Taste: An Interdisciplinary Overview

• Taste is an interdisciplinary concept encompassing gustatory sensation, computational modeling, molecular design, and aesthetic judgment.
• Neural and computational studies transform sensory inputs into quantifiable data using EEG, deep learning, and multimodal analyses.
• Research spans engineered peptides and 3D food printing to acronymic TASTE systems in astronomy, MBSE, recommendation, and sentiment analysis.

TASTE is an interdisciplinary term in contemporary research. In one sense it denotes gustation and its computational, molecular, neural, linguistic, and design-theoretic representations; in another, it functions as an acronym for domain-specific systems in astronomy, model-based systems engineering, recommendation, stance detection, spoken language modeling, aspect sentiment triplet extraction, and agent benchmarking. Across these literatures, taste is treated variously as a receptor-mediated sensory modality, a measurable brain signal, a target for peptide and food design, a spatially programmable property of fabricated foods, a named entity in recipes, an image-inferred multisensory expectation, a crossmodal bridge to sound, and a human-centered construct for modeling stylistic perception (Yue et al., 13 Feb 2025).

1. Conceptual range of taste

In gustatory research, taste peptides are short amino-acid chains that directly elicit taste sensations when they contact taste receptors. In the dataset curated for TastePepAI, 88.54% of known taste peptides are shorter than 15 amino acids, and peptides $\ge 15$ residues were excluded from modeling to reduce noise. The same study emphasizes that taste is rarely single-purpose: among 1131 curated taste peptides, over 30% have multiple taste properties, including dual, triple, and even four-taste combinations, and sequence similarity analysis shows no clear boundary in sequence space between categories such as sweet and umami (Yue et al., 13 Feb 2025).

The gustatory literature in the supplied corpus treats the canonical basic tastes in partly different ways depending on task definition. Taste EEG studies focus on sour, sweet, bitter, and salty as experimental stimuli, with controlled delivery to the tongue and classification of the resulting neural activity (Xia et al., 2023). Ingredient-level food modeling uses five dimensions—sweetness, sourness, bitterness, umami, and saltiness—on the trained-panel SVT 0–100 Spectrum scale (Tagkopoulos et al., 22 Apr 2026). FoodSense, by contrast, operationalizes taste as a scalar evaluative dimension of likely taste quality inferred from food images, with ratings rescaled from a 1–7 Likert scale to 1–5, plus free-text descriptors such as “sweet,” “savory,” “spicy,” and “bland” (Ishraq et al., 15 Apr 2026). This divergence indicates that “taste” can refer either to basic taste quality, to perceived intensity, or to expected hedonic quality, depending on the measurement regime.

The term also extends beyond gustation. In TASTEset, TASTE is a named-entity category in recipe NLP defined as “the flavour (e.g. bittersweet, butter-flavoured, sweet, semi-sweet)” (Wróblewska et al., 2022). In a distinct human-centered design framework, taste denotes stylistic aesthetic perception rather than sensory flavor, operationalized through labels such as aerodynamic, classic, dynamic, elegant, futuristic, luxury, rugged, sleek, and sporty (Hong et al., 23 Jan 2026). This suggests that contemporary research uses taste both for chemical sensation and for structured judgments of style.

2. Taste as a neural and computational signal

Taste-evoked EEG studies in the supplied literature treat gustation as an objectively decodable brain state. One study collected taste EEG from 20 right-handed, non-smoking adults under sour, sweet, bitter, and salty stimulation, using 21 electrodes, 10 s taste windows segmented into five non-overlapping 2 s samples, yielding 1600 labeled samples. A Temporal and Spatial Reconstruction Data Augmentation method coupled with a convolutional network with multi-view channel attention reached an accuracy of 99.56%, an F1-score of 99.48%, and kappa of 99.38%, indicating that short scalp EEG segments contain highly discriminative information about taste condition (Xia et al., 2023).

A complementary study addresses the channel-selection problem for taste EEG. Using 21 EEG channels, a convolutional neural network with channel and spatial attention combined with Grad-CAM was used to rank channels by activation relevance. With all 21 channels, the model achieved 97.99% accuracy and 97.95% F1-score; with 12 channels selected by CAM-Attention, it retained 97.55% accuracy and 97.52% F1-score while reducing FLOPs by 41.18%. The same work reports that different tastes produce distinguishable activation patterns, with prominent contributions from frontal midline, temporal front, occipital, and central regions (Xia et al., 2024).

At a more mechanistic level, taste perception is modeled as a multiscale process spanning receptor-cell biophysics to population spike coding. A hybrid neuron framework combines Hodgkin–Huxley dynamics for taste receptor cells with Izhikevich spiking neurons for large-network simulations, and includes modality-specific receptor dynamics for T1R/T2R, ENaC, and proton channels, GHK-driven ion currents, glutamate release kinetics with alpha-function profiles, AMPA receptor trafficking regulated by phosphorylation, and STDP-like plasticity. At the network level, the framework optimizes temporal spike synchrony and combinatorial population coding, treating taste qualities, mixtures, and hedonic value as structured spike patterns rather than as single labeled-line outputs (Lazovsky et al., 16 Sep 2025).

These results collectively oppose the view that taste assessment must remain purely subjective. The neural studies do not eliminate subjectivity in food evaluation, but they show that controlled taste stimulation produces reproducible neural signatures that can be modeled, visualized, and compressed for downstream computation (Xia et al., 2023).

3. Molecular, formulation, and fabrication approaches to taste design

TastePepAI frames taste as a de novo design target for peptides. Its core generator is a loss-supervised adaptive variational autoencoder trained on 1131 non-redundant taste peptides, with a taste-avoidance mechanism that constructs positive and negative latent spaces and filters candidates by $k$-nearest-neighbor distances. The framework integrates SpepToxPred, an ensemble toxicity predictor for short peptides, and uses physicochemical characterization and experimental validation. In the reported case study, the full pipeline identified 73 novel peptides with sweet, salty, and umami tastes, intended bitterness suppression, purity $>95$–98%, cell viability $>90\%$ at 100 $\mu$M in four human cell lines, and hemolysis $<1.5\%$ at 100 $\mu$M in mouse RBCs (Yue et al., 13 Feb 2025).

Ingredient-level food modeling presents a different formalization. “Predicting food taste with bound-driven optimization” models recipes as composite materials and applies Reuss–Voigt and Hashin–Shtrikman bounds to sweetness, sourness, bitterness, umami, and saltiness on a curated dataset of 70 recipes decomposed into 209 ingredient-level taste references with trained-panel ground truth. The additive bounds systematically under-predict perceived taste: 77% of actual taste values exceeded the HS upper bound, ranging from 26% for bitterness to 97% for saltiness. A hybrid model that augments the HS baseline with eight chemistry-proxy features for Maillard reactions, caramelization, evaporative concentration, protein hydrolysis, and nucleotide synergy reduces mean absolute error by 27–62% for sweetness, sourness, umami, and saltiness while using only 10 interpretable features, and constrained inverse design is demonstrated with Differential Evolution (Tagkopoulos et al., 22 Apr 2026).

Taste can also be engineered spatially rather than only compositionally. TastePrint decouples geometry from taste in 3D food printing by combining extrusion of a single base material with layer-wise airbrushed liquid seasonings. The system integrates a PyQt6 GUI for model slicing and spray specification with a modified Ender 3 food printer equipped with six airbrush channels, five of which were used for sweet, salty, sour, bitter, and umami solutions. Its spray-resolution model achieved $R^2 = 0.86$, the spray-amount model achieved $R^2 = 0.99$, and participants in an exploratory usability study completed taste-pattern design in approximately 15 min on average (Miyatake et al., 24 Mar 2026).

Taken together, these works show that taste design can target molecules, formulations, and spatial distributions. A common implication is that taste is not treated as a fixed byproduct of composition alone, but as an engineerable output whose controllability depends on representation: latent sequence space for peptides, interpretable chemistry proxies for recipes, and layer-wise dose fields for printed foods (Tagkopoulos et al., 22 Apr 2026).

4. Taste in language, multimodality, and human-centered representation

Recipe NLP operationalizes taste as explicit language. TASTEset provides 700 recipes, 3,788 manually annotated ingredient lines, and 13,362 total entities across nine classes: FOOD, QUANTITY, UNIT, PROCESS, PHYSICAL QUALITY, COLOR, TASTE, PURPOSE, and PART. TASTE accounts for 126 occurrences, 31 unique surface forms, 0.94% of all entities, and an average of 0.18 instances per recipe. In baseline evaluations, BERT+CRF achieved the best TASTE F1 at 0.789, slightly above BERT-large at 0.781 and substantially above LUKE at 0.655, indicating that taste descriptors are comparatively sparse and difficult NER targets (Wróblewska et al., 2022).

Image-based modeling extends taste beyond text. FoodSense introduces 66,842 participant-image pairs across 2,987 food images, with numeric ratings and free-text descriptors for taste, smell, texture, and sound. For taste specifically, the released dataset reports 63,978 ratings with mean 3.89 and standard deviation 1.08 on the rescaled 1–5 scale, plus 9,217 unique taste terms. FoodSense-VL is trained to predict ratings and grounded explanations from images; for taste it achieves MAE 0.515, RMSE 0.645, Pearson $r = 0.398$, Spearman $k$0, CCC 0.381, and 3-class ordinal accuracy 0.705, while using a two-stage training procedure to avoid scalar-rating collapse during rationale generation (Ishraq et al., 15 Apr 2026).

Crossmodal work further expands the concept. A fine-tuned MusicGen model conditioned on taste descriptions generated 15-second pieces intended to sound sweet, bitter, sour, or salty. In an online study with $k$1, the fine-tuned model was significantly preferred over the base model for sweet, sour, and bitter prompts, whereas salty showed no significant advantage. Factor analysis of listener ratings linked sweetness to positive valence and warmth, bitterness and sourness to negative affective descriptors, and saltiness to a more complex factor mixing surprise and happiness (Spanio et al., 4 Mar 2025).

The human-centered AI framework for aesthetic taste demonstrates a semantic extension of the term. Using 1000 wheel images, 80 representative stimuli, 575,520 pairwise judgments from 2,398 participants, designer-informed annotations, computer-vision features, and semantic alignment between participant text and GPT-5.2 captions, the framework models style-specific Bradley–Terry scores for labels such as sporty and luxury. It argues explicitly that preference data alone provides limited guidance for concrete design decisions because it does not encode which visual attributes, semantic cues, or design patterns drive those judgments (Hong et al., 23 Jan 2026).

A recurring misconception across these literatures is that preference scores or single labels are sufficient. The supplied work repeatedly rejects that simplification by introducing richer structures: explicit TASTE entities in recipes, multisensory rating–descriptor–rationale tuples, taste-to-sound mappings mediated by affect, and feature-linked stylistic interpretations (Hong et al., 23 Jan 2026).

5. TASTE as an acronymic family of technical systems

A large portion of the recent literature uses TASTE as an acronym rather than as a sensory noun. The acronym is domain-specific and should not be conflated across fields.

TASTE expansion	Domain	Defining contribution
The Asiago Survey for Timing transit variations of Exoplanets	Exoplanet astronomy	Ground-based TTV/TDV survey at the Asiago 1.82 m telescope
The ASSERT Set of Tools for Engineering	Space-systems MBSE	ESA-supported toolset using ASN.1, AADL, SDL, with IF-based formal verification
Text mAtching based SequenTial rEcommendation	Recommender systems	T5-based dual-encoder sequential recommendation via text matching
Textual And STructural Embeddings	Stance detection	GRN fusion of Sentence-BERT content and SDP-derived structural embeddings
Text-Aligned Speech Tokenization and Embedding	Spoken language modeling	Reconstruction-trained speech tokenizer aligned to text tokens
Task Synthesis from Tool Sequence Evolution	Agent benchmarking	Automatic generation of valid, high-coverage, difficult tool-use benchmarks
Tree-LSTM Aspect Sentiment Triplet Extraction	Sentiment analysis	Hybrid neural-symbolic extraction of target–sentiment–cause triplets

In astronomy, TASTE is a focused ground-based program for high-precision, short-cadence transit photometry optimized for TTV and TDV studies. Its first report uses the Asiago 1.82 m telescope and AFOSC to obtain timing accuracies of $k$2 s for HAT-P-3b and $k$3 s for HAT-P-14b, refining ephemerides with a linear relation $k$4 and demonstrating competitiveness with other ground-based efforts (Nascimbeni et al., 2010).

In space-systems engineering, TASTE is an ESA-supported MBSE toolset built around ASN.1 for data modeling, customized AADL for architecture, and SDL with Ada/C/C++ integration for behavior and implementation. The formal-verification extension described in the MoC4Space project translates TASTE models and properties into IF observers and timed-automata-style verification artifacts, enabling checks of Boolean stop conditions, MSC properties, and SDL observers inside the TASTE workflow (Dragomir et al., 2021).

In information retrieval, TASTE reframes sequential recommendation as text matching between verbalized user histories and item descriptions using a T5-based dual encoder. It addresses popularity bias and long-tail cold start, and on Amazon Beauty, Sports, Toys, and Yelp it outperforms DIF-SR with improvements up to about 30% in Recall/NDCG while reducing the fraction of popular items among top-5 recommendations (Liu et al., 2023). In stance detection, TASTE combines Sentence-BERT with SDP-based structural speaker embeddings through a Gated Residual Network and attains the best or tied-best performance on most 4Forums and CreateDebate settings, particularly where conversational structure carries high signal-to-noise ratio (Barel et al., 2024).

In spoken language modeling, TASTE denotes a text-aligned speech tokenizer trained with a reconstruction objective. It directly aligns speech units to ASR text tokens via attention-based aggregation over shallow and last-layer encoder states, preserves paralinguistic information, and operates at about 190 bps. The resulting TASTE-based spoken LLMs are reported as comparable on SALMON and StoryCloze while significantly outperforming other pre-trained SLMs on speech continuation (Tseng et al., 9 Apr 2025). In benchmarking, TASTE is “Task Synthesis from Tool Sequence Evolution,” using an Adaptive Contrastive $k$5-gram model guided by LLM plausibility judgments. Uniform tool sampling produced 6.7% valid sequences, a non-adaptive model 16.7%, and the full method 86.7%; the resulting $k$6-Bench more than doubles the number of unique tool combinations and causes severe performance drops for models that nearly saturate $k$7-Bench (Keren et al., 27 May 2026). In sentiment analysis, TASTE is a hybrid neural-symbolic framework for Tree-LSTM aspect sentiment triplet extraction that does not require triplet-level training data, using dependency-tree sentiment composition plus symbolic rules to extract target–sentiment–cause triplets (Sutherland et al., 2021).

6. Recurring methodological themes and open problems

Several methodological themes recur across these otherwise disparate meanings of TASTE. One is alignment: text-aligned speech tokens in spoken language modeling, textual–structural fusion in stance detection, semantic alignment between consumer descriptions and image captions in aesthetic taste modeling, and named-entity extraction of flavor descriptors in recipes all treat performance as depending on how well symbolic structure is aligned with latent or sensory data (Tseng et al., 9 Apr 2025).

A second theme is contrastive or exclusion-based control. TastePepAI uses positive and negative taste manifolds and a taste-avoidance mechanism to design sweet, salty, and umami peptides while excluding bitterness (Yue et al., 13 Feb 2025). TASTE for agent benchmarks similarly relies on positive and negative $k$8-gram statistics derived from plausible and implausible tool sequences (Keren et al., 27 May 2026). This suggests a broader design pattern in which undesirable functions are represented explicitly rather than only as absence.

A third theme is interpretability under complexity. The hybrid recipe-taste model combines HS baselines with eight chemistry-proxy features rather than relying solely on a 115-feature black-box Lasso (Tagkopoulos et al., 22 Apr 2026). The aesthetic framework insists on designer-informed features and consumer language rather than opaque reward-only models (Hong et al., 23 Jan 2026). The ESA TASTE verification work inserts formal properties, observers, and MSC diagnostics into an automated MBSE chain (Dragomir et al., 2021). The ASTE-oriented TASTE likewise couples Tree-LSTM sentiment parsing with symbolic rules to reduce dependence on expensive joint triplet annotation (Sutherland et al., 2021).

The major limitations are equally recurrent. Taste labels remain subjective and protocol-dependent in peptide datasets, multisensory image annotation, and aesthetic judgments (Yue et al., 13 Feb 2025). Objective neural decoding has so far been demonstrated in small, controlled laboratory settings rather than in ecologically rich eating scenarios (Xia et al., 2023). Additive ingredient models fail when processing chemistry or perceptual synergy is ignored (Tagkopoulos et al., 22 Apr 2026). Benchmark saturation can mask weak generalization, as shown by the sharp decline from $k$9-Bench to $>95$0-Bench (Keren et al., 27 May 2026). In acronymic uses, a persistent source of confusion is simple homonymy: TASTE may refer to gustation, a telescope survey, an MBSE toolset, a recommender, a stance detector, a speech tokenizer, a benchmarking pipeline, or a sentiment extractor, and the technical meaning is entirely domain-dependent.

The supplied literature therefore presents TASTE not as a single stable concept but as a family of formalized objects: sensory modality, linguistic attribute, design target, benchmark-construction strategy, and acronymic system name. What unifies these usages is not a shared ontology, but a repeated attempt to transform a vague, high-level notion—flavor, style, speech alignment, tool-use coverage, or correctness—into structured representations that can be measured, optimized, and validated.