MAVIS: A Multidisciplinary Research Overview

• MAVIS is a context-dependent acronym that spans various domains, referring to instruments, algorithms, datasets, and benchmarks with no single canonical meaning.
• In astronomy, MAVIS denotes a visible-wavelength MCAO instrument for the VLT, designed to achieve diffraction-limited imaging with advanced adaptive optics.
• Other MAVIS applications include multimodal AI, robotics SLAM, surgical scene understanding, quantum-dot control, and smartphone-based datacenter monitoring, showcasing its versatile use.

MAVIS is a context-dependent acronym used across several research communities for unrelated instruments, algorithms, datasets, and benchmarks. In the arXiv record represented here, the name most prominently denotes the MCAO-Assisted Visible Imager and Spectrograph for the VLT Adaptive Optics Facility, but it also appears in multimodal AI, microsurgical scene understanding, visual-inertial SLAM, semiconductor quantum-dot control, and smartphone-oriented datacenter monitoring (McDermid et al., 2020). The term therefore has no single canonical technical meaning outside its immediate disciplinary context.

1. Disambiguation and nomenclature

The principal uses of the acronym in the cited literature are summarized below.

Expansion	Domain	Representative paper
MCAO-Assisted Visible Imager and Spectrograph	adaptive optics / astronomy	(McDermid et al., 2020)
Multi-Camera Augmented Visual-Inertial SLAM	robotics / SLAM	(Wang et al., 2023)
Multimodal AVailability and source Attribution for long-form VISual QA	multimodal QA benchmark	(Song et al., 15 Nov 2025)
MAthematical VISual instruction tuning	multimodal mathematical reasoning	(Zhang et al., 2024)
Multi-Agent Video Retrieval via Structured Video Understanding	video retrieval	(Zhang et al., 8 Jun 2026)
Multi-Objective Alignment via Value-Guided Inference-Time Search	LLM alignment	(Carleton et al., 19 Aug 2025)
Micro-surgical Artificial Vascular anastomosIS	surgical dataset	(Choi et al., 26 Nov 2025)
Modular Autonomous Virtualization System	quantum-dot control	(Rao et al., 2024)
Managing Datacenters using Smartphones	mobile systems	(Shankar et al., 2018)

A common source of confusion is that these systems are unrelated despite sharing the same acronym. The astronomy literature uses MAVIS for a visible-wavelength MCAO instrument on the VLT (Greggio et al., 2021), whereas QMAVIS is a long video-audio understanding pipeline (Lin et al., 10 Jan 2026), and the surgical MAVIS dataset is a component of SurgMLLMBench (Choi et al., 26 Nov 2025).

The record also contains internal numerical discrepancies in some summaries. One benchmark abstract reports 157K visual QA instances, while the detailed exposition reproduces a table totaling 24,000 QA pairs (Song et al., 15 Nov 2025). Likewise, the mathematical-visual instruction-tuning abstract reports 558K diagram-caption pairs for MAVIS-Caption, whereas the detailed exposition states 588 K pairs (Zhang et al., 2024). This suggests that exact dataset cardinalities should be verified against the corresponding paper version when those numbers are operationally important.

2. MAVIS as a visible-wavelength MCAO instrument for the VLT

In astronomy, MAVIS is the Multi-conjugate Adaptive-optics Visible Imager-Spectrograph planned for the VLT Adaptive Optics Facility (McDermid et al., 2020). The Phase A science case describes it as a general-purpose instrument intended to deliver the diffraction limit of an 8 m telescope in the visible, with a mean V-band Strehl ratio of $>10\%$ and goal $>15\%$ across a relatively large $30$ arc second science field, together with both a Nyquist-sampled imager and an integral-field spectrograph spanning $370$–$1000$ nm (McDermid et al., 2020). The corresponding diffraction-limited scale at $550$ nm is given as $<20$ mas, via $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$ (McDermid et al., 2020).

The instrument concept evolved through design and performance studies. The optical trade-off paper describes MAVIS as an instrument proposed for the VLT UT4 Nasmyth platform, exploiting the existing Adaptive Optics Facility with 4 sodium Laser Guide Stars and an adaptive secondary mirror with 1 170 actuators, while adding two post-focal deformable mirrors conjugated at $\sim 4$ km and $\sim 12$ km, and three Natural Guide Stars for tomographic reconstruction and correction of atmospheric turbulence (Greggio et al., 2021). The Phase A science case instead describes an Adaptive Optics Module containing two transmissive post-focus deformable mirrors plus the AOF adaptive secondary mirror, for a total of 5 420 actuators, together with eight $>15\%$0 Shack–Hartmann LGS sensors and three near-infrared-sensing NGS units (McDermid et al., 2020). The performance-estimation paper reflects the later architecture explicitly as three deformable mirrors at $>15\%$1, $>15\%$2, and $>15\%$3 km, with 8 LGS SH WFSs and 3 NGS SH WFSs, totaling 11 wavefront sensors (Agapito et al., 2022).

The associated optical studies emphasize stringent wavefront, distortion, telecentricity, and throughput constraints. For the adaptive-optics module, the optical design paper states a residual static and non-common-path aberration goal of $>15\%$4 nm RMS per science channel, a total optical throughput target $>15\%$5 in $>15\%$6–$>15\%$7 nm and $>15\%$8 in $>15\%$9–$30$0 nm, distortion $30$1, telecentricity $30$2, and minimized field curvature (Greggio et al., 2021). Three post-focal relay classes were studied—refractive, reflective, and catadioptric—with the refractive on-axis lens design identified as the best balance between optical performance, throughput, engineering simplicity, and integration risk (Greggio et al., 2021).

Performance modeling combines analytical and Monte Carlo methods. The 2022 performance-estimation study gives the residual wavefront budget as

$30$3

with representative contributions of $30$4 nm RMS fitting error, $30$5 nm tomography plus generalized fitting plus aliasing, $30$6 nm measurement noise, $30$7 nm temporal error, $30$8 nm sodium elongation/truncation, $30$9 nm LGS tip-tilt jitter, $370$0 nm static calibration and non-common-path “extra,” and $370$1 nm post-AVC vibrations, yielding net $370$2 nm and V-band Strehl $370$3 (Agapito et al., 2022). That study reports that the system-level requirements are met: 10% V-band SR over 30″ and sky coverage $370$4 at the South Galactic Pole with EE50 $370$5 (Agapito et al., 2022).

3. Synthetic calibration, simulation, and astrometry for the astronomical MAVIS

A substantial later literature around the astronomical MAVIS concerns calibration, tomography, and astrometric exploitation. The SynIM library was introduced as a GPU-accelerated Python package for synthetic interaction, projection, and covariance matrices in next-generation adaptive optics, explicitly including MAVIS (Agapito et al., 5 Jun 2026). The MAVIS-focused summary states that visible-band MCAO on an 8 m telescope pushes deformable-mirror actuator counts into the few-thousand range, for example 1600–2500 actuators per DM, and uses 3–6 natural guide-star Shack–Hartmann WFSs with up to $370$6 subapertures. Under sub-$370$7 nm wavefront-error budgets at $370$8, daytime poke-matrix measurements become infeasible, so SynIM replaces empirical interaction matrices with synthetic matrices that incorporate the exact optical geometry (Agapito et al., 5 Jun 2026).

The key geometric device is a single composite affine transform combining DM and WFS shifts, rotations, and magnifications: $370$9 with $1000$0, $1000$1, and $1000$2 defined in homogeneous coordinates (Agapito et al., 5 Jun 2026). According to the paper, applying $1000$3 once to the high-resolution phase screen avoids multiple resamplings, preserves sub-pixel accuracy when $1000$4 are fractional and $1000$5 differs from unity by $1000$6, and prevents stepwise degradation of high-frequency content (Agapito et al., 5 Jun 2026). For Shack–Hartmann slope extraction, the optimized numerical derivative is reported to yield RMS modal error of numerical derivatives of $1000$7, up to 3× better than G-tilt at high spatial frequencies (Agapito et al., 5 Jun 2026).

The MAVIS benchmarks reported for SynIM are operationally consequential. Single-WFS interaction-matrix generation for a $1000$8 SH $1000$9 modes system is given as $550$0 s on a CPU and $550$1 s on an NVIDIA L40S GPU, corresponding to $550$2 speed-up; the GPU memory footprint per DM–WFS pair is $550$3 GB in float32; and a full 3-WFS tomographic dictionary with 27 matrices is computed in $550$4 s (Agapito et al., 5 Jun 2026). Comparison to PASSATA diffractive models yields modal RMS discrepancies $550$5 nm, while end-to-end SPECULA simulations report closed-loop residual WFE of 84.2 nm RMS for SynIM versus 84.7 nm RMS for PASSATA, and Strehl at $550$6 of 0.25 versus 0.24 (Agapito et al., 5 Jun 2026). The same summary states that SynIM integrates with SPRINT so that misregistrations estimated from telemetry slopes can be ingested and the synthetic IM and MMSE reconstructor regenerated in $550$7 s per WFS–DM pair (Agapito et al., 5 Jun 2026).

Astrometry is another defining technical theme. One study frames MAVIS as an instrument for high-precision ground-based astrometry in the visible, using MAVISIM, SuperStar, and DAOPHOT to simulate imaging performance and incorporate telemetry into astrometric analysis (Taheri et al., 2024). That paper expresses the total 2D astrometric error per star as

$550$8

and identifies distortion and pixel-scale calibration as dominant terms in current simulations (Taheri et al., 2024). Under realistic turbulence scenarios, the quoted RMS astrometric residuals are in the $550$9–$<20$0 mas range per epoch, with projected improvements aiming at 50–100 $<20$1as for bright stars ($<20$2), and $<20$3as after multi-epoch averaging (Taheri et al., 2024).

A related MAVISIM study focused on detecting an intermediate-mass black hole in a globular cluster by proper-motion measurements (Monty et al., 2021). It models high-order PSF spatial variability, low-order tip-tilt residuals, and static field distortion, combining them through

$<20$4

For a single 30 s exposure, that work reports $<20$5 for stars in $<20$6 and $<20$7 at $<20$8, and concludes that the bright-star 50 $<20$9as goal is recoverable under the modeled conditions (Monty et al., 2021). In the NGC 3201-like simulation with a central $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$0 IMBH, the recovered inner-bin velocity-dispersion error is $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$1 km/s, with the no-IMBH model lying $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$2 below the IMBH case in the inner $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$3 (Monty et al., 2021).

4. MAVIS in multimodal AI, retrieval, and language-model control

In multimodal AI, MAVIS and QMAVIS designate several unrelated frameworks. QMAVIS—“Q Team-Multimodal Audio Video Intelligent Sensemaking”—is a late-fusion pipeline for long video-audio understanding built from a video LMM, Whisper-Large v3 ASR, caption interleaving, and an aggregation LLM (Lin et al., 10 Jan 2026). It chunks arbitrarily long videos into 30 s clips, converts visual and audio streams into text, constructs

$\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$4

and applies a final LLM over $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$5, recursively if the context window is exceeded (Lin et al., 10 Jan 2026). On VideoMME, QMAVIS is reported at 66.46% Top-1 versus 47.9% for VideoLLaMA2, a +38.75% relative gain; on PerceptionTest and EgoSchema it reaches 57.72% and 65.00%, respectively (Lin et al., 10 Jan 2026). Ablations indicate that removing the aggregation LLM lowers VideoMME to 63.00%, while dropping speech-to-text lowers it to 48.18%, showing that the audio-transcription module contributes roughly 18 pp on that benchmark (Lin et al., 10 Jan 2026).

A different MAVIS is a benchmark for multimodal source attribution in long-form visual question answering (Song et al., 15 Nov 2025). The abstract describes it as the first benchmark for multimodal source attribution systems and states that the dataset comprises 157K visual QA instances with fact-level citations to multimodal documents (Song et al., 15 Nov 2025). The detailed exposition, however, constructs the benchmark from Wikipedia and arXiv documents and reproduces a split table totaling 24,000 QA pairs, with answers evaluated for informativeness, groundedness, and fluency (Song et al., 15 Nov 2025). The groundedness metric is defined as

$\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$6

and the reported findings are that multimodal RAG improves informativeness and fluency over unimodal RAG but remains weaker on groundedness for image documents than for text documents (Song et al., 15 Nov 2025). The same paper reports a trade-off: Context-Balanced Prompting raises groundedness from 55.4 to 67.1 while lowering informativeness from 68.1 to 65.8 (Song et al., 15 Nov 2025).

Another MAVIS in this cluster is “MAthematical VISual instruction tuning with an automatic data engine”, a multimodal mathematical reasoning pipeline (Zhang et al., 2024). The abstract states that MAVIS-Caption contains 558K diagram-caption pairs and MAVIS-Instruct contains 834K visual math problems with CoT rationales, whereas the detailed exposition gives 588 K pairs for MAVIS-Caption and again 834 K problems for MAVIS-Instruct (Zhang et al., 2024). The training curriculum has three stages in the detailed exposition: contrastive fine-tuning of CLIP-Math, diagram-language alignment through a projection layer into Mammoth2-7B, and supervised instruction tuning on MAVIS-Instruct (Zhang et al., 2024). Reported results include 27.5% overall accuracy on MathVerse for MAVIS-7B, versus 16.5% for InternLM-XComposer2-7B and 24.5% for LLaVA-NeXT-110B, and 66.7% top-1 on GeoQA (Zhang et al., 2024).

The acronym also appears in retrieval and alignment. “MAVIS: Multi-Agent Video Retrieval via Structured Video Understanding” reformulates video retrieval as a cooperative pipeline with a structured semantic library, planner, specialized scene/object/action agents, and a Logic-aware Debate veto mechanism (Zhang et al., 8 Jun 2026). On MSR-VTT, the paper reports R@1 = 78.6, R@5 = 94.9, R@10 = 96.3, and an efficiency figure of $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$7 G FLOPs, versus 100 000 G FLOPs for brute-force ViCLIP, i.e. $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$8 speedup (Zhang et al., 8 Jun 2026). “MAVIS: Multi-Objective Alignment via Value-Guided Inference-Time Search” instead denotes an inference-time alignment framework for LLMs, combining objective-specific value models $\theta_{\mathrm{diff}} \simeq 1.22 \lambda / D$9 through

$\sim 4$0

with theoretical monotonic-improvement guarantees under KL-regularized policy iteration and empirical Pareto-front improvements over MORLHF, Rewarded Soups, and MOD on HH-RLHF and Summarize-from-Feedback (Carleton et al., 19 Aug 2025).

5. MAVIS in surgical scene understanding

In surgical AI, MAVIS denotes the Micro-surgical Artificial Vascular anastomosIS dataset introduced within SurgMLLMBench (Choi et al., 26 Nov 2025). It consists of 10 652 static frames extracted from 19 full-procedure videos, with original videos at 1920×1080 px sampled at 1 fps, covering microsurgical vascular anastomosis on 1 mm artificial vessels (Choi et al., 26 Nov 2025). The workflow taxonomy is three-level: Stages (6), Phases (4 per stage), and Steps (up to 8); segmentation uses 8 visual classes: Background material, Forceps, Scissors, Vascular clamps, Needle holder, Vessel, Needle, and Thread (Choi et al., 26 Nov 2025).

The annotations are explicitly multimodal. Pixel-level segmentation masks were manually drawn in COCO format by expert microsurgeons and second-checked by a different expert, particularly for fine structures such as needles and threads (Choi et al., 26 Nov 2025). Structured VQA uses five question types—workflow queries, instrument count, instrument type, instrument action, and dataset source—with fixed templates and atomic-string answers so that exact-match accuracy is well defined (Choi et al., 26 Nov 2025). The evaluation metrics are strict string-matching accuracy for VQA and mean Intersection over Union for segmentation: $\sim 4$1

The baseline results underscore the difficulty of microsurgical imagery. On MAVIS, LLaVA (fine-tuned) reaches 75.7% stage accuracy, 69.4% phase accuracy, 44.8% step accuracy, and 68.5% count accuracy; OMG-LLaVA (no FT) reaches 53.96% segmentation mIoU, whereas OMG-LLaVA (FT) drops to 36.89% mIoU (Choi et al., 26 Nov 2025). The paper attributes these challenges to extremely small and thin instruments, low contrast between vessel walls and background, specular highlights and shadowing under microscope lighting, and rapid, subtle motions requiring precise temporal and spatial grounding (Choi et al., 26 Nov 2025). Within SurgMLLMBench, MAVIS is described as the sole provider of Stage recognition samples (10 652 frames) and the micro-surgical training domain balancing laparoscopic and robotic sources (Choi et al., 26 Nov 2025).

6. MAVIS in SLAM, quantum-dot control, and infrastructure monitoring

The acronym also names several systems in robotics, condensed-matter device control, and mobile systems. In robotics, “MAVIS: Multi-Camera Augmented Visual-Inertial SLAM using SE$\sim 4$2 Based Exact IMU Pre-integration” is an optimization-based SLAM system for a rigid multi-camera rig of 4 fisheye cameras plus a 6-axis IMU (Wang et al., 2023). Its central technical contribution is exact IMU pre-integration on the extended state manifold $\sim 4$3, with closed-form $\sim 4$4 and $\sim 4$5 terms in the discrete exponential update, intended to improve robustness under fast rotational motion and extended integration intervals (Wang et al., 2023). The paper reports first place in all vision-IMU tracks of the Hilti SLAM Challenge 2023, with 1.7 times the score compared to the second place (Wang et al., 2023).

In quantum hardware, MAViS—the Modular Autonomous Virtualization System—is a fully autonomous, multi-layer framework for defining virtual gates in large two-dimensional quantum-dot arrays (Rao et al., 2024). It decomposes gate virtualization into five sequential layers: sensor-gate virtualization, plunger orthogonalization, plunger normalization, barrier coarse virtualization in the OFF regime, and barrier fine virtualization in the ON regime (Rao et al., 2024). The system uses an ensemble of five pre-trained CNN classifiers on $\sim 4$6 charge-stability diagrams, Hough transforms for line detection, and Gaussian fitting on interdot probability maps, all trained on simulated “QFlow 2.0” data (Rao et al., 2024). Demonstrated on a 3-4-3 germanium hole-spin array with 10 plungers, 12 barrier gates, 4 sensor plungers, and 8 screening gates, MAViS acquired $\sim 4$7 maps and required $\sim 4$8 h to fully virtualize all 10 dots (Rao et al., 2024). Reported post-virtualization metrics are residual plunger lever arms $\sim 4$9, OFF-regime barrier-to-plunger shifts $\sim 12$0 mV over $\sim 12$1 mV pulses, and ON-regime residual shift $\sim 12$2 mV for $\sim 12$3 pulses up to 100 mV (Rao et al., 2024).

An earlier systems paper uses MAVIS for Managing Datacenters using Smartphones (Shankar et al., 2018). It proposes a three-tier architecture of publishers on datacenter nodes, a server farm of CEP engines using IBM InfoSphere Streams, and mobile subscribers on Android or iOS (Shankar et al., 2018). Its design goals include end-to-end alert latency under 10 s, support for thousands of physical hosts and hundreds of thousands of subscriptions, low smartphone overhead, dynamic subscriptions, and contextual data access (Shankar et al., 2018). In an evaluation using monitoring data from approximately 3,000 machines and $\sim 12$4K subscriptions, one 8-core CEP server handled $\sim 12$5K subscriptions with average CPU $\sim 12$6, while alert packets were approximately 500 bytes and a 10 min history fetch cost $\sim 12$7 KB (Shankar et al., 2018).

Across these domains, the recurring acronym does not imply methodological continuity. A plausible implication is that “MAVIS” now functions less as a uniquely identifying technical term than as a local project name whose meaning is determined by disciplinary context, citation, and surrounding vocabulary.