference lies in the absence term and the learned propensities. The CASMI comparison (in silico tools; Comprehensive comparison of in silico MS/MS fragmentation tools of the CASMI contest: database boosting is needed to achieve 93% accuracy) found that a consensus of such tools reaches only 22% Top-1 without database boosting. Still, this is the most serious threat. Our third ablation is exactly this comparison. If it ties the full verifier, we will report the grounding test and the abstention calibration as the contribution and drop the claim of a better scorer.
Why not just train a learned cross-encoder reranker with hard isomer negatives? A network that sees spectrum-candidate pairs and is trained on same-formula decoys could plausibly match our accuracy on in-distribution data. We will train one on the same training split and candidate pools. The grounding test decides the comparison. If the cross-encoder matches our accuracy and also shows a large grounding gap, it is a genuine competitor, and our verifier's features become a way to audit it. If its preferences survive diagnostic-peak deletion, its accuracy is borrowed from priors, and it should degrade on the CASMI and scaffold-disjoint splits.
Why not just prompt a strong vision-language model or SpectraLLM to reason step by step? MATH-Vision reports that chain-of-thought hurt several multimodal models, and Chart101 reports collective elementary-perception accuracy below 25%. Both findings suggest that free-form reasoning over spectra will inherit perception errors. The oracle-perception condition tests this directly. If SpectraLLM chooses correctly once it is handed correct fragment annotations, the bottleneck is perception and our enumerator fixes it. If it still fails, deduction is the bottleneck, and explicit executable rationales are needed.
Proposed Experiments
The first decisive check is the grounding test itself, run across GEMS, FOAM, DiffMS, MIST retrieval, a SpectraLLM-style model, the cross-encoder, and our verifier on MassSpecGym's scaffold-disjoint split and on NIST. We expect existing systems to show a small grounding gap and the verifier to show a large one by construction. The informative result is the size of the gap for the learned systems. If they turn out well grounded, the motivating hypothesis is wrong and we say so.
The second check reranks GEMS on NIST and FOAM on MassSpecGym, reporting Top-1, Top-10, pairwise isomer accuracy, and MCES distance. Results are stratified by molecular size, since GEMS degrades beyond 40 heavy atoms and enumeration cost grows with size.
The third check is selective prediction. We plot accuracy against coverage and report the false-positive rate at 80% coverage. This matters for forensic settings such as synthetic cannabinoids, where near-identical spectra make abstention more valuable than a guess.
The fourth check tests where the rationales fail. On a held-out sample, we examine whether the cited diagnostic masses correspond to fragments annotated in the reference data. We report the fraction of verifier errors caused by rearrangement fragments, which bond-cleavage enumeration cannot reach. That fraction bounds how far a purely structural verifier can go.
References
- Plot Twist: Multimodal Models Don't Comprehend Simple Chart Details
- Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset
- MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery
- GEMS: Molecular Structure Identification via Geodesic Navigation of the Isomer Manifold
- Generative structural elucidation from mass spectra as an iterative optimization problem
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra
- SpectraLLM: Uncovering the Ability of LLMs for Molecule Structure Elucidation from Multi-Spectra
- MassSpecGym: A benchmark for the discovery and identification of molecules
- GeoSQA: A Benchmark for Scenario-based Question Answering in the Geography Domain at High School Level
- Synthesize Step-by-Step: Tools, Templates and LLMs as Data Generators for Reasoning-Based Chart VQA
- AbductiveMLLM: Boosting Visual Abductive Reasoning Within MLLMs
- Comprehensive comparison of in silico MS/MS fragmentation tools of the CASMI contest: database boosting is needed to achieve 93% accuracy