Seahorse: Cross-Disciplinary Insights

• SEAHORSE is a multifaceted designation spanning NLP datasets, astronomical phenomena, STPP benchmarks, and symbolic artifacts in HCI, illustrating cross-disciplinary naming practices.
• The multilingual summarization dataset and unified STPP framework provide rigorous benchmarks, enhancing evaluation metrics and promoting cross-domain transfer in machine learning.
• Astronomical uses of SEAHORSE name objects by morphology while its symbolic representation in therapy underscores its diverse practical applications across research areas.

SEAHORSE is a designation that appears across several unrelated research literatures. It names a multilingual, multifaceted dataset for summarization evaluation; a unified benchmarking framework for spatiotemporal point processes; several astronomical objects whose projected morphology resembles a seahorse; and, in HCI studies, a symbolic or animated marine character used in therapeutic and environmental storytelling (Clark et al., 2023, Miettinen, 2017, Aalaila et al., 1 Jul 2026).

1. Cross-disciplinary uses of the designation

Domain	Entity	Description
Natural language processing	SEAHORSE	Multilingual, multifaceted summarization-evaluation dataset with 96,645 human-rated summaries (Clark et al., 2023)
Galactic star formation	Seahorse Nebula / G304.74+01.32	Filamentary infrared dark cloud with seahorse-like morphology in Herschel images (Miettinen, 2017)
Radio astronomy	Seahorse radio relic	Brightest radio relic in PSZ2 G200.95-28.16, named for its bent, elongated morphology (Pal et al., 2024)
Extragalactic astronomy	Cosmic Seahorse	Strongly lensed submillimeter galaxy WISE J122651.0+214958.8 (Sulzenauer et al., 2021)
H II region studies	RN A “Seahorse”	Rosette Nebula globule hosting a Class I YSO and outflow activity (Mäkelä et al., 2017)
Spatiotemporal machine learning	SEAHORSE	Unified benchmarking framework for STPPs with raw-coordinate likelihood reporting (Aalaila et al., 1 Jul 2026)
HCI and therapy	“Seahorse” family symbol	Therapeutic image representing the mother in AI-infused family storymaking (Liu et al., 2024)

In astronomy, the designation usually follows projected morphology or an established nickname. In NLP and STPP research, it functions as the formal name of a dataset or framework. In HCI, it operates as a symbolic or embodied character. This distribution of usages suggests that the term is less a single technical concept than a recurrent naming convention spanning benchmark design, morphology-based astronomical nomenclature, and symbolic interaction design.

2. SEAHORSE in multilingual summarization evaluation

SEAHORSE was introduced as a dataset for multilingual, multifaceted summarization evaluation in response to two specific problems: the scarcity of human evaluation data beyond English and the methodological weakness of training evaluation models on data that overlap with benchmark test sets. The dataset contains 96,645 summaries with human ratings along six dimensions of text quality—comprehensibility, repetition, grammar, attribution, main ideas, and conciseness—covering 6 languages, 9 systems, and 4 datasets. The languages are German, English, Spanish, Russian, Turkish, and Vietnamese; the datasets are XSum, XL-Sum, MLSum, and WikiLingua; and the systems include human reference summaries together with T5-base, T5-base-250, T5-xxl, mT5-small, mT5-small-250, mT5-xxl, PaLM 1-shot, and PaLM finetuned (Clark et al., 2023).

The annotation protocol is explicitly reference-free. Rather than comparing a candidate summary with a reference summary, raters judge the summary directly using six questions, Q1–Q6. Q1 asks whether the summary can be read and understood; if the answer is “No,” the remaining questions are skipped. Q2 evaluates whether the summary is free of unnecessarily repeated information; Q3 grammatical correctness; Q4 whether all information is fully attributable to the source article; Q5 whether the summary captures the main idea(s) of the source; and Q6 whether the source information is represented concisely. This design reflects the paper’s central premise that summarization quality is not one-dimensional: a system output can be fluent but repetitive, grammatical but unsupported, or faithful but incomplete.

SEAHORSE serves both as a benchmark and as training data for learned metrics. The paper fine-tunes a separate mT5-based classifier/regressor for each quality dimension. The model is trained as a text-to-text system that takes the article and summary as input, separated by "premise:" and "hypothesis:" tags, and outputs either '0' or '1'. Training uses a filtered subset with usable binary labels, after removing duplicates and non-Yes/No ratings, leaving 88,280 items for metric training and testing. This metric family is referred to as mt5SEAHORSE.

Evaluation on the benchmark side uses Pearson correlation and area under the ROC curve (AUC), explicitly because both are threshold-free and consistent with prior meta-evaluation practice. The learned SEAHORSE metrics outperform majority class and ROUGE-L on the SEAHORSE test set across all dimensions. Repetition is the easiest dimension to predict, and it also has the highest annotator agreement. The paper further reports 8,920 duplicated annotations and an overall pairwise agreement of 82%, while noting that agreement is easier on comprehensibility, repetition, and grammar than on attribution, main ideas, and conciseness.

A principal empirical claim is cross-domain transfer. SEAHORSE-trained metrics generalize zero-shot to mFACE, where Q1 is mapped to quality, Q4 to attribution, and Q6 to informativeness, and they outperform other non-mFACE-trained methods in almost all settings, including languages not seen during SEAHORSE training. The attribution metric also transfers to TRUE, especially on summarization datasets, and remains competitive on dialogue datasets such as BEGIN, Q2, and DialFact. In that sense, SEAHORSE functions not only as a benchmark for summarization evaluation but also as a large-scale supervision source for multilingual, reference-free evaluation models.

3. The Seahorse Nebula and the Seahorse infrared dark cloud

In Galactic star-formation studies, the “Seahorse Nebula” denotes the filamentary infrared dark cloud G304.74+01.32. The nickname derives from Herschel far-infrared/submillimetre images, in which the cloud appears as a long, curved filament with a prominent head-like brightening and a sinuous body, accompanied by elongated, perpendicular striations. SABOCA 350 $\mu$m mapping at 9″ resolution shows that the cloud contains a dense internal filament with mean width $0.18\pm0.05$ pc. On larger scales, LABOCA 870 $\mu$m imaging yields a projected length of $8.9\pm2.3$ pc, mean effective width $0.35\pm0.09$ pc, total mass $773\pm131\ M_\odot$, and line mass $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$. The paper reports that $36\%\pm16\%$ of LABOCA clumps are fragmented into SABOCA cores, that the WISE data suggest a YSO host fraction of $65\%\pm18\%$, and that the filament is thermally supercritical by a factor of $\gtrsim3.5$ on the LABOCA scale and $0.18\pm0.05$0 on the SABOCA scale (Miettinen, 2017).

The same cloud was later studied chemically with an APEX survey near 170 GHz toward 14 positions along the full filament. Six transitions were detected at $0.18\pm0.05$1: SO$0.18\pm0.05$2, H$0.18\pm0.05$3CN$0.18\pm0.05$4, H$0.18\pm0.05$5CO$0.18\pm0.05$6, SiO$0.18\pm0.05$7, HN$0.18\pm0.05$8C$0.18\pm0.05$9, and C$\mu$0H$\mu$1. SO, H$\mu$2CO$\mu$3, and HN$\mu$4C were detected in every source; C$\mu$5H and H$\mu$6CN had detection rates of 92.9% and 85.7%; and SiO was detected only in SMM 3, with a 7.1% detection rate. Three clumps—SMM 5, SMM 6, and SMM 7—showed SO, H$\mu$7CN, H$\mu$8CO$\mu$9, HN$8.9\pm2.3$0C, and C$8.9\pm2.3$1H in absorption. The most significant evolutionary trends are the decline in C$8.9\pm2.3$2H abundance and the decline in the $8.9\pm2.3$3 ratio as clumps evolve from IR dark to IR bright and then to an H II region. The strongest abundance correlations are between C$8.9\pm2.3$4H and HN$8.9\pm2.3$5C, with $8.9\pm2.3$6, and between HN$8.9\pm2.3$7C and H$8.9\pm2.3$8CO$8.9\pm2.3$9, with $0.35\pm0.09$0 (Miettinen, 2020).

A further study constructed far-IR to submillimetre SEDs for 12 dense cores in the cloud using WISE, IRAS, Herschel, SABOCA, and LABOCA data. Modified blackbody fitting gave mean cold and warm dust temperatures of $0.35\pm0.09$1 K and $0.35\pm0.09$2 K, mean mass $0.35\pm0.09$3, mean luminosity $0.35\pm0.09$4, mean H$0.35\pm0.09$5 number density $0.35\pm0.09$6, and mean surface density $0.35\pm0.09$7. All 12 cores have virial parameter $0.35\pm0.09$8, hence are gravitationally bound. Seven of 12 cores (58%) lie above both the Kauffmann & Pillai and Baldeschi mass-radius thresholds and also satisfy $0.35\pm0.09$9; five of 12 cores (42%) are fragmented into two components with projected separations of about 0.09–0.21 pc. The fragmentation analysis is heterogeneous: SMM 6a is consistent with thermal Jeans fragmentation, while the other fragmented cores require either underestimated cold dust temperatures or an effective sound speed including non-thermal motions (Miettinen, 2020).

Taken together, these studies present the Seahorse as a nearby IRDC with hierarchical fragmentation, internal chemical evolution, multiple evolutionary stages from IR-dark clumps to an H II region, and several dense cores that appear capable of high-mass star formation.

4. Other astronomical objects named Seahorse

The name also denotes the brightest radio relic in the low-mass merging cluster PSZ2 G200.95-28.16. In this context the relic, labeled R1 in the paper, is renamed “Seahorse” because its bent, elongated morphology resembles a seahorse, with a bent head to the east, a notch in the body, an extended tail, and two bright filaments separated by the notch. Observations with the upgraded GMRT, MeerKAT, and Chandra characterize it as a polarized diffuse synchrotron source in the cluster outskirts with largest linear size 1.53 Mpc, largest width 237 kpc, average linear polarization fraction $773\pm131\ M_\odot$0 at 1283 MHz, integrated spectral index $773\pm131\ M_\odot$1, and 1.4 GHz radio power $773\pm131\ M_\odot$2. The injection spectral index is identified as $773\pm131\ M_\odot$3, implying a DSA shock Mach number of $773\pm131\ M_\odot$4. The relic shows spectral steepening toward the cluster centre, smoothly aligned B-vectors along both filaments, and depolarization across the shock width. At the same time, no X-ray shock is detected at the relic location, and the relic is an outlier in the radio relic power–cluster mass relation (Pal et al., 2024).

In extragalactic astronomy, WISE J122651.0+214958.8 is dubbed the “Cosmic Seahorse.” It is a strongly lensed submillimeter galaxy at $773\pm131\ M_\odot$5, magnified by a foreground galaxy cluster at $773\pm131\ M_\odot$6 with $773\pm131\ M_\odot$7. The main arc spans about 8.2 arcsec. From MAGPHYS fitting, the paper derives an apparent infrared luminosity $773\pm131\ M_\odot$8 and an apparent IR-based star-formation rate $773\pm131\ M_\odot$9. The molecular-gas result that motivates the source’s significance is the exceptionally low line-luminosity ratio $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$0, interpreted as evidence for Milky Way-like low-excitation molecular gas. The authors conclude that this is best explained by an extended, main-sequence star-formation mode rather than a compact merger-driven starburst, and they describe the object as a possible missing link between starbursts and lower-luminosity systems during cosmic noon (Sulzenauer et al., 2021).

A smaller-scale morphological usage appears in the Rosette Nebula, where globule RN A is nicknamed the “Seahorse.” The globule lies on the northwestern molecular shell and is illuminated and eroded by the UV field of NGC 2244 while also hosting embedded star formation. Its morphology includes a tail extending southwest, a shorter head toward the northeast, a bar-like feature in the belly, and a nearly north-south bright ridge ending in a bowshock cataloged as MHO 3142. The bright rim emission is largely fluorescent H$87\pm27\ M_\odot\,\mathrm{pc}^{-1}$1 excited by the Rosette UV field, with about 20–40% of the Ks-band flux in the rims contributed by the H$87\pm27\ M_\odot\,\mathrm{pc}^{-1}$2 1–0 S(1) line at 2.12 $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$3m. The strongest evidence for active star formation is RN A IRS 1, a Class I YSO embedded deep inside the globule and identified as the outflow-driving source. CO observations show blueshifted line wings, the outflow cavity wall is bright in the 2.124 $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$4m H$87\pm27\ M_\odot\,\mathrm{pc}^{-1}$5 line and at 8.0 $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$6m, and H$87\pm27\ M_\odot\,\mathrm{pc}^{-1}$7 structures suggest a possible parsec-scale outflow of about 1 pc (Mäkelä et al., 2017).

5. Benchmarking frameworks and seahorse-inspired optimization

A distinct computational usage appears in the 2026 STPP literature, where SEAHORSE is introduced as a unified benchmarking framework for spatiotemporal point processes. Its central abstraction is a common encode–evolve–decode interface: the history encoder compresses the event history $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$8, the state evolution propagates that representation through continuous time, and the decoder exposes the predictive law for the next event. The framework standardizes dataset handling, model construction, training modes, hyperparameter tuning, and evaluation artifacts under a single executable benchmarking contract. A central fairness mechanism is raw-coordinate likelihood reporting: even if a model trains in normalized coordinates, benchmark-facing NLL values are corrected back to raw coordinates through the log-Jacobian relation. The default experimental policy uses fixed 70/10/20 train/validation/test splits averaged over three seeds. SEAHORSE unifies classical factorized models, temporal neural baselines with spatial heads, neural STPP likelihood models, and sample-based generative models, and it exposes execution modes fit, tune, bench, and evaluate. It is paired with HawkesNest, a synthetic stress-test suite based on a multivariate Hawkes process backbone, and the paper’s main diagnostic result is that increasing structural complexity reveals model-family inductive biases rather than yielding a universally best method (Aalaila et al., 1 Jul 2026).

A related but separate line of work uses the seahorse motif in metaheuristic optimization. The Sea Horse Optimizer (SHO) is a swarm-based algorithm inspired by spiral locomotion, Brownian-like motion, predation, and breeding behavior in seahorses. Its movement phase employs a logarithmic helical equation together with Levy flight and Brownian motion. The paper “A new approach for solving global optimization and engineering problems based on modified Sea Horse Optimizer” proposes mSHO, which replaces the original exploitation mechanism with a three-step local-search scheme: neighborhood-based local search, global non-neighborhood-based search, and wandering around the existing search region, also described as circumnavigation of the current search region. The reported results assign mSHO total rank 1 on the CEC’2020 test suite and best objective values for nine engineering design problems, including pressure vessel design, speed reducer design, tension/compression spring, welded beam design, three-bar truss engineering design, industrial refrigeration system, multi-product batch plant, cantilever beam problem, and multiple disc clutch brake problems (Hashim et al., 2024).

6. Symbolic, therapeutic, and behavioral uses in HCI

In family expressive arts therapy, the seahorse appears not as a formal model name but as a therapeutic image. The title phrase “When He Feels Cold, He Goes to the Seahorse” comes from a child’s symbolic family story in which the purple fish represented the child, the seahorse represented the mother, and the jellyfish represented the father. In the therapist’s reading, the phrase signified that when the child “feels cold,” he seeks comfort and attachment from the mother, while the father is more distant. The study ran over five weeks with five one-hour storymaking sessions involving seven families and 18 participants—8 children, 7 mothers, and 3 fathers—guided by a professional expressive arts therapist. The material setup combined colored pencils, markers, paint sticks, clay, non-woven fabric, yarn, drawing paper, LEGO, personal toys, and Midjourney. The workflow repeatedly moved through physical making, photographing and prompting, AI generation, printing and materializing, and further physical making. The analysis used the Expressive Therapies Continuum, presented as $87\pm27\ M_\odot\,\mathrm{pc}^{-1}$9, and the seahorse example is located primarily at the Cognitive/Symbolic level as a family self-symbol that reveals attachment, comfort, and disconnection (Liu et al., 2024).

In OceanChat, the seahorse is one of three animated marine characters—Beluga whale, Moon jellyfish, and Three-Spot seahorse—used in a $36\%\pm16\%$0 between-subjects experiment with $36\%\pm16\%$1. The character appears both in a Static Character Narrative and in a Conversational Character Narrative powered by OpenAI’s gpt-4o-mini, with text-to-speech audio, concise 2–3 sentence responses, reflective questions, and a final urgent and hopeful call to action after three user turns. Quantitatively, relative to the jellyfish baseline, the seahorse had a significant negative effect on perceived anthropomorphism ($36\%\pm16\%$2, $36\%\pm16\%$3), perceived animacy ($36\%\pm16\%$4, $36\%\pm16\%$5), and climate change impact awareness ($36\%\pm16\%$6, $36\%\pm16\%$7). Effects on likeability and safety were only marginal, and most downstream outcomes—including climate change belief, psychological distance, climate policy adoption, self-reported pro-environmental behavior, sustainable choice preferences, and willingness to share climate change information—were not significant. The paper therefore treats the seahorse as a boundary case showing that species selection matters, and it uses the result to motivate “graduated anthropomorphism,” a balance between species authenticity and human-like relational cues (Pataranutaporn et al., 5 Feb 2025).

Across these literatures, SEAHORSE functions as a benchmark name, an astronomical nickname, a morphology-based label, and a symbolic or embodied character. The commonality is not a shared technical ontology but a recurring research practice: the same term is repurposed to identify complex structures—datasets, filaments, relics, galaxies, globules, frameworks, and characters—whose scientific importance lies in their internal organization rather than in the name itself.