Source-linked AI summary
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
TL;DR
Open-ended smell descriptions are difficult to translate into physical aromas, and existing reproduction approaches require sensor measurements or molecular specifications. AromaGen maps natural-language descriptions to a 12-odorant wearable system and supports iterative feedback; participants distinguished generated targets above chance, and feedback improved perceived similarity.
Problem
Existing aroma-reproduction approaches require sensor measurements or precise molecular specifications, while translating natural-language smell descriptions into perceptible compositions remains unresolved.
Source: §1 · p. 1, §2.3 · p. 3Method
AromaGen maps free-form natural-language descriptions to compositions over 12 base odorants selected from a semantic olfactory space and rendered by a wearable system with iterative language feedback.
Source: §1 · p. 2, §3 · p. 4, §3.1 · p. 4Results
Participants distinguished generated target smells above chance in a 3-AFC task, and natural-language feedback improved perceived similarity to physical targets.
Source: §6 · p. 21Takeaways & Limitations
AromaGen demonstrates that multimodal LLM olfactory knowledge, grounded in human composition data and a physical odorant basis, can support language-guided generation of smells associated with intended real-world targets.
Source: §6 · p. 21The 50-target benchmark skews toward food and beverage smells, and participants came primarily from two U.S. university communities, limiting generalization across smell types and cultural backgrounds.
Source: §5.4 · p. 21
Abstract
from arXiv · showhide
Smell's connections with food, memory, and social experience have motivated researchers to bring olfaction into interactive systems. Multimodal AI opens new possibilities for generating smells from natural language, yet it remains unclear whether pretrained models encode sufficient olfactory knowledge to translate language into perceptible compositions. We present AromaGen, an AI-powered 12-channel wearable olfactory system that maps free-form natural-language descriptions to 12 base odorants selected to cover a semantically derived olfactory space, while supporting iterative refinement through natural-language feedback. In a between-subjects study (N=60) across 50 real-world benchmark smells, participants distinguished AromaGen-generated target smells above chance in a three-alternative forced-choice task, with retrieval-augmented AI composition performing comparably to human-expert composition and zero-shot AI composition. Natural-language feedback significantly improved the perceived similarity of both AI- and human-expert compositions. Our findings demonstrate the feasibility of using pretrained multimodal AI, with human aroma composition data, to generate perceptible olfactory experiences from natural language.
1 Introduction
AromaGen addresses the challenge of translating open-ended smell descriptions into physically renderable compositions from a compact odorant basis. The introduction presents its 12-channel system, language-based refinement, and evaluation against real-world smell targets.
1 Introduction
- Prior olfactory interfaces rely on fixed scent reservoirs or predefined generation patterns, while broad real-world smell reproduction from a compact odorant basis remains an open challenge.Multimodal models may encode language–olfaction relationships, but whether that knowledge suffices to generate perceptible compositions remains unresolved.Source: §1 · pp. 1–2
- AromaGen maps everyday smell descriptions to compositions over 12 base odorants, with a neck-worn dispenser rendering mixtures that users can iteratively refine through natural-language feedback.The system combines a pretrained multimodal model with olfactory-specific context, human-generated composition data, and retrieval-augmented generation over previously evaluated formulations.Source: §1 · p. 2
- Because olfaction has no agreed set of primaries, AromaGen aligns its 12 channels to olfactory clusters derived from PCA of English smell descriptors.The descriptor pipeline is adapted from Hörberg.Source: §1 · p. 2
- A between-subjects evaluation compares AI-generated compositions with olfactory-expert formulations across 50 real-world benchmark smells, including ratings before and after optional natural-language feedback.The three-alternative forced-choice task tests whether participants associate a generated smell with its intended target rather than whether it is perceptually indistinguishable from that target.Source: §1 · p. 2
- The paper frames AromaGen as grounding multimodal LLM olfactory knowledge in human composition data and a physical odorant space to support language-guided smell generation.Its contributions combine the composition approach, a wearable 12-channel system, and an evaluation of recognition and refinement.Source: §1 · p. 2
2 Related Work
Related work spans AI for olfaction, the structure of smell language, aroma reconstruction, and olfactory interfaces. AromaGen differs by using latent olfactory knowledge in multimodal models to turn natural-language descriptions into physically renderable formulations.
2 Related Work
- The paper situates its approach across multimodal AI, olfactory data modeling, aroma reproduction, and olfactory interfaces.It builds on interactive AI agents that could perceive and generate smells and learn from user feedback, while noting that smell remains largely absent from AI research.Source: §2 · p. 3, §2.1 · p. 3
2.1 Olfaction as an AI Modality
- Rather than predicting perception from molecular features, AromaGen takes a different approach: it leverages olfactory knowledge already encoded in multimodal LLMs to translate free-form language into actionable aroma formulations.Source: §2.1 · p. 3
2.2 Olfactory Language and Semantics
- Despite sparse English aroma vocabulary, olfactory descriptors form a semantic space, and LLMs can recover olfactory-semantic relationships that correlate with human perceptual judgments.The authors adapt Hörberg et al.’s descriptor pipeline to derive 12 clusters used by the hardware palette and evaluation.Source: §2.2 · p. 3
2.3 Aroma Reproduction and Approximation
- Existing aroma-reproduction methods search mixtures over finite component palettes, but require physical sensor measurements or precise molecular specifications rather than everyday language.Examples include sensor-based odor recording and optimization or decomposition methods.Source: §2.3 · p. 3
2.4 Olfactory Interfaces
- Olfactory interfaces span diverse delivery mechanisms and applications, but typically use fixed repertoires that preclude user-specified aroma generation at runtime.AromaGen couples an LLM formulation pipeline with a multichannel olfactory display to provide described aromas on demand.Source: §2.4 · pp. 3–4
3 AromaGen
AromaGen maps natural-language descriptions to mixtures of 12 base odorants and renders them through a wearable display. Its semantic palette, retrieval-augmented composition, and optional feedback loop support smell generation and refinement.
3.1.1 Constructing the Describable Olfactory Space.
- Two stable semantic axes—pleasantness and concrete source versus abstract evaluation—organize the describable olfactory space used to guide base-odorant selection.The axes emerged across vocabulary-selection thresholds; researchers grouped descriptors into 12 perceptually informed clusters.Source: §3.1.1 · p. 4, §3.1.2 · p. 4, Figure 2 · p. 5
3.1.3 Refining the 12 Base Odorants.
- Internal testing found some single oils too specific for their clusters, so six base odorants became fixed composites of two or three essential oils.For example, birch tar oil’s barbecue-like quality suited some smoky targets but was less suitable for coffee and fresh bread.Source: §3.1.3 · p. 4
- The final 12-channel palette replaced Body & Animalic with Savoury & Umami, addressing reference-smell safety concerns and insufficient savory representation.A chef recommended a dedicated savory and umami base; the cluster was excluded because the study required participants to smell real-world references.Source: §3.1.3 · p. 5
3.2 AI Modeling of Aroma Composition
- AromaGen treats composition as structured prediction: an LLM maps a natural-language smell description to a ratio vector over the 12 base odorants.The supported mechanisms are zero-shot composition, retrieval-augmented composition, and natural-language iterative refinement.Source: §3.2 · pp. 5–6
3.2.2 Retrieval-Augmented Composition.
- Retrieval-augmented composition supplies the model with five semantically similar evaluated examples, using human ratings and feedback as precedents rather than copying the nearest formulation.Higher-rated examples serve as positive precedents, while lower-rated examples provide negative evidence and feedback about changes to avoid.Source: §3.2.2 · p. 7
3.2.3 Iterative Refinement.
- Users can smell a generated mixture and refine it with natural-language feedback; the model revises the current formulation while preserving uncriticized aspects.The original target, current ratios, interaction history, and latest feedback inform each revision.Source: §3.2.3 · p. 7
3.3 Interaction Design
- The interaction combines natural-language specification, physical rendering through a multichannel display, and optional refinement from users’ perceptual feedback.The experimenter-facing interface accepts typed or spoken input, visualizes odorant weights, and sends compositions to the wearable.Source: §3.3 · p. 8, §3.3.1 · p. 8
3.3.3 Odor Delivery Mechanism.
- Because only one channel can activate at a time, the controller approximates a composition with interleaved odor pulses allocated roughly in proportion to predicted ratios.Each pulse lasts 0.5 seconds, and presentations with more active odorants take longer than those with fewer.Source: §3.3.3 · p. 9
4 Methods
The methods evaluate AromaGen’s fixed 12-odorant system through expert elicitation, a comparative study of target identification and perceived similarity with natural-language refinement, and an exploratory study of free-form generation. Together, these studies examine benchmark reconstruction and how users engage with language-driven olfactory generation, including its potential for creative or personally meaningful experiences.
4.1 Expert Elicitation Study.
- Three olfactory experts independently composed all 50 benchmark smells from the same 12 bases, iteratively adjusting odorant ratios after comparing generated aromas with physical targets.They could use any number of bases, rated similarity on a 7-point scale, and were encouraged to finish within three refinement rounds.Source: §4.1 · p. 9, §4.1.4 · p. 10
4.1.2 Benchmark Smell Preparation.
- Physical benchmark odors were refreshed according to their material properties, and experimenters checked each source’s detectability immediately before presentation, replacing insufficient samples.The 12 base odorants were prepared in labeled vials and AromaGen cartridges were refreshed before expert elicitation.Source: §4.1.2 · p. 10, §4.1.3 · p. 10
4.1.5 Expert Elicitation Results.
- Experts’ lowest-rated compositions varied across clusters, and the final baseline selected each target’s highest within-expert-normalized formulation, with contributions nearly equal across experts.Expert 1’s lowest cluster was Fermented & Sour (M=2.50), Expert 2’s was Roasted & Smoky (M=4.50), and Expert 3’s was Putrid & Decay (M=2.75); the final set included 17, 16, and 17 formulations, respectively.Source: §4.1.5 · p. 11
4.2.2 Study Design.
- In the comparative study, 60 participants were assigned across three 20-person conditions: RAG-based AromaGen with optional feedback, human-expert compositions with optional feedback, and zero-shot AromaGen without feedback.The zero-shot arm used the same LLM, prompt, temperature, and output representation as the RAG-based arm but omitted retrieved examples and refinement.Source: §4.2.1 · p. 11, §4.2.2 · p. 11
4.2.3 Procedure.
- Each participant evaluated 12 targets, one per olfactory cluster, through 3-AFC identification and a 7-point similarity rating, followed by optional natural-language refinement in the two feedback conditions.The 3-AFC candidates were the generated target, a generated distractor, and a physical distractor; feedback was transcribed verbatim and used with the target descriptor and current composition to generate a revision.Source: §4.2.2 · p. 11, §4.2.3 · pp. 12–13
4.2.5 Analysis.
- Analyses used crossed participant-and-target random intercepts for identification, similarity, and feedback changes; a separate five-person exploratory study examined free-form generation beyond the benchmark.Identification used a binomial GLMM, ratings used linear mixed models, and interviews documented participants’ overall experiences.Source: §4 · p. 9, §4.2.5 · pp. 13–14
4.3.1 RQ1: People Identified Generated Smells Above Chance.
- 68.1% of benchmark trials were correctly identified, and all three composition conditions exceeded chance without significant pairwise accuracy differences.AromaGen scored 70.8%, Expert 69.6%, and AromaGen zero-shot 63.7%; the study does not establish equivalence or perceptual indistinguishability.Source: §4.3.1 · p. 14, §4.3.2 · p. 14, Figure 10 · p. 15
4.3.3 RQ3: Effect of Spoken Feedback on Perceived Similarity.
- Feedback increased mean similarity by 0.56 points for AromaGen and 0.26 for Expert across all trials, with larger gains on trials where participants requested revisions.On revised trials, similarity rose by 0.82 for AromaGen and 0.54 for Expert; the difference in feedback benefit between conditions was not significant (p = .058).Source: §4.3.3 · pp. 14–15
4.3.4 Qualitative analysis of feedback patterns.
- Participants refined smells by requesting directional perceptual changes and relational comparisons, often expressing uncertainty and combining increases with reductions.They used “more” 278 times versus “less” 60 times; feedback could steer smells without naming odorants or specifying compositional changes.Source: §4.3.4 · p. 15, Figure 11 · p. 15
4.3.5 Across Category Analyses.
- Identification varied by smell family: chemical and solvent smells were easiest at 83%, while sweet and gourmand were hardest at 45%; familiarity showed no accuracy or similarity association.Correct identification was associated with 0.70 points greater confidence and higher similarity ratings (4.06 vs. 3.51).Source: §4.2.5 · p. 14, §4.3.5 · p. 16, Figure 12 · p. 17
4.3.6 Compositional Analyses.
- AromaGen and zero-shot compositions diverged similarly from expert compositions, with total variation distances of 40.5 and 41.6, respectively.Some targets converged across conditions, while others showed distinct ingredient choices, including black pepper, cucumber, aloe vera, pickled cucumber, and nacho with sour cream.Source: §4.3.6 · p. 16, Figure 13 · p. 18
4.4 Exploratory Free-form Aroma Generation Study.
- The exploratory study recruited five experienced participants for free-form generation, allowing them to create ten chosen smells, refine them through language, and discuss capabilities and failures in interviews.Participants had creative or smell-related backgrounds; interview analysis was descriptive and not intended to support generalizable claims.Source: §4.4 · p. 16, §4.4.1 · p. 16, §4.4.2 · p. 17
4.4.4 Preliminary Findings from Free-form Aroma Generation.
- Among 32 of 50 free-form targets that participants refined, mean similarity rose from 2.94 to 4.09, and none ended with a lower rating.Across all 50 targets, final mean similarity was 4.16; reconstruction varied substantially and was lower-fidelity than for the benchmark smells.Source: §4.4.4 · pp. 17–18
- In the exploratory study, generated smells sometimes prompted imaginative exploration and personal associations, including memories, though these small-sample cases do not establish general effects.Participants described difficulty communicating individualized memories, while one participant connected a generated “Sister’s hug” aroma with homesickness and memories of her sister.Source: §4.4.4 · p. 19
- Future work should examine how distinctiveness, complexity, linguistic describability, available odorants, and user expectations jointly shape reconstruction quality.These preliminary findings suggest directions for investigation rather than definitive conclusions about AromaGen’s free-form capabilities.Source: §4.4.4 · p. 19
- Further study should investigate users’ mental models of language-driven olfactory generation and whether generated smells support creative exploration or personally meaningful experiences.Interpret these observations in light of the small sample and participants’ prior AromaGen experience.Source: §4.4.4 · p. 19
5 Discussion
AromaGen divides olfactory creation between an LLM that composes candidate smells and users who evaluate and steer them, a division that may reduce expertise demands. The discussion also identifies limits in the odorant palette, refinement interface, and study scope.
5.1 Human–LLM Collaboration for Olfactory Creation
- Users can evaluate candidate smells even when composing or explicitly identifying a target is difficult, motivating a collaboration in which the LLM composes and users steer.Users describe perceptual discrepancies in natural language rather than specifying odorants and ratios.Source: §5.1 · p. 19
5.2 Toward Accessible Personalized Aroma Generation
- Natural-language descriptions and feedback could reduce the expertise, materials, and manual formulation effort needed to create and personalize aromas.An illustrative internal session produced formulations for only nine targets in one hour using manual pipetting and mixing; it was not part of the controlled evaluation.Source: §5.2 · p. 20, §5.1 · p. 20
5.3 Familiarity Is Not the Same as Knowing a Smell
- Self-reported familiarity predicted neither 3-AFC performance nor perceived similarity, but the measure may reflect familiarity with a labeled concept rather than independent odor recall.Participants smelled the physical target and received its identity immediately before judging candidates.Source: §5.3 · p. 20
5.4 Limitations and Future Work
- Single-channel pulses blend odorants only after release, limiting how faithfully the neck-worn device can approximate a target smell.Concurrent release through a mixing chamber or multiple outlets is proposed as a future hardware direction.Source: §5.4 · p. 20
- The 12-odorant basis cannot represent some targets, and the study did not separate palette-coverage failures from poor composition.The authors identify a richer or strategically chosen palette, including task-specific cartridge swaps, as a possible response.Source: §5.4 · pp. 20–21
- Verbal refinement captures users’ evaluations only to the extent they can describe them, so candidate-mixture preference judgments could reduce the language demand.The passage notes that English speakers rarely have dedicated smell terms and often find familiar aromas difficult to name.Source: §5.4 · p. 21
- Retrieval stores perceptual feedback as in-context examples but does not update model weights, so the model cannot improve its olfactory representations across sessions.The study provides 278 spoken refinement rounds and paired before/after similarity ratings as potential preference-training data.Source: §5.4 · p. 21
- The evaluation’s scope is limited by a food- and beverage-skewed 50-target benchmark, participants primarily from two U.S. university communities, and unmeasured fatigue and session-order effects.Essential oils are also imperfect proxies for real-world sources, potentially affecting composition quality and verbal feedback.Source: §5.4 · p. 21
6 Conclusion
AromaGen maps free-form descriptions to physical compositions over a 12-odorant basis and supports natural-language refinement. Participants distinguished generated targets above chance, while the findings do not establish perceptual indistinguishability or superiority to human experts.
6 Conclusion
- AromaGen translates free-form descriptions into physically rendered compositions over 12 odorants, with iterative natural-language refinement.The authors report above-chance discrimination and improved perceived similarity with feedback, but do not claim perceptual indistinguishability or superiority to human experts.Source: §6 · p. 21
7 GENERATIVE AI DISCLOSURE
The authors disclose using generative AI tools to refine author-drafted manuscript phrasing, with all generated or refined text reviewed, revised, and approved by the authors.
7 GENERATIVE AI DISCLOSURE
- Generative AI tools refined phrasing in author-drafted text, and the authors reviewed, revised, and approved all generated or refined text.The authors take responsibility for the paper’s accuracy and integrity.Source: §7 · p. 21
A Describable Olfactory Space Construction
The authors adapted a corpus-based approach to build a semantic representation of English olfactory descriptions from a large text corpus.
A Describable Olfactory Space Construction
- The construction adapted Hörberg et al.’s corpus-based approach, using 2.97 billion tokens and extracting 92,120 sentences associated with olfactory and gustatory seed terms.Source: A Describable Olfactory Space… · p. 23
A.1 Olfactory Descriptor Extraction
The descriptor-extraction process identified candidate olfactory words from corpus contexts around seed terms and yielded 3,558 candidates.
A.1 Olfactory Descriptor Extraction
- The process retained content words found within a ±4-word window of seed terms when they appeared at least 10 times.Candidates were then evaluated using an Olfactory Association Index (OAI), measuring their association with olfactory contexts relative to their overall occurrence.Source: §A.1 · p. 23
- The extraction process yielded 3,558 candidate descriptors.Source: §A.1 · p. 23
A.2 Semantic Representation and Dimensionality Reduction
A 300-dimensional CBOW Word2Vec model represented descriptors, and PCA was applied to a standardized semantic distance matrix to identify dominant dimensions in English olfactory vocabulary.
A.2 Semantic Representation and Dimensionality Reduction
- Descriptor embeddings from a 300-dimensional CBOW Word2Vec model were compared using Pearson correlations to form pairwise semantic distances.Source: §A.2 · p. 23
- The descriptor-by-descriptor distance matrix was standardized before PCA identified dominant dimensions in English olfactory descriptors.Source: §A.2 · p. 23
A.3 Robustness Across Olfactory Association Thresholds
The analysis tested whether semantic dimensions persisted across different OAI cutoffs and assessed component stability with bootstrap iterations and external semantic variables.
A.3 Robustness Across Olfactory Association Thresholds
- The analysis was repeated using the top 5%, 10%, 15%, 20%, 25%, and 30% of descriptors ranked by OAI to test dimension stability as less olfactory-specific vocabulary was included.Source: §A.3 · p. 23
- Component stability was evaluated through 200 bootstrap iterations, with components interpreted against external semantic variables including valence, arousal, and sensory experience.Other variables included sensory-modality associations, concreteness, and imageability.Source: §A.3 · p. 23
A.4 Resulting Semantic Dimensions
The supplement reports two stable semantic dimensions and describes expert input and prompt procedures for composing, retrieving, and revising odorant blends. These procedures distinguish concrete-smell matching from abstract-scene composition and use prior evaluations and user feedback as composition guidance.
A.4 Resulting Semantic Dimensions
- Two stable dimensions organize olfactory descriptions by pleasantness and by concrete odor sources versus abstract or evaluative descriptions.The paper retains these dimensions as an operational two-dimensional representation of AromaGen’s describable olfactory space.Source: §A.4 · p. 23
B Expert Interview Protocol
- Expert interviews addressed mixture versus sequential release, perception duration, ingredient counts, everyday-aroma coverage, and odorant volatility.The protocol also asked whether essential oils are commonly used in fragrance blending and whether the base odorants cover formative-study categories.Source: B Expert Interview Protocol · pp. 23–24
Base Composition Prompt
- The base-composition prompt targets perceptual similarity for concrete smell requests and emotional resonance for abstract moods, memories, or sensory scenes.For abstract requests, the prompt directs the model to deconstruct the description into ordered beats and map each beat to odorants.Source: Base Composition Prompt · p. 24
RAG Prompt Augmentation
- Retrieved compositions provide prior ratios, similarity ratings, and feedback; their influence depends on both their rating and semantic similarity to the current request.Closely matching precedents are starting points, while weak precedents are treated as evidence of an unsuccessful or incomplete composition.Source: RAG Prompt Augmentation · pp. 24–25
Iterative Refinement Prompt
- The revision prompt prioritizes the latest feedback, preserves uncriticized odorants, and favors targeted ratio changes before rebalancing the blend.It requires all 12 odorants in the output and ratios summing to exactly 1.00.Source: Iterative Refinement Prompt · p. 25
D Participant Demographics
The passage reports demographic characteristics for study participants and describes the material types used for AromaGen’s odorants.
D Participant Demographics
- Among 63 respondents, ages ranged from 18 to 64; 37% did not primarily speak English, and self-rated smell expertise averaged 2.75/5 (SD 1.02).Source: D Participant Demographics · p. 25
- Forty-three participants identified as women and 20 as men; the largest reported ethnic-background groups were Asian and East Asian (43%), Hispanic and Latino (25%), and White and European (21%).Source: D Participant Demographics · p. 25
- Demographic data were available for 20 participants in the AI+RAG and Expert conditions and 19 in the zero-shot AI condition.One zero-shot AI participant did not complete the intake form.Source: Table 4 · p. 25, D Participant Demographics · p. 25
F Odorant Material Types
- The odorant materials include essential oils, tinctures, fragrance oils, pure aroma chemicals, and powdered plant material, with differing sources and aroma characteristics.Source: F Odorant Material Types · p. 25
G Comparative Study Sampling Procedure
Each participant sampled one target from every olfactory cluster, with balanced target presentation and distractors drawn from other clusters.
G Comparative Study Sampling Procedure
- Each participant completed 12 trials, sampling one target from each cluster; balanced sampling distributed the 50 benchmark smells approximately evenly across participants.Each cluster contained four or five benchmark targets.Source: G Comparative Study Sampling Procedure · p. 25
- Targets were sampled without replacement within each cluster until all items had been presented, then the pool reset; trial order was randomized for each participant.Source: G Comparative Study Sampling Procedure · p. 25
- Each trial paired the target with one generated and one physical distractor from other clusters, with distractor sampling balanced across trials.Repeated target presentations were generally paired with different distractors.Source: G Comparative Study Sampling Procedure · p. 26
H Hardware Specification
The supplied passage identifies Figure 15 as a hardware specification sheet but gives no hardware specifications.
H Hardware Specification
- Figure 15 is titled “Hardware Specification Sheet.”Source: Figure 15 · p. 26