NIRVANA: Multifaceted Research Artifacts

• NIRVANA is a multi-context designation for research artifacts ranging from high-resolution astronomical interferometers to advanced ML systems, simulation codes, and educational data repositories.
• In astronomy, LINC-NIRVANA employs innovative interferometric and adaptive optics methods to achieve near-infrared diffraction-limited imaging with multiple guide stars.
• In computational astrophysics and machine learning, NIRVANA codes optimize simulations, video compression, and large model pruning, significantly reducing computational costs while improving performance.

NIRVANA is not a single technical object but a recurring designation used for several unrelated research artifacts. In astronomy, it most prominently denotes LINC-NIRVANA, a Fizeau-type interferometric imager for the Large Binocular Telescope equipped with a layer-oriented multi-conjugate adaptive optics module. In computational astrophysics, it denotes the public NIRVANA code and NIRVANA-III, used for magnetohydrodynamics, solar-wind modeling, and self-gravitating shearing-box simulations. In machine learning and data systems, the name has been used for a video implicit-representation compressor, a dataset on student use of ChatGPT during essay writing, a diffusion-serving cache, a structured pruning method for LLMs, a specialized generalist model with task-aware memory, and a multi-modal analytics framework (Bergomi et al., 2018, Louis et al., 2015, Maiya et al., 2022, Jelson et al., 8 Apr 2026, Agarwal et al., 2023, Ai et al., 17 Sep 2025, Jiang et al., 30 Oct 2025, Zhu et al., 25 Nov 2025).

1. Naming scope and major referents

The term appears across distinct subfields with no shared technical lineage implied by the name alone. In bibliographic practice, the arXiv identifier is therefore the most reliable disambiguator.

Designation	Domain	Characteristic description
LINC-NIRVANA (LN)	Astronomical instrumentation	high resolution, near infrared imager; Fizeau-type interferometric imager on the Large Binocular Telescope
NIRVANA / NIRVANA-III	Computational astrophysics	public finite-volume MHD code; shearing-box Poisson solvers with vertical vacuum boundary conditions
Nirvana	Galaxy kinematics	“Nonaxisymmetric Irregular Rotational Velocity ANAlysis” for MaNGA velocity fields
NIRVANA	Video INR compression	Neural Implicit Representations of Videos with Adaptive Networks and Autoregressive Patch-wise Modeling
NIRVANA	Educational AI dataset	dataset capturing how university students use generative AI while writing an analytical essay
Nirvana	ML systems	approximate-caching for diffusion serving, structured pruning, task-aware memory, and LLM-native analytics

The astronomical usages are tied to observatory hardware and physics codes, whereas the machine-learning usages denote algorithmic systems, datasets, or model families (Zanger et al., 2024, Schleich et al., 2023).

2. LINC-NIRVANA as an interferometric AO instrument

LINC-NIRVANA is a high resolution, near infrared imager mounted at the bent-Gregorian focus of the Large Binocular Telescope. Its final goal is to perform Fizeau interferometric imaging by combining the two 8.4 m primaries into an equivalent baseline of 22.8 m. In the simple two-aperture Fizeau mode, the theoretical angular resolution is $\theta \simeq \lambda/(2B)$ with $B = 22.8\,\mathrm{m}$; the representative values given are $\theta_H \simeq 7.2\,\mathrm{mas}$ at H-band and $\theta_{K'} \simeq 10\,\mathrm{mas}$ at K′-band (Bergomi et al., 2018).

Physically, LN is described as a massive bench of approximately $6 \times 4 \times 4.5$ meters with more than 250 optical components. Its principal subsystems are four multi-pyramid wavefront sensors, forty “Star Enlargers,” an interferometric beam-combining module with a fringe tracker, and a NIR science camera. Up to 20 Natural Guide Stars can be used simultaneously: 12 in the ground layer and 8 in the high layer. The optical relay splits the field so that the Ground-layer Wavefront Sensor sees an annulus from $2.8'$–$6'$ and the High-layer Wavefront Sensor scans the central $2'$ (Bergomi et al., 2018).

The adaptive-optics architecture is a “Multiple Field-of-View, Layer-Oriented” Multi-Conjugate AO scheme. Ground-layer correction is provided by the adaptive secondary mirror, specified as 672 actuators in the alignment overview and approximately 672 actuators in the commissioning summary, while the high layer is corrected by an internal Xinetics deformable mirror with 349 actuators, conjugated at about 7.1 km. The ground-layer loop runs at $\leq 1\,\mathrm{kHz}$, and the high-layer loop refines the residuals on the central field. The near-IR imager covers roughly a $10''\times10''$ field at a 3 mas/pixel scale, with diffraction-limited imaging from a single 8.4 m aperture or combined interferometric operation through the beam combiner (Santhakumari et al., 2021).

A further defining feature is the project’s modular or decoupled test philosophy. The architecture separates ground-layer AO on each aperture, mid-high-layer AO around 349-actuator Xinetics mirrors, a near-infrared fringe tracker, a Fizeau beam combiner, and the cryogenic NIR science camera. This decoupling was adopted explicitly for risk mitigation and commissioning efficiency, with Pathfinder serving as the first on-sky subsystem (Kopon et al., 2014).

3. Rotation control, calibration, commissioning, and reconstruction in LINC-NIRVANA

Because the telescope is on an alt-azimuth mount, LINC-NIRVANA must stabilize the focal plane on the pyramids while accommodating field rotation. The Ground-layer Wavefront Sensors sit on a rotating bearing, and the mid-high sensors use K-mirror optical derotators. In both cases, the focal plane is fixed with respect to the pyramids, but the pupil image rotates on the wavefront-sensor detector. The corresponding deformable-mirror projection onto the WFS therefore changes continuously, producing a misalignment between actuator geometry and sub-aperture geometry (Arcidiacono et al., 2010, Arcidiacono et al., 2018).

The control problem is expressed with the unrotated relation $B = 22.8\,\mathrm{m}$0, where $B = 22.8\,\mathrm{m}$1 is the vector of measured WFS slopes and $B = 22.8\,\mathrm{m}$2 the DM commands. If the pupil is rotated by an angle $B = 22.8\,\mathrm{m}$3, the required update is

$B = 22.8\,\mathrm{m}$4

with $B = 22.8\,\mathrm{m}$5 the slope-space rotation operator. In the calibration procedure, interaction matrices are measured at several angles $B = 22.8\,\mathrm{m}$6, counter-rotated to a common pupil orientation, averaged as

$B = 22.8\,\mathrm{m}$7

and converted into a regularized least-squares reconstructor

$B = 22.8\,\mathrm{m}$8

For on-sky use, $B = 22.8\,\mathrm{m}$9 is precomputed for each $\theta_H \simeq 7.2\,\mathrm{mas}$0 step and uploaded to the BCU in real time as the bearing or K-mirror moves (Arcidiacono et al., 2018).

The numerical-rotation paper quantifies the operational envelope. Maximum on-sky pupil rotation speed is given as $\theta_H \simeq 7.2\,\mathrm{mas}$1. In commissioning tests with the First-Light AO system, the loop remains essentially unaffected for rotation errors $\theta_H \simeq 7.2\,\mathrm{mas}$2 up to $\theta_H \simeq 7.2\,\mathrm{mas}$3 and can stay closed up to $\theta_H \simeq 7.2\,\mathrm{mas}$4. The strategy described is to pre-store two control matrices on the AdSec BCU, compute $\theta_H \simeq 7.2\,\mathrm{mas}$5 in software when the measured angle passes a threshold such as $\theta_H \simeq 7.2\,\mathrm{mas}$6, upload the new matrix, and swap pointers. The real-time CPU requirement is on the order of $\theta_H \simeq 7.2\,\mathrm{mas}$7–$\theta_H \simeq 7.2\,\mathrm{mas}$8, and one new upload per $\theta_H \simeq 7.2\,\mathrm{mas}$9–$\theta_{K'} \simeq 10\,\mathrm{mas}$0 is typically sufficient. The same study reports H-band Strehl $\theta_{K'} \simeq 10\,\mathrm{mas}$1–$\theta_{K'} \simeq 10\,\mathrm{mas}$2 at $\theta_{K'} \simeq 10\,\mathrm{mas}$3, a drop of less than $\theta_{K'} \simeq 10\,\mathrm{mas}$4 at $\theta_{K'} \simeq 10\,\mathrm{mas}$5, and a $\theta_{K'} \simeq 10\,\mathrm{mas}$6–$\theta_{K'} \simeq 10\,\mathrm{mas}$7 drop at $\theta_{K'} \simeq 10\,\mathrm{mas}$8 (Arcidiacono et al., 2010).

The broader commissioning campaign confirmed the optical train and gradually closed the AO loops. First photons in November 2016 confirmed the system, and by March–June 2017 on-sky tests yielded Ground-Layer AO results on the SX arm with open-loop FWHM $\theta_{K'} \simeq 10\,\mathrm{mas}$9 improving to closed-loop FWHM $6 \times 4 \times 4.5$0 in K′. In Multi-Field-of-View MCAO, the reported sequence was open-loop FWHM $6 \times 4 \times 4.5$1, GWS-only closed-loop FWHM $6 \times 4 \times 4.5$2 with 40 modes, and full GWS+HWS closed-loop FWHM $6 \times 4 \times 4.5$3 with $6 \times 4 \times 4.5$4 modes. The later commissioning summary reports Strehl ratio up to $6 \times 4 \times 4.5$5–$6 \times 4 \times 4.5$6 on axis at K′ under $6 \times 4 \times 4.5$7 seeing, with uniform $6 \times 4 \times 4.5$8 over a $6 \times 4 \times 4.5$9 field, control bandwidth around 200 Hz, and total latency $2.8'$0 (Bergomi et al., 2018, Santhakumari et al., 2021).

Several commissioning lessons were made explicit. Major challenges included vibration-induced PSF elongation at 9 Hz and 16 Hz, changing WFS geometry due to derotation, temperature excursions from $2.8'$1 to $2.8'$2, flexure-induced misregistration, and a detector “aurora effect” on CCD39. The solutions described include OVMS+ accelerometer feed-forward, pseudo-synthetic interaction matrices, thermal shrouds, real-time WFS CCD tracking in XYZ to maintain registration within 0.1 pixel, and refined guide-star acquisition algorithms that acquire 4 GWS + 3 HWS stars of approximately $2.8'$3 in $2.8'$4 (Santhakumari et al., 2021).

LINC-NIRVANA also motivated specialized inverse methods for high-dynamic-range interferometric imaging. For simulated observations of a young stellar object jet, a multi-component Richardson–Lucy formulation decomposes the unknown object as $2.8'$5, with $2.8'$6 a known-position point source and $2.8'$7 the diffuse jet. A Tikhonov-like penalty is applied to $2.8'$8, and alternating multiplicative updates are used. The reported outcome is a reconstruction of the original jet intensity distribution with an error smaller than 10%, together with strong reduction of artifacts relative to standard multi-frame Richardson–Lucy deconvolution (Camera et al., 2012).

4. Astrophysical simulation and inference codes named NIRVANA

In computational astrophysics, NIRVANA denotes a public finite-volume code for three-dimensional, compressible magnetohydrodynamics. In the convection study used for local correlation tracking, the code solves the continuity, momentum, induction, and energy equations for an ideal gas of constant molecular weight in Cartesian coordinates, with periodic horizontal boundaries and impermeable, stress-free vertical boundaries. The cited setup uses a domain $2.8'$9, $6'$0, a $6'$1 mesh, Rayleigh number $6'$2, and magnetic and thermal Prandtl numbers of 0.1. Temperature slices at $6'$3 were then used to test apodizing windows for local correlation tracking: the triangular and trapezoidal windows perform the best and worst, respectively, the narrowest width $6'$4 px gives correlations of about 0.64 in both velocity components, and the highest-velocity bin reaches correlation $6'$5 while low-speed cells are poorly recovered at about 0.43 (Louis et al., 2015).

The same code base was extended in the NIRwave study by adding an explicit wave-turbulence-driven coronal-heating term $6'$6 to the MHD energy equation. The adapted model assumes an axisymmetric dipole magnetic field with polar strength $6'$7 and constrains the solar-wind solution with Parker Solar Probe data. The steady-state solution reconstructs a bimodal solar wind with slow and fast wind speeds of $6'$8 and $6'$9, respectively, and a global mass-loss rate of $2'$0 solar masses per year. The number density is reported to be in good agreement with derived empirical constraints, while larger deviations remain for radial velocity and temperature; the NIRwave model is also reported to be in better agreement with the observational constraints than a polytropic wind model generated with NIRVANA (Schleich et al., 2023).

NIRVANA-III extends the name into self-gravitating shearing-box calculations. For Poisson’s equation with shear-periodic boundary conditions in the plane and vertical vacuum boundary conditions, two spectral solvers are described: the Superposition Analytical-Spectral Hybrid Approach and the Vico-Greengard-Ferrando method with hybrid boundary conditions. The latter is described as slightly more accurate, and the overall method exhibits relative errors down to $2'$1–$2'$2 with only $2'$3–$2'$4 cells, with approximately third-order convergence in the key tests. Implemented with P3DFFT pencil decomposition, the solver scales to 4096 CPU cores and remains below approximately 6% of total runtime (Restrepo et al., 11 Oct 2025).

A separate astronomical software usage is the Python code Nirvana, expanded as “Nonaxisymmetric Irregular Rotational Velocity ANAlysis.” It is a physically motivated Bayesian forward-modeling framework for two-dimensional galaxy velocity fields, using a parameter vector that includes kinematic center, systemic velocity, inclination, major-axis position angle, bisymmetric position angle, and radial profiles for first- and second-order modes. Applied to SDSS-IV/MaNGA, it produced models and rotation curves for 1263 unique barred galaxies and a matched unbarred control sample. The reported barred sample exhibits elevated non-circular motions, with second-order amplitudes at roughly $2'$5 of about $2'$6 in gas versus about $2'$7 in the control sample, and about $2'$8 in stars versus about $2'$9 in the control sample (Zanger et al., 2024).

5. NIRVANA in machine learning systems, compression, and serving

In video compression, NIRVANA denotes an implicit neural representation method that models video as groups of frames and predicts spatio-temporal patch volumes rather than individual pixels or full frames. The default implementation uses groups of length $\leq 1\,\mathrm{kHz}$0, patch size $\leq 1\,\mathrm{kHz}$1, and an architecture with $\leq 1\,\mathrm{kHz}$2 SIREN MLP layers of hidden size 512 plus a lightweight convolutional upsampling decoder. Networks are trained autoregressively across groups, with the current group initialized from the previous group and the quantized residual stored rather than the full parameter set. The training loss combines MSE with an entropy term over quantized integer latents. On the UVG-HD benchmark, the reported result is 37.70 dB PSNR at 0.86 BPP, versus 37.36 dB at 0.92 BPP for NeRV, with encoding time reduced from about 80 h to 6.7 h and decoding speed improved from 11.0 fps to 65.4 fps. On UVG-4K, it maintains similar PSNR and BPP with a sixfold speedup, and on long YouTube-8M clips it remains near 35 dB while NeRV degrades from 33.4 to 30.5 dB (Maiya et al., 2022).

For diffusion-model serving, Nirvana denotes an approximate-caching system that reuses intermediate latent noise states from prior generations for similar prompts. The architecture includes a CLIP text embedding generator, a One-Class SVM match predictor, a vector database, a cache-selector heuristic over $\leq 1\,\mathrm{kHz}$3 skipped steps, intermediate-state storage, the diffusion model itself, and an offline LCBFU cache maintainer. The LCBFU score is defined as $\leq 1\,\mathrm{kHz}$4, combining frequency of reuse with computational benefit. On the reported production workloads, Nirvana with the predictor achieves hit-rate $\leq 1\,\mathrm{kHz}$5, compute saving 21%, latency reduction 19.8%, and cost reduction 19%; without the predictor, hit-rate is about 93% and compute saving 23%. Throughput is reported as Nirvana$\leq 1\,\mathrm{kHz}$6 versus Vanilla, while image-quality deltas relative to the full model remain small (Agarwal et al., 2023).

For large-language-model compression, NIRVANA denotes “NTK-InfoRmed adaptiVe neuron/AttentioN heAd pruning.” Its saliency is based on the first-order Taylor term

$\leq 1\,\mathrm{kHz}$7

and its rationale is to preserve the Neural Tangent Kernel under Adam-like dynamics while allocating sparsity adaptively between MLP neurons and attention heads through a parameter $\leq 1\,\mathrm{kHz}$8. Calibration data are selected by minimizing a KL divergence proxy between dense and pruned outputs. On Llama3.1-8B at 20% sparsity, the reported zero-shot result is perplexity 13.38/19.77/26.20 and average accuracy 55.09, improving after LoRA recovery tuning to 12.37/18.58/25.00 and average accuracy 61.51. On T5-base with 50% MLP pruning, the reported downstream results include 90.69 on MRPC, 82.45 on CoLA, 94.27 on SST-2, and 85.99 on MNLI. Inference on an A100 at 20% pruning yields 0.28 s latency versus 0.33 s for the dense 8.03 B model (Ai et al., 17 Sep 2025).

Another ML usage is the 1.3 B-parameter Specialized Generalist Model Nirvana, which replaces full self-attention with a hybrid of Sliding-Window Attention and linear attention, and introduces a Task-Aware Memory Trigger together with a Specialized Memory Updater. The model is described as $\leq 1\,\mathrm{kHz}$9 in sequence length and performs cross-layer online gradient descent on low-dimensional fast codes at inference time. In the reported evaluations, Nirvana obtains Wiki perplexity 16.05, LAMBADA perplexity 11.56, and average reasoning accuracy 56.51%, while on a Single Needle-In-Haystack retrieval benchmark it reaches 99.1% average accuracy. In the MRI study, with a frozen Nirvana backbone and lightweight codecs, the reported reconstruction numbers at undersampling $10''\times10''$0 are SSIM $10''\times10''$1, PSNR $10''\times10''$2 dB, and NMSE $10''\times10''$3 (Jiang et al., 30 Oct 2025).

A further systems usage is the multi-modal analytics framework Nirvana, which exposes programmable semantic operators—map, filter, reduce, and rank—and combines an agentic logical optimizer with a cost-aware physical optimizer. The logical optimizer uses natural-language transformation rules and random-walk-based search; the physical optimizer uses an improvement-score metric to choose among LLM backends and employs evaluation pushdown and computation reuse. Across three real-world benchmarks, the reported effect is an end-to-end runtime reduction of 10%–85% and an average system cost reduction of 76%, with higher answer accuracy than the stated baselines on the Movie, Estate, and Game workloads (Zhu et al., 25 Nov 2025).

6. NIRVANA as a dataset for student–AI writing behavior

In education and HCI, NIRVANA denotes a dataset designed to reproduce how students use generative AI for essay writing. The dataset includes 77 U.S. university students who completed a 30-minute analytical argumentative essay based on an ACT-style prompt about automation and human labor. The writing environment was a custom web editor with an integrated in-house ChatGPT interface built on GPT-3.5-turbo, allowing unrestricted, student-driven queries (Jelson et al., 8 Apr 2026).

The dataset records four principal modalities: keystroke-level logs including insertions, deletions, replacements, and cursor movements; copy/paste, cut, and drag events; full timestamped ChatGPT transcripts; and pre- and post-task surveys covering demographics and several psychometric instruments. The editor was built with CodeMirror 5 and the CodeMirror-Record library, and all events were logged with millisecond timestamps. The paper defines essay-level features such as word count $10''\times10''$4, readability through Flesch–Kincaid Reading Ease and Grade Level, and query frequency $10''\times10''$5. It also defines the Human Contribution Ratio,

$10''\times10''$6

and the Human Edit Ratio,

$10''\times10''$7

where $10''\times10''$8 and $10''\times10''$9 are human additions and deletions, and $B = 22.8\,\mathrm{m}$00 and $B = 22.8\,\mathrm{m}$01 are ChatGPT-pasted words and deleted pasted words (Jelson et al., 8 Apr 2026).

K-means clustering on $B = 22.8\,\mathrm{m}$02 with $B = 22.8\,\mathrm{m}$03 yielded four writing profiles: Lead Authors ($B = 22.8\,\mathrm{m}$04, mean HCR 0.98, mean HER 0.95), Collaborators ($B = 22.8\,\mathrm{m}$05, mean HCR 0.53, mean HER 0.62), Vibe Writers ($B = 22.8\,\mathrm{m}$06, mean HCR 0.01, mean HER 0.06), and Drafters ($B = 22.8\,\mathrm{m}$07, mean HCR 0.09, mean HER 0.64). The accompanying replay interface presents the editor history and ChatGPT conversation side by side, with timeline annotations for inquiry, copy, and paste events and a word-count-over-time graph. Empirically, query count is positively correlated with word count ($B = 22.8\,\mathrm{m}$08, $B = 22.8\,\mathrm{m}$09), time spent ($B = 22.8\,\mathrm{m}$10, $B = 22.8\,\mathrm{m}$11), and Dale–Chall readability grade ($B = 22.8\,\mathrm{m}$12, $B = 22.8\,\mathrm{m}$13). The study further reports that Most students (Lead Authors) wrote independently despite access to AI, while perceived ownership was higher for Lead Authors than for Collaborators and Vibe Writers (Jelson et al., 8 Apr 2026).