Mantis: Biology, Astronomy, AI & More

Updated 5 July 2026

Mantis is a term that encompasses both the biological order Mantodea and a codename for diverse scientific projects across astronomy, AI, robotics, and cybersecurity.
Biological studies on mantids reveal key insights into genome evolution, rapid species radiation, and agricultural importance through chromosome-scale analyses.
Innovative applications include state-of-the-art observatory systems, multimodal machine learning models, and energy-efficient robotics, achieving impressive performance metrics.

Mantis denotes both praying mantises in biology and a recurrent acronym or codename in contemporary scientific research. In the biological literature it refers to members of the order Mantodea, including Mantis religiosa; in technical literature it names systems for wide-field radio astronomy, photometric-redshift estimation, conversational search, multimodal language generation, multi-image instruction tuning, time-series classification, disease forecasting, robotics, cybersecurity, medical imaging, and mixed-signal sensing hardware (Liu et al., 29 Apr 2025, Cappellen et al., 2016).

1. Biological referent: Mantodea

Praying mantises are predatory insects of the order Mantodea. The genomic study of five species—Mantis religiosa, Tenodera sinensis, Deroplatys truncata, Hymenopus coronatus, and Metallyticus violaceus—places them in a lineage of roughly ~2500 species with major roles in agriculture, traditional medicine, bionics, and entertainment (Liu et al., 29 Apr 2025). That study reports chromosome-scale reference genomes and identifies transposable element expansion as the major force governing genome size in Mantodea. It also infers that the Mantodea ancestor may have had only one X chromosome, with later translocations between the X chromosome and an autosome in the lineage of Mantoidea, and it finds a lower evolutionary rate for the metallic mantis than for the other sampled mantises (Liu et al., 29 Apr 2025).

The same work argues that Mantodea underwent rapid radiation after the K–Pg mass extinction event, which may have contributed to confusion in species classification (Liu et al., 29 Apr 2025). In this biological sense, “mantis” names a zoological lineage that is both evolutionarily distinctive and methodologically useful for chromosome evolution, repeat-driven genome expansion, and rapid diversification studies.

2. Astronomy and astrophysical usage

In radio astronomy, MANTIS stands for Mid-Frequency Aperture Array Transient and Intensity-Mapping System. It is proposed as a wide-field mid-frequency aperture-array science demonstrator and precursor to an MFAA-based SKA telescope on the South African SKA site. The preliminary specifications given are a 450–1450 MHz frequency range, $>500\ \mathrm{MHz}$ bandwidth, 1500–2500 m $^2$ collecting area, A/T = 38–63 m $^2$ /K, SEFD = 74–44 Jy, 200 deg $^2$ at 1 GHz field of view, and transient buffering (Cappellen et al., 2016). Its main science drivers are transients, HI intensity mapping, pulsars, and SETI, and it is explicitly framed as the first MFAA station on the SKA site and as a reference point for MFAA science, costing, and performance toward SKA Phase 2 (Cappellen et al., 2016).

In astronomical machine learning, “Mantis Shrimp” names a multi-survey computer-vision model for photometric redshift estimation that fuses GALEX, Pan-STARRS, and UnWISE imagery. The model uses a ResNet50 modified for 9-channel input, recasts redshift estimation as classification over $C=200$ bins between 0.0 and 1.0, and is trained on a spectroscopic compilation of roughly $4.2\times10^6$ galaxies, with preliminary experiments using about 7% of the total (Engel et al., 2024). On the held-out test set reported in the paper, the multi-instrument CNN achieves MAD $=0.0244$ , Bias $=1e{-2}$ , and $\eta=13.84\%$ , with ablations showing that IR helps substantially while UV contributes little in the reported setup (Engel et al., 2024). The broader significance is that “Mantis” in astronomy names both a future wide-field radio instrument and a multisurvey image-fusion model, but in two entirely unrelated subfields.

3. Language, multimodal, and information-seeking systems

In NLP and multimodal learning, the label has been reused for several unrelated resources and methods.

System	Domain	Salient properties
MANtIS (Penha et al., 2019)	Conversational search	80,324 multi-turn, multi-domain, grounded Stack Exchange dialogues; 118,349 response-ranking contexts; 1,356 conversations with 6,701 intent-labeled utterances
MAnTiS (Sollami et al., 2021)	Multimodal NLG	GPT-2 medium with ResNet-152 image conditioning via a multimodal prefix; product description generation from images and titles
MANTIS (Jiang et al., 2024)	Multi-image LMMs	721K multi-image instruction examples plus 268K single-image reasoning examples; Mantis-SigLIP averages 62.6 on five multi-image benchmarks
MANTIS (Zanwar et al., 2023)	Mental-health text classification	Hybrid and stacked ensembles using Mental RoBERTa, PsychBERT, ClinicalBERT, and a BiLSTM over 168 engineered features; best 89.31 validation F1 and 83.76 test F1
MANTIS (Li et al., 2022)	Lexical simplification	Unsupervised extension of LSBert using RoBERTa-medium, 30 candidates, weighted ranking, and textual-entailment filtering; 5.9% accuracy improvement over LSBert

The MANtIS conversational-search dataset is defined by two states—information-need elucidation and information presentation—and was introduced as a large-scale benchmark for mixed-initiative, grounded information-seeking dialogue (Penha et al., 2019). MAnTiS for natural language generation, by contrast, is a prefix-based multimodal conditioning method in which modality-specific encoders project into the LLM embedding space before generation; on fashion description generation it outperforms Context-Attn and Pseudo-Self on BLEU4, CIDEr, METEOR, and ROUGE-L (Sollami et al., 2021). The 2024 MANTIS work on interleaved multi-image instruction tuning argues that multi-image ability need not be acquired through massive noisy interleaved pre-training, and reports that Mantis-SigLIP exceeds Idefics2 by about 9 absolute points on the paper’s five multi-image benchmarks while remaining competitive on single-image evaluation (Jiang et al., 2024).

4. Temporal foundation models

In time-series learning, Mantis is a classification-oriented foundation model rather than a forecasting-oriented one. The original model is a lightweight ViT-based encoder pretrained with a contrastive objective on 7 million time series samples. It uses a fixed input length of 512, produces 32 tokens of dimension 256, has about 8 million parameters, and is evaluated both in frozen-feature and full fine-tuning regimes. The paper emphasizes that Mantis achieves the lowest calibration error on average over 131 datasets among the compared TSFMs, while also supporting multivariate adapters for channel compression and interaction modeling (Feofanov et al., 21 Feb 2025).

MantisV2 is presented as a direct successor aimed at closing the frozen-versus-fine-tuned gap in zero-shot time-series classification. The work introduces Mantis+, pretrained entirely on 2 million synthetic time series, and then refines the architecture into MantisV2, which has 4.2M parameters originally and 2.2M after pruning. It combines synthetic-only pretraining, intermediate-layer extraction, revised token aggregation, self-ensembling over interpolation scales, and cross-model embedding fusion (Feofanov et al., 19 Feb 2026). On UCR, the paper reports 0.8195 average accuracy for MantisV2 with Random Forest and 0.8360 with logistic regression; the strongest fusion, MantisV2 + TiConvNext, reaches 0.8494, effectively matching fine-tuned MantisV2 at 0.8500 (Feofanov et al., 19 Feb 2026).

A different temporal use appears in infectious-disease forecasting. There, Mantis is a simulation-grounded foundation model trained entirely on over 400 million simulated days of outbreak dynamics and deployed in strict zero-shot mode on real epidemiological tasks (Dudley et al., 17 Aug 2025). The model is designed to generalize across diseases, regions, and outcomes, including cases, hospitalizations, deaths, and an unseen modality at inference time, ILI (Dudley et al., 17 Aug 2025). The paper states that Mantis outperformed 39 comparison models across six diseases, including the CDC COVID-19 Forecast Hub models tested, while also providing calibrated uncertainty, mechanistic interpretability through back-to-simulation attribution, and actionable 8-week forecasts (Dudley et al., 17 Aug 2025).

5. Robotics and embodied intelligence

In biomimetic underwater robotics, mantis-shrimp inspiration appears through latch-mediated spring actuation. The reported system uses a soft bistable actuator with an integrated latch mechanism to reproduce LaMSA-style slow energy loading and fast release with a single motor (Bi et al., 26 May 2026). The robot has mass $350\,\mathrm{g}$ , a 167 RPM, 4 W DC motor, and four fins. Experiments report stable periodic flapping, maximum thrust $^2$ 0, impulse $^2$ 1, and 30 mm net vertical displacement per cycle, together with vertical ascent, diagonal forward motion, lateral translation, and turning through fin-angle modulation (Bi et al., 26 May 2026).

For autonomous mobile agents, Mantis names a methodology for deploying spiking neural networks under memory, energy, and adaptability constraints. Its three components are SNN operation optimization, an enhanced online STDP rule with adaptive learning rates, weight decay, and balanced threshold potential, and a memory- and energy-aware model-selection procedure (Putra et al., 2022). On the reported MNIST and Fashion-MNIST experiments, the methodology yields 3.32× memory reduction and 2.9× inference energy saving for an 8-bit SNN relative to a 32-bit baseline, while also reducing training energy by up to 2.7× (Putra et al., 2022).

In robot learning, Mantis is also a Vision-Language-Action model with Disentangled Visual Foresight. It combines a Qwen2.5-VL backbone, trainable latent-action and action queries, a connector, a Sana-based DiT head for future-frame prediction, and a diffusion action head (Yang et al., 20 Nov 2025). The total model size is 5.8B parameters, and the main empirical result is a 96.7% average success rate on LIBERO after fine-tuning, with further real-world gains over $^2$ 2 in instruction following, unseen-instruction generalization, and reasoning-oriented task variants (Yang et al., 20 Nov 2025). Across these works, “Mantis” in robotics denotes systems that use dense embodied supervision—whether biomechanical, spiking, or multimodal—to improve control.

6. Security, malicious infrastructure, and software systems

In systems research, Mantis originally denoted a framework for predicting program execution time through program instrumentation, sparse polynomial regression, and program slicing. It extracts loop counts, branch counts, method invocation counts, variable values, and exception counts, then fits a predictive model of execution time and generates compact feature evaluators through slicing (Chun et al., 2010). The reported result is more than 93% accuracy with less than 10% training data, with around 6.1% prediction error on Lucene and 5.5% on ImageJ in the representative experiments (Chun et al., 2010).

In cybersecurity defense, Mantis is a framework that counters LLM-driven cyberattacks by exploiting prompt-injection susceptibility in the attacker’s own agent (Pasquini et al., 2024). It uses decoy services plus an Injection Manager to plant hidden adversarial instructions into service responses after exploit-like interaction. The paper evaluates both passive tarpit and active counterstrike modes and reports about 95.4% defender success while reducing attacker success to under 3% across the tested agents, targets, and model backends (Pasquini et al., 2024). The active mode can induce the attacker’s LLM-agent to run a fetch-and-execute command that establishes a reverse shell back to the defender, whereas the passive mode diverts the agent into an infinite exploration loop (Pasquini et al., 2024).

In malicious-domain detection, MANTIS is a content-agnostic system that detects attacker-created domains at hosting time by monitoring low-reputation hosting infrastructure (Deniz et al., 13 Feb 2025). It builds a heterogeneous graph over apex domains, FQDNs, IPs, /24 subnets, and ASNs, and combines graph structure with lexical, hosting, and IP features (Deniz et al., 13 Feb 2025). The operational headline figures are 99.7% precision, 86.9% recall, 0.1% FPR, and about 19K new malicious domains per day, with lead times of days to weeks before popular blocklists and an average detection time of 3 days 17 hours after first PDNS appearance versus over 8 days for VT in the reported comparison (Deniz et al., 13 Feb 2025).

7. Medical imaging and near-sensor hardware

In quantitative MRI, MANTIS stands for Model-Augmented Neural neTwork with Incoherent k-space Sampling. It is a deep-learning framework for accelerated knee T2 mapping that combines direct U-Net-based map prediction with model-augmented k-space consistency (Liu et al., 2018). The method uses the physical mono-exponential decay model inside training and was evaluated at acceleration factors $^2$ 3 and $^2$ 4. At $^2$ 5, the paper reports nRMSE $^2$ 6 and SSIM $^2$ 7 for MANTIS, compared with $^2$ 8 and $^2$ 9 for CNN-Only; at $^2$ 0, the values are $^2$ 1 and $^2$ 2 versus $^2$ 3 and $^2$ 4 (Liu et al., 2018).

In hardware, MANTIS is a mixed-signal near-sensor convolutional imager SoC that integrates a $^2$ 5 3T APS array, DS3 units, analog memory, MAC units, switched-capacitor amplifiers, and SAR ADCs for feature extraction and face RoI detection (Lefebvre et al., 2024). Its distinguishing combination is $^2$ 6 4b-weighted filters, multi-scale operation, and double sampling. The paper reports peak energy efficiencies normalized to 1b operations of 4.6 TOPS/W at the accelerator level and 84.1 TOPS/W at the SoC level, with feature-map RMSE ranging from 3% to 11.3% (Lefebvre et al., 2024). In the demonstrated face RoI detector, the chip attains a false negative rate of 11.5%, discards 81.3% of image patches, and reduces off-chip transmitted data by 13× relative to the raw image (Lefebvre et al., 2024).

This suggests that “Mantis” functions less as a standardized technical term than as a recurrent project name reused across disciplines. In current arXiv usage, it denotes a zoological lineage and a heterogeneous set of instruments, datasets, models, and systems whose only commonality is nomenclature rather than method or domain.