SCENT: Olfactory Signaling, Computation, & Interaction

• Scent is an olfactory stimulus foundational to early mammalian development, social communication, and biological signaling with diverse applications.
• It integrates complex physical dynamics, where advection and turbulence enable effective odor tracking beyond simple diffusion models.
• Computational and experimental frameworks merge to model, generate, and operationalize scent for interactive systems and machine olfaction.

Scent denotes an olfactory stimulus and, more broadly, a class of biological, physical, computational, and interactive phenomena centered on odor perception, odor signaling, and odor-mediated inference. In the literature considered here, scent appears as the earliest functional sensory modality in mammalian development, a substrate for territorial and social communication, a difficult physical signal whose environmental transport cannot be reduced to simple diffusion, and an increasingly important computational object represented through molecular graphs, electronic noses, and electron ionization mass spectra. The same literature also uses SCENT as the name of several technical frameworks, especially in machine olfaction and multimodal learning (Wagner et al., 2019, Biswas et al., 2024, Tsonis et al., 7 Jul 2026, Zhang et al., 26 May 2026).

1. Biological modality and perceptual structure

In mammals, olfaction is the first sensory modality to develop during fetal life, and neonatal survival depends heavily on odor-guided behaviors such as nipple localization, feeding, maternal attachment, nest recognition, huddling, and avoidance of danger (Wagner et al., 2019). This early functional status makes scent unusually important in preweaning animals, where hearing and vision are less useful and behavioral repertoires are limited. The neonatal mouse assay described in the literature exploits exactly this fact: isolated pups emit abundant ultrasonic vocalizations, and odor cues modulate those calls in ways that can be used as an olfactory readout rather than as a trained discrimination task (Wagner et al., 2019).

At the level of human report, smell experience is described as subjective, difficult to verbalize, and strongly tied to memory, emotion, and personal context. Everyday odor descriptions often mix sensory attributes, source-based references, affective terms, and autobiographical associations rather than relying on a stable standardized vocabulary (Zhong et al., 2024). This complicates both psychophysical measurement and AI modeling. A plausible implication is that scent is not merely a chemical input channel; it is also a language-poor perceptual domain in which semantic report and perceptual structure only partially coincide.

2. Social, territorial, and ecological signaling

Scent functions as an active social signal in multiple species. In free-ranging dogs, urine scent marks are not treated as simple elimination by-products or purely sexual signals; they mediate territorial defence, social assessment, intrasexual competition, and likely aspects of group identity and cohesion (Biswas et al., 2024). Dogs investigated scent-mark cloths longer than controls, and neighbouring-group scent marks longer than own-group scent marks. Males showed higher territorial response scores than females, and overmarking was predominantly observed in males, especially toward male scent marks and neighbouring-group marks (Biswas et al., 2024). The strongest distinct response pattern was elicited by neighbouring-group male scent, indicating that scent identity is interpreted jointly through sex and group membership.

A complementary theoretical account appears in a PDE model of territorial pattern formation mediated purely by scent marking. In that model, territories can emerge without any attractive force toward a den, nest, or previously visited center; conspecific avoidance via active scent marks is sufficient, provided that scent does not decay faster than it takes the animal to traverse the terrain and that animals respond to scent averaged over a finite perceptual neighborhood rather than pointwise scent alone (Potts et al., 2015). The model explicitly distinguishes territory, represented by active scent fields, from home range, represented by utilization distributions. This establishes a non-speculative link between individual scent-marking rules and population-level spatial segregation (Potts et al., 2015).

3. Physical transport, detectability, and navigation by scent

As a physical signal, scent is difficult to track if one assumes only molecular diffusion. The diffusion analysis in the literature formalizes odor tracking through two simultaneous conditions: the concentration must exceed a detection threshold, $C(x) > C_T$, and the concentration ratio across a spatial offset must exceed a directional threshold, $\frac{C(x)}{C(x+\Delta)} > R$ (McCaul et al., 2020). For realistic parameters, purely diffusive models cannot satisfy both conditions simultaneously: fields that are detectable at useful distances have gradients that are too shallow to follow, whereas fields with strong gradients require implausibly large source fluxes. The paper therefore concludes that realistic odor tracking requires full fluid dynamics, especially advection and turbulence, rather than diffusion alone (McCaul et al., 2020).

This physical constraint directly motivates engineering approaches to olfactory navigation. “Olfactory Inertial Odometry” defines an OIO framework that extrapolates principles from SLAM and VIO to scent-guided robotics by combining fast-sampling olfaction sensors with inertial kinematics and robot motion information (France et al., 3 Jun 2025). The demonstrated system used a real 5-DoF robot arm, metal oxide and electrochemical sensing, bout-style temporal signal processing, and source localization algorithms based on closed-loop gradient following, an RSSI-inspired belief map, and Expected SARSA. In the reported trials, the reinforcement-learning controller was fastest on average, though also most variable, and all tasks localized the source without timing out (France et al., 3 Jun 2025). This suggests that scent navigation becomes tractable when odor measurements are interpreted as motion-conditioned time series rather than as isolated concentration readings.

4. Experimental olfactory assessment in development and disease

A particularly concrete laboratory implementation of scent measurement is the neonatal mouse sono-olfactometer assay. The apparatus consists of a sealed pup chamber with inside dimensions of $18 \times 12 \times 12$ cm, an odor-delivery system with controlled flow rates, and an ultrasonic recording chain centered on the 40–120 kHz range of mouse pup vocalizations (Wagner et al., 2019). The protocol uses three analytically distinct periods within a 5-minute isolation session: a baseline minute without odorant, a one-minute odor exposure period, and a 90-second exhaust/washout period chosen to allow complete elimination of exhaust air containing odorants (Wagner et al., 2019).

The assay uses odor-evoked inhibition of isolation-induced ultrasonic vocalizations as an index of odor detection. Two inhibitory scents were used: citral, prepared as “1 ml citral ad 10 ml mineral oil dilution,” and adult male mouse scent operationalized as 10 g of soiled bedding from a group of 6 unfamiliar adult OF1 males (Wagner et al., 2019). Calls were counted manually in Audacity and expressed as calls/min; within-animal odor effects were analyzed with the Wilcoxon matched-pairs signed rank test, with significance annotations including $P < 0.01$ and $P < 0.0001$ (Wagner et al., 2019).

Biologically, the key result is that control pups decreased call emission in response to both citral and male scent, whereas pups congenitally infected with murine CMV showed impaired USV inhibition to both odor types, indicating defective olfactory perception (Wagner et al., 2019). The assay detected CMV-related olfactory dysfunction by day 6 after birth, before hearing deterioration appears. Its strengths are that it is rapid, non-trained, suitable for very young pups, and compatible with BSL-1, BSL-2, or BSL-3 constraints; its limitations are that the readout is indirect, depends on robust baseline calling, and can be confounded by arousal, sickness behavior, or general vitality (Wagner et al., 2019).

5. Computational representation of scent

Recent computational work treats scent as a representation-learning problem. In molecular QSOR, graph neural networks were shown to outperform classical descriptor-based baselines on a 5,030-molecule expert-labeled dataset with 138 odor descriptors, and the resulting 63-dimensional penultimate-layer embedding behaved as a transferable odor space rather than as a task-specific classifier (Sanchez-Lengeling et al., 2019). Transfer to unseen descriptors and to the DREAM olfaction benchmark supported the claim that learned graph embeddings capture meaningful perceptual structure (Sanchez-Lengeling et al., 2019).

Generative odorant discovery extends this representational perspective into molecular design. A QSAR-guided VAE trained on approximately $5 \times 10^5$ ChemBL molecules used an external odor classifier to shape latent space toward odor probability, yielding 100% validity via rejection sampling, 94.8% unique structures, and 74.4% “Uncharted Scaffold Hops” distinct from the training-set cores (Pearce et al., 28 Dec 2025). The generated set had Fréchet ChemNet Distance $\approx 6.96$ to an external unseen odorant set, compared with $\approx 21.6$ to the ChemBL baseline, suggesting concentration in an odorant-like region of chemical space rather than simple valid-molecule generation (Pearce et al., 28 Dec 2025).

Machine olfaction has also shifted from explicit molecular structure to alternative sensing modalities. One SCENT framework aligns electron ionization mass spectrometry embeddings with frozen molecular structure embeddings so that only EI-MS is needed at inference time; on GS-LF odor prediction, SCENT improved over the MS-only EIMS2Vec baseline from $87.98$ to $89.99$ Micro-AUC, from $\frac{C(x)}{C(x+\Delta)} > R$0 to $\frac{C(x)}{C(x+\Delta)} > R$1 Weighted-AUC, and from $\frac{C(x)}{C(x+\Delta)} > R$2 to $\frac{C(x)}{C(x+\Delta)} > R$3 Adjusted Precision@5 (Zhang et al., 26 May 2026). Another SCENT framework, “Semantic Context-aware e-Nose Transformer,” addresses vision–olfaction alignment on the New York Smells dataset by using VLM-generated descriptors of the touched object, environmental context, and plausible ambient smell cues as a semantic bridge. With full object + context + inferred smell supervision, it reports $\frac{C(x)}{C(x+\Delta)} > R$4 for smell-to-image, $\frac{C(x)}{C(x+\Delta)} > R$5 for smell-to-text, and $\frac{C(x)}{C(x+\Delta)} > R$6 for joint smell-to-image-text retrieval, outperforming vision-only or bridge-based baselines (Tsonis et al., 7 Jul 2026). Together, these systems recast scent as a multimodal latent variable that can be grounded by chemistry, spectra, language, and context.

6. Scent in interactive systems, media, and human–AI alignment

Scent is also being operationalized as an interaction modality. In audiovisual odor prediction, OlfactProfile treats “the right scent” for media as observer-dependent rather than content-determined, using a benchmark of 1,350 clips, a 99-class scent vocabulary, and three odor tracks—Foreground Odor, Background Odor, and Emotion Odor (Lou et al., 16 Jun 2026). Its central finding is negative as well as positive: olfactory profiles are not beneficial by default, and naive profile concatenation or uniform profile modulation can degrade performance, whereas structured field-wise profile conditioning improves prediction, with the strongest gains on Background Odor and Emotion Odor (Lou et al., 16 Jun 2026).

In virtual reality, multisensory portal experiments used scene-congruent scents such as coffee, orange, flowers, and pizza, delivered by a wearable four-scent near-nose device with active ventilation (Bak et al., 14 Sep 2025). The fully multisensory VAO condition produced the highest accuracy at 98.519%, the fastest completion time at 12.926 s, and the highest confidence at 6.844; by contrast, visual + olfactory cues alone improved means numerically over visual-only but did not produce significant pairwise gains over vision-only on accuracy or time (Bak et al., 14 Sep 2025). Participants nevertheless reported that olfactory cues were harder to distinguish, more fatiguing, and vulnerable to lingering-odor interference.

Workplace olfactory physicalization pushes scent further into peripheral interaction. AuraDesk translated smartwatch-derived physiological cues into localized desk-side scent output through an eight-channel atomization device and a hybrid mapping that combined local Arousal–Valence inference with constrained actuation, including a minimum 15-minute cooldown and one active channel at a time (Hu et al., 1 Apr 2026). In a one-day in-situ deployment with 25 knowledge workers, participants often interpreted scent not as an explicit alert but as a subtle atmospheric cue supporting awareness, micro-break taking, and environmental attunement (Hu et al., 1 Apr 2026).

Human–AI perceptual alignment remains limited. In a 40-participant “sniff and describe” study, an embedding-based system achieved a 27.50% success rate for direct scent description and 37.50% for comparative scent description, with marked biases toward lemon and peppermint and persistent failure on rosemary (Zhong et al., 2024). The study distinguishes semantic validity from perceptual similarity: AI guesses could be rated valid relative to a verbal description while still being perceptually dissimilar to the target scent. This suggests that current LLM embedding spaces encode some odor semantics without fully reproducing human olfactory organization (Zhong et al., 2024).

7. Acronymic and metaphorical uses of “SCENT”

In current technical literature, SCENT is not a single standardized term but a recurrent acronym attached to distinct methods in different domains.

Usage	Expansion or meaning	Domain
SCENT (Zhang et al., 2018)	Sequence-to-sequence ConvErsion NeTwork	Voice conversion
SCENT (Park et al., 16 Apr 2025)	Scalable Conditioned Neural Field for SpatioTemporal Learning	Continuous scientific data
SCENT (Zhang et al., 26 May 2026)	Spectrum-to-Chemical Embedding alignmeNT	EI-MS-based olfactory prediction
SCENT (Tsonis et al., 7 Jul 2026)	Semantic Context-aware e-Nose Transformer	Vision–language–olfaction learning

Not all uses are acronymic. In “Follow the Scent,” scent is a metaphor for the stable trace left by legacy EUI-64 interface identifiers in IPv6 edge routers; the paper explicitly states that it does not define a system called SCENT (Rye et al., 2021). In “the scent of long-range order,” the phrase denotes a remnant of zero-temperature ordered structure that persists locally in low-temperature prethermal dynamics rather than literal odor (Alba et al., 2017). In ASRank, “answer scent” names a generated textual prior over what an answer should represent and functions as a re-ranking signal rather than an olfactory concept (Abdallah et al., 25 Jan 2025). This multiplicity of uses makes SCENT both a genuine olfactory term and a productive cross-domain metaphor for traces, latent structure, and guided search.