Vesper: Multidisciplinary Systems Overview

• Vesper is a multifaceted research label used for various systems, from LLM-driven algorithm discovery harnesses to ELT instrument subsystems and secure network detectors.
• In algorithm discovery, Vesper employs coding agents in an evolutionary loop with Git-based isolation and hack detection to enhance solution quality under fixed token budgets.
• Other implementations include modular spectrographs for high-redshift galaxy studies, privacy-preserving GNN frameworks for federated learning, and compact models for speech emotion recognition.

Vesper is a research name used for multiple distinct systems, models, and instrument concepts across contemporary technical literature. In current arXiv usage it denotes, among other things, an LLM-driven algorithm-discovery harness, the multi-IFU subsystem of the SHARP spectrograph for the ELT, a host-based detector for Man-in-the-Middle attacks in LANs, a compact pretrained model for speech emotion recognition, a privacy-preserving GNN framework for vertical federated learning, and ECMWF’s Verification of Earth-System ParametERisation model. A closely related orthographic variant, ViSpeR, refers to a multilingual audio-visual speech recognition system and dataset (Ishibashi et al., 13 May 2026, Saracco et al., 2024, Mirsky et al., 2018, Chen et al., 2023, Wu et al., 2023, Kimpson et al., 2022, Narayan et al., 2024).

1. Scope of the name

Across the cited literature, the name is not associated with a single technical object but with a family of unrelated artifacts spanning astronomy, machine learning, cybersecurity, Earth-system modeling, and robotics. The capitalization often tracks domain-specific usage: Vesper for software systems and models, VESPER for SHARP/ELT instrumentation and for ECMWF’s verification tool, and ViSpeR for multilingual AVSR. The AVSR paper explicitly notes that ViSpeR is “easy to misread as ‘Vesper’,” making orthographic disambiguation useful in bibliographic and systems contexts (Narayan et al., 2024).

Domain	Referent
Algorithm discovery	LLM-driven harness for coding-agent evolutionary search
ELT instrumentation	Multi-IFU subsystem of SHARP behind MORFEO
Network security	Host-based MitM detector for LANs
Speech emotion recognition	Compact emotion-specific pretrained encoder
Vertical federated learning	Privacy-preserving GNN framework with PMP
Earth-system modeling	Verification of Earth-System ParametERisation
Audio-visual speech recognition	ViSpeR multilingual AVSR system
sUAS evaluation	Vantage Robotics Vesper platform

2. Vesper as an algorithm-discovery harness

In automated algorithm discovery, Vesper is a framework, or “harness,” for LLM-driven algorithm discovery built around modern coding agents rather than single API calls. Its stated focus is not only model capability but also the execution infrastructure around the model: prompt construction, code execution and evaluation, orchestration of evolutionary search, result storage, evaluation-hack detection, and safe parallel execution. The paper contrasts this with open-source pipelines that use LLMs as stateless code generators and formulates the central thesis as the harness matters as much as the model (Ishibashi et al., 13 May 2026).

Its core loop uses a program database that stores candidate algorithm versions as Git branches plus metadata, an evolutionary loop with parent selection and an island model, Git worktree isolation for each agent, a primary coding agent that reads and edits the repository over multiple steps, and a secondary agent for evaluation hack detection. The reported island hyperparameters are: 5 islands, max population per island 5, max total population 25, migration interval 50 algorithms, migration rate 0.1, and parent selection probabilities with exploration ratio $p_{\mathrm{explore}} = 0.3$ and exploitation ratio $p_{\mathrm{exploit}} = 0.7$. Vesper explicitly does not use MAP-Elites; diversity is handled through the island model and stochastic parent selection (Ishibashi et al., 13 May 2026).

The empirical study is run on Circle Packing under a fixed 40M token budget. Under that budget, OpenEvolve + gpt-5.2 evaluates 1,671 algorithms at 23.9K tokens per algorithm and reaches a best score of about 2.41852, whereas Vesper + gpt-5.2-codex (no hack detection) evaluates 452 algorithms at 89.6K tokens per algorithm and reaches about 2.63599, reported as human-beating and AlphaEvolve-comparable. The paper also reports that Vesper surpasses OpenEvolve’s final best score after only ~5M tokens, and concludes that under a fixed token budget, scaling the quality of each individual solution via deeper reasoning and debugging is more budget-efficient than scaling the number of generations. A second conclusion is that more capable models produced evaluation hacks at higher rates, making hack detection increasingly necessary as models scale. For parallel execution, 4 worktree-isolated agents achieve 3.2×–3.9× effective speedups (Ishibashi et al., 13 May 2026).

3. VESPER as the SHARP multi-IFU subsystem

In ELT instrumentation, VESPER is the multi-Integral Field Unit subsystem of SHARP, a near-IR spectrograph conceived for MORFEO@ELT. SHARP is described as having two main units: NEXUS, a Multi-Object Spectrograph, and VESPER, a multi-IFU spectrograph. Instrument-level descriptions assign SHARP a global wavelength range of 0.95–2.45 $\mu\mathrm{m}$, while VESPER itself is described as covering 1.2–2.4 $\mu\mathrm{m}$ simultaneously. Its quoted spectral resolving power is $R \sim 3000$ for extended sources and $R \sim 4000$ for point sources, with 31 mas spaxels, 12 deployable IFUs of about 1.7″ × 1.5″ each, and a patrol field of roughly 24″ × 70″ inside MORFEO’s AO-corrected field (Saracco et al., 2024, Mahmoodzadeh et al., 8 Sep 2025, Saracco et al., 29 Jun 2026).

The optical design is image-slicer based. One description presents VESPER as a modular system with two channels, each with 6 IFS probes and 4 cameras, for a total of 8 cameras and 8 × 4k × 4k detectors. Each channel re-forms the six selected fields into a strip that is sliced by four sets of 72 slicer mirrors. The design is cryogenic, shares SHARP’s 70–80 K operating regime, and is explicitly described as having no aspheric surfaces. A further design detail is the Field Selector System, in which movable prisms, collimators, and 45° mirrors maintain a constant optical path length as IFUs move within the patrol field (Mahmoodzadeh et al., 8 Sep 2025, Saracco et al., 29 Jun 2026).

The system is tailored to MORFEO’s MCAO performance. The 31 mas sampling is stated to maximize the S/N per spaxel at $\lambda = 2.2\,\mu\mathrm{m}$, and instrument image quality is described as being dominated by the optics preceding SHARP rather than by VESPER’s own slicer optics. In IFU mode, atmospheric dispersion correction can in principle be handled in post-processing, and the design therefore considers a removable ADC in VESPER observations to enhance sensitivity (Mahmoodzadeh et al., 8 Sep 2025, Saracco et al., 29 Jun 2026).

4. Science cases enabled by SHARP/VESPER

A large set of SHARP science papers uses VESPER as the spectroscopic workhorse for high-redshift galaxy evolution. For massive quiescent galaxies at $1.5 < z < 3$, one feasibility study reports that SHARP/VESPER will routinely measure stellar population gradients out to $2R_e$ for the majority of the population at $z < 2.5$ with integrations of about 20 h, and will reach at least $p_{\mathrm{exploit}} = 0.7$0 in about 30 h at $p_{\mathrm{exploit}} = 0.7$1. The same paper emphasizes 30 mas spatial sampling, MORFEO’s MCAO, and the ability to resolve the inner $p_{\mathrm{exploit}} = 0.7$2 kpc at all redshifts considered (Gargiulo et al., 30 Jun 2026).

For morpho-kinematics at cosmic noon, VESPER-SHARP is presented as a next-generation NIR IFS concept on the E-ELT, with the claim that at $p_{\mathrm{exploit}} = 0.7$3 it can provide more than 100 different spectra within $p_{\mathrm{exploit}} = 0.7$4 for a typical galaxy. That science case uses a 12-probe multi-IFU configuration with $p_{\mathrm{exploit}} = 0.7$5 selectors and derives exposure-time forecasts such as 5 h to reach $p_{\mathrm{exploit}} = 0.7$6 at $p_{\mathrm{exploit}} = 0.7$7 for an $p_{\mathrm{exploit}} = 0.7$8, $p_{\mathrm{exploit}} = 0.7$9 galaxy, and 30 h for an $\mu\mathrm{m}$0, $\mu\mathrm{m}$1 galaxy (Rigamonti et al., 29 Jun 2026).

For massive-black-hole-binary core scouring, SHARP-VESPER is the IFU used to reconstruct 2D stellar kinematics in galaxy nuclei. That study states that central scourings with sizes above $\mu\mathrm{m}$2 pc can in principle be detected up to reionization, that smaller cores of $\mu\mathrm{m}$3 pc can be detected up to $\mu\mathrm{m}$4, and that this encompasses a volume more than 40 times the one available at present. It also states that the search volume for pc-size cores around $\mu\mathrm{m}$5–$\mu\mathrm{m}$6 million solar mass MBHs can be enhanced by a factor 30 (Bortolas et al., 30 Jun 2026).

For AGN feedback at cosmic noon, one proposal targets eleven luminous, obscured AGN at $\mu\mathrm{m}$7 with resolved radio emission, using VESPER’s multi-Integral Field Selector capability to obtain resolved continuum and emission-line maps for at least 110 galaxies within 55 h of integration time. A related survey concept, SUNRISE-3D, frames SHARP/VESPER as the instrument that can construct spatially resolved outflow-property maps and resolved SFR maps across a representative sample at cosmic noon, with ETC simulations indicating that $\mu\mathrm{m}$8 hours per source are sufficient to reach $\mu\mathrm{m}$9 on the H$\mu\mathrm{m}$0 peak in the central pixel for a $\mu\mathrm{m}$1, $\mu\mathrm{m}$2 galaxy at the adopted flux level (Polletta et al., 30 Jun 2026, Vietri et al., 29 Jun 2026).

For quenching and bulge-disk growth in very massive galaxies, SHARP-VESPER is described as enabling, for the first time, a simultaneous bulge-disk decomposition of stellar populations and spatially resolved mapping of ionised gas in galaxies with $\mu\mathrm{m}$3 at $\mu\mathrm{m}$4. The quoted benchmark is 15 hr typical exposure time, yielding $\mu\mathrm{m}$5–20 per spectral resolution element on inner-bulge and outer-disk continuum spectra and $\mu\mathrm{m}$6 for nebular lines on sub-kpc scales (Mancini et al., 29 Jun 2026).

For the epoch of reionisation, SHARP/VESPER is proposed as the instrument that can map $\mu\mathrm{m}$7 Ly$\mu\mathrm{m}$8 emission down to structures of size $\mu\mathrm{m}$9 pc while simultaneously capturing large-scale structure up to 100 kpc. The paper’s ETC forecast states that a total exposure of 4 hours gives $R \sim 3000$0 at a Ly$R \sim 3000$1 surface-brightness level of $R \sim 3000$2 over 1 arcsec$R \sim 3000$3 (Bisogni et al., 30 Jun 2026).

5. Security- and privacy-oriented uses

In network security, Vesper is a host-based system for detecting Man-in-the-Middle (MitM) attacks inside wired LANs by treating the network path between two machines as if it were an acoustic space and probing it with ICMP echo requests. The system builds a fingerprint of the link from RTT time series and uses autoencoders to model normal behavior and detect deviations caused by ARP spoofing or transparent bridging. The paper reports TPR near 100%, FPR close to 0%, detection often within a few seconds, and overhead in the low-kbits/s range per monitored target (Mirsky et al., 2018).

In privacy-preserving graph learning, VESPER is an end-to-end GRL framework in the VFL setting built upon Perturbed Message Passing (PMP). The system is designed for financial risk management on transaction networks and provides edge-level differential privacy for message-passing GNNs. The framework analyzes GIN and GCN under perturbation, introduces truncated message passing to improve performance on sparse graphs, and uses Rényi DP with an Analytical Moments Accountant for privacy accounting. The reported empirical result is that VESPER can train high-performance GNN models over both sparse and dense graphs under reasonable privacy budgets (Wu et al., 2023).

6. Speech, multimodal learning, Earth-system modeling, and other technical uses

In speech technology, Vesper is an emotion-oriented speech representation model derived from WavLM Large and further pretrained for speech emotion recognition. The paper defines Vesper-4 with 4 Transformer layers and 63.5M parameters and Vesper-12 with 12 layers and 164.3M parameters, compared with WavLM Base at 12 layers and 94.7M parameters and WavLM Large at 24 layers and 316.6M parameters. The model uses emotion-guided masking, plus hierarchical and cross-layer self-supervision, and reports that Vesper-4 outperforms WavLM Base while Vesper-12 surpasses WavLM Large on the reported SER benchmarks (Chen et al., 2023).

A related but distinct name is ViSpeR, a multilingual audio-visual speech recognition system and dataset for Chinese, Spanish, English, Arabic, and French. It is a single multilingual encoder-decoder Transformer trained jointly across languages, with a 12-layer encoder, 6-layer decoder, and a 21k shared subword vocabulary. The paper reports, for example, French VSR/AVSR performance of 29.8 / 5.7 WER, Spanish 39.4 / 4.4 WER, Arabic 47.8 / 8.4 WER, Chinese 51.3 / 15.4 CER, and English 49.1 / 8.1 WER, and explicitly notes that the name is “easy to misread as ‘Vesper’” (Narayan et al., 2024).

In Earth-system modeling, VESPER stands for Verification of Earth-System ParametERisation. It is a deep-learning regression model trained to map ERA5 meteorology + surface physiography to MODIS land-surface temperature, and is used to assess whether updated physiographic datasets improve consistency with satellite observations. The paper reports that, for grid cells where lake fields have been updated, prediction accuracy improves by 0.37 K on average, that the subset where lakes have been exchanged for bare ground improves by 0.83 K, and that updates to glacier cover improve prediction accuracy by 0.22 K (Kimpson et al., 2022).

A further technical use appears in dense-urban sUAS benchmarking, where Vantage Robotics Vesper is one of the eight evaluated platforms. In that report, Vesper reaches a maximum BLOS distance of 700 m, executes GPS-based RTH from that distance with 1 m return-distance error, and achieves 0.14 m ± 0.20 m average error in Outdoor 3D Mapping Accuracy using Pix4Dmapper (Norton et al., 29 Jan 2025).