In research published in Discover Chemistry, Takuya Ehiro and Reiko Yamashita systematically evaluated GPT-family models as zero-shot generators of odor descriptors, meaning the models received no task-specific training whatsoever. Working with a curated dataset of 3,493 odorant molecules drawn from the Leffingwell PMP 2001 database and distributed via the Pyrfume repository, the team asked the models to describe odors, rate pleasantness on a scale from minus four to plus four, and even estimate their own reliability. The results reveal that these language systems, trained on vast human-generated text spanning chemistry, perfumery, and food science, have absorbed a remarkable amount of collective olfactory knowledge.

The most striking finding concerns concentration. In human perception, the same compound can smell entirely different depending on how much of it reaches the nose. Indole, for example, is floral at trace levels and fecal at high concentrations. When the researchers prompted the models with low- and high-concentration conditions for four such odorants, including indole, diacetyl, skatole, and cis-3-hexen-1-ol, the newer GPT-5 models reproduced these hedonic shifts with impressive fidelity. GPT-5-mini and GPT-5 assigned positive pleasantness scores to indole and skatole at low concentrations and collapsed to the bottom of the scale at high concentrations, exactly mirroring the documented perceptual transitions. The older GPT-4o-mini failed these tests, suggesting that advances in model capability translate directly into better encoding of concentration-percept relationships.

The team also probed whether language models could capture cultural dimensions of smell, something no structure-based approach can attempt. They assigned the models distinct cultural personas through role-play prompting, simulating sensory scientists from East Asia, North America, and Western Europe, then asked them to evaluate methional, a key aroma compound in soy sauce and kimchi, and methyl salicylate, the wintergreen compound beloved in American candy but associated with medicine in Europe. The models produced trends in the predicted directions, with East Asian personas rating methional more favorably, but the differences were not statistically significant. The authors frame this as a proof of concept rather than evidence of culturally grounded odor modeling, noting that it echoes cross-cultural research showing molecular identity matters far more than cultural background in odor pleasantness.

To benchmark descriptor quality, the researchers compared LLM-generated odor descriptions against expert-assigned ground-truth labels using two vectorization schemes: embedding-based similarity in a continuous semantic space and count-based matching against 180 canonical descriptor columns. A message-passing neural network pretrained on data including the evaluation set achieved the highest scores, with a mean cosine similarity of 0.814 under embedding-based vectorization, but this comparison was inherently asymmetric, since the neural network had seen the test data during training while the language models operated entirely zero-shot. All LLM conditions substantially exceeded a frequency-informed random baseline of 0.677, demonstrating genuine chemical-olfactory knowledge rather than simple label-frequency matching. Notably, under top-10 sampling, the neural network’s performance dropped to a level comparable to or below GPT-5-mini and GPT-5.

Perhaps the most revealing experiments involved scrambling the inputs. When the researchers paired each compound’s correct chemical name with a SMILES string from a different molecule, 76.3 percent of the generated descriptors were semantically closer to the name-source compound than to the SMILES-source compound, establishing that the chemical name is the primary carrier of odor-relevant information. Yet the models were not ignoring structure entirely: in the name-masked condition, they still produced chemically plausible descriptors from SMILES strings alone, and in 5.6 percent of scrambled cases they explicitly flagged the mismatch between name and structure, evidence of cross-referencing between the two representations. This aligns with earlier findings that IUPAC nomenclature, which systematically encodes functional groups and substitution patterns, outperforms other molecular string formats for language model prediction tasks.

Can a chatbot’s self-reported confidence be trusted? Partially, the study suggests. The models’ expressed reliability scores correlated positively with prediction accuracy across name-containing prompt conditions, with correlations reaching 0.45 against the minimum cosine similarity at low concentration, and negatively with prediction variability. The correlation vanished for SMILES-only prompts, indicating that confidence is only informative when the model has sufficient semantic context. Meanwhile, principal component analysis of the embedding space revealed that hedonic valence dominates the first principal component, accounting for over 42 percent of variance, meaning pleasantness is the single largest organizing axis of the models’ internal odor representations.

To verify that the models’ pleasantness judgments rest on genuine chemistry rather than linguistic coincidence, the team trained an XGBoost surrogate model to predict LLM-assigned pleasantness scores from molecular fingerprints alone. It achieved an R-squared of 0.810 at high concentration, and even under a demanding scaffold-based split that excluded all unseen molecular cores from training, performance remained at 0.754. SHAP analysis surfaced chemically interpretable patterns: acyclic ether and aromatic fragments contributed positively to predicted pleasantness, while nitrogen-containing and carbonyl substructures contributed negatively. An independent effect-size analysis confirmed the picture, with thiol groups showing the strongest negative association and ethers and esters the strongest positive ones, directions fully consistent with established olfactory literature.

The practical payoff came in predicting instrumental odor sensor measurements. Using a Shimadzu FF-2020 electronic odor identification system, the researchers measured similarity scores across nine gas categories for 44 compounds, then tested whether features derived from LLM outputs could improve prediction of these readings. For aliphatic hydrocarbons, LLM-derived features achieved an R-squared of 0.567 with XGBoost, compared with just 0.222 for conventional molecular fingerprints; for organic acids, the LLM features reached 0.429 against 0.189 for fingerprints. The study is candid about limitations, including the small sensor sample, cross-contamination in the nine-dimensional axis projection, and the models’ heavy dependence on chemical names, but the direction is clear.

What emerges is a portrait of language models as complementary players in computational olfaction rather than replacements for structure-based methods. They cannot yet beat a supervised graph neural network on its own benchmark, but they offer something no fingerprint ever could: sensitivity to concentration, hedonic framing, and potentially cultural context, all accessible through nothing more than prompt text. As generative models continue to improve, the boundary between reading about smell and predicting it grows thinner, and the perfume lab, the flavor house, and the environmental sensor industry may all find a new kind of colleague waiting inside their chat windows.

Subject of Research: Zero-shot evaluation of generative large language models for predicting odor quality and odor sensor measurements from molecular information

Article Title: Evaluating generative large language models as zero-shot semantic odor descriptor generators with applications to odor sensor prediction

Article References: Ehiro, T., & Yamashita, R. (2026). Evaluating generative large language models as zero-shot semantic odor descriptor generators with applications to odor sensor prediction. Discover Chemistry, 3(1), Article 570. https://doi.org/10.1007/s44371-026-01015-7

Image Credits: AI Generated

DOI: 10.1007/s44371-026-01015-7

Keywords: large language models, odor prediction, olfactory perception, zero-shot learning, GPT-5, chemical nomenclature, molecular fingerprints, pleasantness, odor sensors, machine learning, quantitative structure-odor relationship, sensory science

The study, published as an open-access paper in Food Chemistry: X, was led by Huijing Lan and colleagues including senior author Yan Xu, with analytical work carried out at Jiangnan University. The team selected Nongxiangxing Baijiu as a proof-of-concept system for good reason: its aroma arises from a large and diverse set of volatile compounds generated during solid-state fermentation, distillation, and aging, and its evaluation tradition explicitly prizes complexity-related attributes such as richness, layering, and persistence. Unlike cross-style comparisons, where differences may simply reflect broad stylistic variation, comparing products within the same aroma style asks a sharper question: can chemistry distinguish products sharing the same characteristic compounds but differing in perceived depth?

The researchers analyzed 21 commercial Nongxiangxing Baijiu samples at 52% alcohol by volume, drawn from eight representative Chinese brands and spanning high-, medium-, and low-tier products as declared by their manufacturers. They quantified 140 volatile compounds chosen from prior quantitative and gas chromatography-olfactometry studies, using three complementary analytical methods tailored to different compound classes and abundances. High-abundance compounds such as major esters and alcohols were measured by direct-injection gas chromatography with flame ionization detection. Compounds requiring headspace enrichment were captured by solid-phase microextraction coupled to GC-MS, while trace-level analytes such as pyrazines, lactones, and sulfur compounds were concentrated by dispersive liquid-liquid microextraction and quantified by tandem mass spectrometry. Calibration curves, built in matrix-matched ethanol solutions with isotope-labeled internal standards, showed coefficients of determination from 0.9131 to 1.0000 and recoveries between 80.1% and 129.6%.

The heart of the work lies in how those raw concentration tables were transformed into perceptually meaningful dimensions. The framework rests on three pillars. The first, functional complexity, asks how broadly the volatile profile spans odor quality space. Each of the 140 compounds was passed through a machine-learning odor prediction model, the principal odor map framework of Lee and colleagues, which generated probability scores for 138 odor descriptors from molecular structures supplied as SMILES strings. As a consistency check, the ten highest-ranked predictions for each compound were compared with annotations in the Good Scents database: at least one database annotation appeared among the top predictions for 99.29% of compounds. The 138 descriptors were then collapsed into twelve semantic aroma categories, including fruity, floral, woody, baked, and herbal, and Shannon entropy was used to measure how evenly a compound’s predicted odor information spreads across those categories.

The second pillar addresses time. Odor persistence, the tendency of an odorant to linger, was predicted for every compound by a hybrid graph neural network trained on a curated fragrance dataset that uses the historical relative-persistence scale of Poucher. The model combined a two-layer graph attention network with a multilayer perceptron, fed by six physicochemical descriptors plus a 2048-bit molecular fingerprint, and achieved a mean absolute error of 5.43 with a coefficient of determination of 0.93 in five-fold cross-validation. High-persistence predictions clustered among medium-chain fatty acids and their esters, lactones, aromatic aldehydes, phenolics, and terpenoid alcohols, compounds such as octanoic acid, decanoic acid, ethyl caprate, beta-damascenone, gamma-decalactone, and vanillin. Notably, several of these overlap with compounds reported in Baijiu empty-cup aroma studies, which examine the fragrance that lingers in a finished glass, lending supporting consistency to the predictions.

The third pillar is structural. Drawing on four established molecular complexity metrics, SCScore, the CM and Cse indices of Proudfoot, and Spacial-Score, the team scored each compound for atom connectivity, branching, ring systems, heteroatom composition, and stereochemical richness, then combined the four normalized metrics with equal weight. Terpenes, terpenoid alcohols, lactones, phenolics, and related aromatics scored highest, reflecting their ring systems and multiple functional groups. This echoes an earlier finding that structurally more complex odorants tend to evoke a broader range of olfactory notes, suggesting molecular architecture may constrain the perceptual space a volatile compound can occupy. Across all three pillars, compound-level scores were aggregated to the sample level using log-transformed concentrations as weights, a choice designed to retain abundance information while preventing a handful of dominant esters from swamping the whole profile.

When the three components were summed into TACI, the 21 samples ranged from 438.69 to 513.57, with the highest value in a high-tier product of brand NXX4 and the lowest in the low-tier product of brand NXX6. Six of the eight brands showed a monotonic increase in TACI with commercial tier, and in several brands the higher-tier products displayed broader contributions beyond the sweet, alcoholic, and fruity core, extending into woody, floral, vegetal, and herbal territory. But the association was not universal. In brand NXX7, the middle-tier product ranked highest on every component index, while NXX6 showed almost no tier separation. The authors are careful to note that commercial grades reflect market positioning rather than standardized quality benchmarks, and that aging and blending can reshape volatile composition in ways that do not necessarily raise the index.

The crucial test came against human perception. Nine professional Baijiu tasters, each with more than eight years of experience, scored six representative samples on a nine-point complexity scale under blind, randomized conditions, having been instructed to judge complexity as an overall attribute of aromatic diversity, layering, harmony, and persistence rather than mere intensity. The correlation between TACI and the panel’s mean scores was strikingly positive: Pearson’s r of 0.84 with a coefficient of determination of 0.70. With only six sample-level observations, however, the authors insist this is an exploratory association, not confirmatory validation. The result is nevertheless a tantalizing sign that an algorithmic digest of gas chromatograms can approximate something as elusive as what a master taster senses in a glass.

The study is candid about its limits. The odor descriptor predictions come from a pretrained black-box model that may not fully translate to perception in a complex ethanol matrix; persistence values derive from fragrance references rather than Baijiu-specific measurements; the three components share the same concentration matrix, which inflates their mutual correlations; and equal-weight integration was chosen as a transparent baseline rather than a claim about perceptual equivalence. OAV-based weighting was deliberately avoided because matrix-matched odor thresholds were not consistently available, though the authors acknowledge that future work should test odor thresholds, alternative descriptor groupings, and component weights. Cumulative uncertainty across the pipeline was not quantified. Still, the framework offers something the field has lacked: a transparent, scalable, chemistry-grounded index that complements rather than replaces sensory evaluation. If validated on larger and more diverse collections, the approach could extend beyond Baijiu to other aroma-rich fermented beverages, giving producers and researchers alike a common quantitative language for one of flavor science’s most stubborn abstractions.

Subject of Research: A computational volatile-based aroma complexity index for strong-aroma-type Chinese Baijiu

Article Title: A volatile-based aroma complexity index for Nongxiangxing baijiu

Article References: Lan, H., Zheng, J., Zhao, D., Su, J., Lu, Y., Chen, S., & Xu, Y. (2026). A volatile-based aroma complexity index for Nongxiangxing baijiu. Food Chemistry: X, 39, Article 104440. https://doi.org/10.1016/j.fochx.2026.104440

Image Credits: AI Generated

DOI: 10.1016/j.fochx.2026.104440

Keywords: baijiu, aroma complexity, food chemistry, volatile compounds, gas chromatography, machine learning, odor prediction, Shannon entropy, sensory evaluation, flavor science, distilled spirits, odor persistence