HELIOS: Diverse Research Systems

• HELIOS is a recurring research designation that spans varied fields such as quantum processors, LiDAR simulations, exoplanetary radiative transfer, and smart-energy systems.
• It integrates advanced methodologies including GPU-initiated asynchronous disk IO, Bayesian inversion, and adaptive model selection to improve performance and efficiency.
• Distinct acronym expansions and context-specific frameworks disambiguate HELIOS, enabling targeted applications from high-fidelity simulation to low-power perceptual systems and dark matter detection.

HELIOS is a recurrent research designation used for multiple, technically unrelated systems, instruments, simulators, and analytical frameworks. In the cited literature, the name appears in computer systems for terabyte-scale graph neural network training (Sun et al., 2023), trapped-ion quantum computing (Ransford et al., 7 Nov 2025), exoplanetary radiative transfer and retrieval (Malik et al., 2016, Lavie et al., 2016), remote-sensing simulation (Winiwarter et al., 2021), robotics and LiDAR perception (Jung et al., 31 Jan 2025, Ashton et al., 26 Sep 2025), smart-energy language modeling (Jiang et al., 22 Dec 2025), and several other domains. The shared label therefore denotes a family of names rather than a single technical lineage.

1. Scope, capitalization, and acronymic variation

The literature uses several orthographic variants: Helios, HELIOS, HeLiOS, HELIOS++, and HELIOS-Retrieval. Some papers expand the name explicitly. In optics, HELIOS denotes HomogEneous and Layered medIa Optical Scattering (Mavrikakis et al., 26 Feb 2026). In ultralight dark-matter detection, HeLIOS denotes the Helium ultraLIght dark matter Optomechanical Sensor (Hirschel et al., 2023). In technology assessment, HELIOS denotes Hybrid Evaluation of Lifecycle and Impact of Outstanding Science (Garbayo, 22 Aug 2025). In heterogeneous LiDAR place recognition, HeLiOS is glossed as Heterogeneous LiDAR Overlap-based Spherical (Jung et al., 31 Jan 2025).

The same label is also used for domain-specific successor systems or specialized modules. HELIOS++ is a C++ re-implementation and extension of the original Java-based HELIOS laser-scanning simulator (Winiwarter et al., 2021). HELIOS-Retrieval is an open-source retrieval code integrated into the Exoclimes Simulation Platform and built around a forward model, chemistry module, and nested sampling (Lavie et al., 2016). This suggests that, in practice, the surrounding acronym expansion or research area is necessary to disambiguate the term.

2. Computer systems, machine learning, and code generation

In systems research, Helios denotes an out-of-core GNN training system designed for terabyte-scale graphs on a single machine (Sun et al., 2023). Its architecture combines a GPU-initiated asynchronous disk IO stack, a GPU-managed heterogeneous cache, and a deep GNN-aware pipeline. The IO stack allows the GPU to directly access graph data on SSD, requires about 30% GPU cores to reach almost maximal disk IO throughput, and avoids wasting GPU cores between IO submission and completion. The reported outcome is training throughput comparable with in-memory systems, including performance up to 6.43x over state-of-the-art GPU-managed baselines and over 182x over CPU-managed baselines on terabyte-scale graphs (Sun et al., 2023).

A different HELIOS addresses efficient LLM inference serving through adaptive model and early-exit selection (Kumar et al., 14 Apr 2025). Its workflow shortlists candidate LLMs, evaluates them on a small subset of prompts, uses early-exit telemetry to load only a limited number of layers, and periodically reassesses whether another model can satisfy the current service-level objective more efficiently. The reported gains are 1.48× throughput, 1.10× energy-efficiency, 1.39× lower response time, and 3.7× improvements in inference batch sizes relative to the baseline when optimizing for the respective objectives (Kumar et al., 14 Apr 2025).

Helios is also the name of a heterogeneity-aware federated learning framework for the straggler issue (Xu et al., 2019). It identifies device training capability and applies soft-training through rotating neuron training so that straggling devices train dynamically compressed models while preserving collaboration. The framework distinguishes hardware heterogeneity from information heterogeneity, re-weights aggregation by the fraction of active neurons, and reports up to 2.5× speed-up together with accuracy gains over the listed baselines in IID and Non-IID settings (Xu et al., 2019).

In program analysis, HELIOS reframes LLM-based decompilation as a structured reasoning problem (Achamyeleh et al., 21 Jan 2026). The prompt structure exposes FUNCTION_CONTEXT, CFG_OVERVIEW, BLOCK_DETAILS, and RAW_DECOMPILED_CODE, optionally followed by compiler diagnostics. On HumanEval-Decompile for x86_64, the framework raises average object-file compilability from 45.0% to 85.2% for Gemini 2.0 and from 71.4% to 89.6% for GPT-4.1 Mini; with compiler feedback, compilability exceeds 94% (Achamyeleh et al., 21 Jan 2026). A separate generative-model usage appears in long-video synthesis, where Helios is a 14B autoregressive diffusion model that runs at 19.5 FPS on a single NVIDIA H100 GPU and supports minute-scale generation without relying on KV-cache, sparse or linear attention, or quantization (Yuan et al., 4 Mar 2026).

3. Remote sensing, robotics, perception, and wearable interaction

The original HELIOS and its successor HELIOS++ belong to topographic 3D laser-scanning simulation (Winiwarter et al., 2021). HELIOS was a Java-based ray-tracing LiDAR simulator supporting ALS, TLS, MLS, and ULS, whereas HELIOS++ is a C++ framework with modular platform, scene, and scanner components defined in XML, a survey object, support for mixed scene parts, and Python bindings via pyhelios. The simulation pipeline includes kD-tree acceleration, beam divergence modeled as a 2D Gaussian power distribution discretized into sub-rays, and optional full-waveform generation. Benchmarks reported reduced runtimes by up to 83% and memory requirements reduced by up to 94%, with HELIOS++ being up to 5.8× faster while using as little as 6% of the original memory (Winiwarter et al., 2021).

In LiDAR place recognition, HeLiOS is a deep network for heterogeneous sensors with different fields of view and scanning patterns (Jung et al., 31 Jan 2025). Its pipeline combines sparse 3D convolutions, local spherical transformer blocks, a dual-headed aggregation module, overlap-based data mining, and an overlap-guided triplet loss with adaptive margins. The reported heterogeneous-place-recognition result on HeLiPR Roundabout is AR@1 up to 0.912 versus 0.870 for MinkLoc3Dv2, with additional long-term validation showing near-perfect retrieval across a four-year gap (Jung et al., 31 Jan 2025).

In embodied AI, HELIOS is a hierarchical representation and search objective for language-specified pick-and-place mobile manipulation in partially observed scenes (Ashton et al., 26 Sep 2025). It maintains 2D occupancy and semantic value maps together with sparse 3D Gaussian object models carrying Dirichlet posteriors over semantic classes, and it optimizes an exploration–exploitation objective over frontiers and candidate objects. On the OVMM benchmark, the reported state-of-the-art result is FindObj 42.3%, Pick 30.5%, FindRec 18.6%, Place 6.3%, and SR 3.2%, with zero-shot transfer to a real-world office environment on a Spot robot (Ashton et al., 26 Sep 2025).

Another Helios targets always-on smart eyewear through event-based gesture recognition (Bhattacharyya et al., 2024). The hardware uses a 3 mm × 4 mm / 20 mW event camera, and the CNN pipeline on an NXP Nano UltraLite consumes less than 350 mW. The system recognizes seven classes of gestures with 91% accuracy and reports 60 ms real-time latency across 20 users (Bhattacharyya et al., 2024). The work therefore places the HELIOS label on both high-fidelity simulation and low-power interactive perception.

4. Quantum processors, spacecraft observations, and dark-matter detection

Quantinuum Helios is a 98-qubit trapped-ion quantum processor based on the quantum charge-coupled device (QCCD) architecture (Ransford et al., 7 Nov 2025). It uses ^{137}\mathrm{Ba}^{+} hyperfine qubits, a rotatable ion storage ring, two quantum operation regions connected by a junction, and a runtime with real-time compilation of dynamic programs. Averaged over all operational zones, the reported average infidelities are $2.5(1)\times10^{-5}$ for single-qubit gates, $7.9(2)\times10^{-4}$ for two-qubit gates, and $4.8(6)\times10^{-4}$ for state preparation and measurement. Random circuit sampling is used to argue that the processor operates beyond the reach of classical simulation (Ransford et al., 7 Nov 2025).

The name Helios also refers to the historical Helios spacecraft and the scientific reuse of its data. A re-analysis of Helios dust measurements identified a cluster of seven impacts at true anomaly $135^\circ \pm 1^\circ$, and model comparison found that four detected particles are compatible with cometary trails from 45P/Honda-Mrkos-Pajduvsakova and 72P/Denning-Fujikawa (Krüger et al., 2020). The inferred trail density is about $10^{-8}$ to $10^{-7}\,\mathrm{m}^{-3}$. In a separate Helios study, full inversion of 15 solar relativistic electron events measured by the E6 instrument yielded radial mean free paths ranging from 0.02 AU to 0.27 AU, with five short injections and ten long-lasting injections (Pacheco et al., 2019).

In dark-matter instrumentation, HeLIOS uses superfluid ^4\mathrm{He} acoustic modes and a superconducting re-entrant microwave cavity to search for ultralight bosonic dark matter (Hirschel et al., 2023). The detector exploits high mechanical quality factors for resonant amplification and uses helium pressurization to tune the acoustic resonance, with measured tuning of $\Delta f/f_0 \approx 3.35\%/\mathrm{bar}$ and up to 54% total shift at solidification (Hirschel et al., 2023). The cited outlook is that future generations could explore unconstrained scalar and vector ultralight-DM parameter space after an hour of integration time (Hirschel et al., 2023).

5. Exoplanetary atmospheres, radiative transfer, and electromagnetic scattering

In exoplanet science, HELIOS is an open-source, GPU-accelerated radiative-transfer code for self-consistent one-dimensional, plane-parallel atmospheres (Malik et al., 2016). It solves the transfer equation in the two-stream approximation with non-isotropic scattering, uses HELIOS-K for opacity computation, combines opacities with a correlated-$k$ approximation, and adopts analytical equilibrium chemistry. Reported runtime is 2–15 minutes on a laptop NVidia GeForce 750M GPU and 0.5–4 minutes on a Tesla K20 for a self-consistent model with 101 layers and 300 wavelength bins; a 3000-bin post-processed spectrum takes <30 s (Malik et al., 2016).

Subsequent HELIOS work revised the hemispheric two-stream formalism, broadened the opacity treatment, and assessed path-length and scattering corrections (Malik et al., 2019). The code intercompares with models such as COOLTLUSTY and PHOENIX, and the study reports that the scattering correction improves the accuracy of reflected stellar light by ~10%. It also emphasizes that metal oxides, hydrides, alkali atoms, and ionized hydrogen can drive non-monotonic temperature-pressure profiles in very hot atmospheres, and it publishes a grid of cloud-free radiative-convective models for self-luminous planets (Malik et al., 2019).

HELIOS-Retrieval extends the name into Bayesian atmospheric inversion (Lavie et al., 2016). Its forward model uses an exact solution of the radiative-transfer equation in the pure-absorption limit, considers both unconstrained and equilibrium chemistry, and applies nested sampling to compare models through Bayesian evidence. In the HR 8799 system, chemical equilibrium is reported as disfavored by the Bayesian evidence for HR 8799b, c and d, and the retrieved abundance patterns are interpreted as inconsistent with gravitational instability without late-time accretion and consistent with a core-accretion scenario with differential ice accretion (Lavie et al., 2016).

A separate HELIOS in computational electromagnetics is an open-source surface integral equation software for light scattering in homogeneous, periodic, and stratified environments (Mavrikakis et al., 26 Feb 2026). It implements the PMCHWT formulation, discretizes boundaries with triangular meshes and RWG basis functions, uses Ewald's transformation for 2D lattices, and applies a matrix-friendly layered-media Green's tensor with a tabulation–interpolation scheme. The software architecture combines a C++ core with a Python interface that manages preparation, solver execution, post-processing, and visualization (Mavrikakis et al., 26 Feb 2026).

6. Energy systems, fusion design, and technology assessment

In smart-energy AI, Helios is a domain-specialized LLM initialized from Qwen-2.5-7B and trained with the EnerSys multi-agent data-construction framework (Jiang et al., 22 Dec 2025). EnerSys produces EnerBase, EnerInstruct, and EnerReinforce, while EnerBench evaluates objective and subjective performance. The reported benchmark highlights state that Helios reaches 79.09% average on objective tasks, substantially outperforming similar-size general models and approaching GPT-4 in many cases, while also improving performance on explanation, question answering, and energy system modeling (Jiang et al., 22 Dec 2025).

Helios is also the name of a preconceptual stellarator fusion power-plant design based on a planar coil stellarator architecture (Swanson et al., 8 Dec 2025). The design uses a two-field-period quasi-axisymmetric stellarator with aspect ratio 4.5, 12 large plasma-encircling coils, and 324 smaller field-shaping coils. A maximum of 20 T on-coil is enforced, the minimum plasma-to-coil separation is 1.2 m, and the plant is designed to produce 1.1 GW of thermal power and 390 MW of net electric power. The maintenance concept removes entire toroidal sectors between the encircling coils, with a biennial outage of approximately 84 days and an estimated 88% capacity factor (Swanson et al., 8 Dec 2025).

In innovation studies, HELIOS v2.0 is a dynamic framework for technological maturity forecasting (Garbayo, 22 Aug 2025). It ingests five indicators—Investment, Publications, Patents, Adoption, and Regulation—then applies nonlinear sigmoid normalization, S-curve forecasting through Logistic, Gompertz, or Bass models, lifecycle-dependent dynamic weights, nonlinear aggregation through the Choquet integral or OWA, and Monte Carlo uncertainty quantification. The framework is explicitly designed to provide probabilistic forecasts, identify inflection points, and support strategic planning, R&D investment evaluation, and policy-making (Garbayo, 22 Aug 2025). Taken together with the smart-energy LLM and the stellarator design, this usage places HELIOS not only in computational and experimental systems, but also in sector-specific reasoning and decision-support infrastructure.