AromaGen: Interactive Generation of Rich Olfactory Experiences with Multimodal Language Models
Awu Chen, Silang Wang, Yunge Wen, Jianing Yu, Zinab Eisa, Ana Julia Santos Lopes, Jas Brooks, Hiroshi Ishii, Paul Pu Liang
cs.HC, cs.AI
2026-04-02
Maps natural-language smell requests to 12-odorant blends on a neck-worn display; 70.8% blind identification matches expert blends, and language feedback adds +0.56 similarity.
HCI researchers have spent decades trying to put smell into interactive systems: scent in VR, story-driven aromas in games, olfactory health interventions. The hardware never moved past fixed cartridges and pre-programmed release patterns, and users cannot request a smell at runtime. The blocker is structural: vision has RGB primaries and hearing has frequency decomposition, but olfaction has no known set of primaries, so synthesizing arbitrary smells from a small basis has no textbook recipe.
AI research on smell has mostly been classification, from electronic-nose sensors, GC–MS measurements, or molecular structure. Generation systems such as "odor recorders" need sensor readings or exact molecular specs of a target sample, so none of them accept the language people actually use to describe smells. Whether a pretrained multimodal LLM holds enough olfactory knowledge to turn "the smell of a cheeseburger" into a physically realizable recipe was an open question.
Define the generative space from language. From the 2.97-billion-token UMBC WebBase corpus the team extracted 92,120 smell-related sentences and 3,558 candidate descriptors, then ran Word2Vec embeddings through PCA. Two stable axes emerged: pleasant versus unpleasant, and concrete odor source versus abstract evaluation. The 317 strongest descriptors were grouped into 12 clusters (floral, citrus, woody, herbal, spice, sweet, roasted-smoky, fermented, putrid, chemical, clean, umami), and professional perfumers chose a representative essential oil for each; internal testing later turned six channels into fixed composites of two or three oils, with the roasted-smoky channel split evenly between birch tar, coffee, and clove bud. A body/animalic cluster was dropped for lab hygiene and safety reasons and replaced with savory/umami.
LLM composition. The task is structured prediction: natural language in, a 12-dimensional non-negative ratio vector summing to 1 out, typically two to six active odorants. The zero-shot baseline uses a system prompt describing each base. The full pipeline adds RAG: three researchers first evaluated AI compositions for all 50 benchmark smells, smelling each physical reference, rating similarity on a 7-point scale, and giving feedback over iterative revisions. Each database entry stores the target, ratio vector, rating, and feedback. A new request retrieves the top-5 most similar entries as context, with high-rated compositions as positive precedents and low-rated ones plus their feedback as negative evidence. The database keeps growing as people use and correct the system. After smelling a result, a user can say "less sweet" and the model revises the existing vector in context, closing a describe-generate-smell-refine loop.
Hardware. The neck-worn 12-channel device can open only one channel at a time, so a composition is rendered as 10 rounds of interleaved 0.5-second pulses, with each odorant's pulse count proportional to its predicted ratio.
Three experts (a barista and taste researcher, an executive pastry chef, a perfumer) each composed all 50 benchmark smells with the same 12 bases, averaging 3.33 hours each; the best formulation per target, picked by z-score, formed the human baseline. Then 60 participants, 20 per condition, each ran 12 trials of a three-alternative forced-choice task: after smelling the real target, pick the closest of the generated target, a generated distractor, and a physical distractor. Chance is 33.3%.
| Condition | 3-AFC accuracy | Similarity before feedback | After feedback |
| AromaGen (RAG + feedback) | 70.8% | 4.09 | 4.65 |
| Human expert + feedback | 69.6% | 3.79 | 4.05 |
| AromaGen zero-shot | 63.7% | 3.77 | no feedback stage |
All three conditions beat chance (p<.001); no pairwise difference was significant. The paper stresses this does not establish AI-expert equivalence, only that within these 50 smells and the fixed palette, swapping expert recipes for AI ones did not hurt recognizability. Feedback delivered real gains: AromaGen participants gave feedback on 68% of trials, which rose from 3.50 to 4.32 (+0.82); expert-condition feedback covered 48% of trials, rising from 3.48 to 4.02 (+0.54). Whether feedback helped the AI more just missed significance (p=.058).
Clusters diverged widely: chemical/solvent smells were identified 83% of the time and sweet/gourmand only 45%; mint fluoride mouthwash went 17/17 while cinnamon roll managed 4/14. AI and expert recipes were far from identical (total variation distance around 40) yet landed in similar places. In a five-person free-form study, 32 refined targets climbed from 2.94 to 4.09 (+1.16), with none ending lower.
The core evidence: olfactory knowledge latent in a pretrained multimodal LLM, grounded by a compact physical palette and a modest amount of human evaluation data, reaches expert-level composition without training an olfactory model. The value is not beating perfumers; it is collapsing hours of manual formulation into a sentence plus a few rounds of feedback, which changes the economics for VR, retail, and narrative experiences that need smell on demand. For model researchers it adds evidence that text-trained models carry usable crossmodal structure, here for a sense with almost no training data. Measured against the earlier language-to-blend paradigm, this is solid systems progress, not a new paradigm.