EyeVLA: EVLA Upgrade & Vision-Language Systems

• EyeVLA is a multidomain term referring to an upgraded radio astronomy array for high-resolution cosmic studies and a unified vision-language action model for embodied perception.
• The EVLA upgrade delivers significant improvements in sensitivity, frequency coverage, and correlator performance, enabling breakthroughs in transient surveys and galaxy mapping.
• The vision-language system integrates a robotic eyeball with reinforcement learning and fine-tuning strategies to jointly process visual, linguistic, and action tokens for optimized scene understanding.

EyeVLA refers to distinct but technically ambitious systems in both radio astronomy instrumentation and computer vision for embodied perception. In radio astronomy, EyeVLA is an alternative pronunciation and abbreviation for the Expanded Very Large Array (EVLA), a comprehensive modernization of the Karl G. Jansky Very Large Array focused on advancing centimeter-wave radio science. Separately, in computer vision, EyeVLA designates a unified Vision-Language-Action model with an actively controlled robotic “eyeball” for instruction-driven scene understanding, and a fine-tuned vision-LLM for radio source analysis tasks. The following sections detail these distinct but technologically focused instantiations.

1. Expanded Very Large Array (EVLA, “EyeVLA”): Definition and Motivation

The Expanded Very Large Array (EVLA, “EyeVLA”) is a large-scale upgrade of the original VLA centered on a network of 27 radio dishes at the Karl G. Jansky Very Large Array in New Mexico. Initiated to overcome the three-decade stagnation of the VLA in sensitivity, continuity of frequency coverage, and velocity resolution, the EVLA delivers comprehensive, end-to-end improvements in sensitivity, spectral flexibility, digital signal transport, and correlator performance (Dougherty et al., 2010, Perley et al., 2011). The main scientific drivers for this modernization are high-precision studies of cosmic magnetism, deeply embedded star formation, wide-field surveys of transient phenomena, and time-resolved mapping of galaxy evolution.

2. Hardware Architecture and Performance Specifications

Receiver and Signal Chain Upgrades

The EVLA retrofitted each antenna with eight cryogenic receiver bands (L, S, C, X, Ku, K, Ka, Q), enabling continuous coverage from 1 to 50 GHz with up to 8 GHz instantaneous bandwidth per polarization. Baseband digitizers at each antenna digitize four 2-GHz-wide intermediate frequency pairs per polarization (totaling up to 16 GHz), and a fiber-optic transmission system routes these signals to the central correlator, eliminating analog calibration instabilities.

WIDAR Correlator

The WIDAR (Wide-band Interferometric Digital ARchitecture) correlator processes the full 16 GHz from all antennas simultaneously, providing a minimum of 16,384 spectral channels per baseline (expandable to over 4 million), flexible sub-band definition, and full-Stokes polarization products. Forty-eight independently tunable sub-band pairs can each be assigned arbitrary frequency range and channelization, supporting concurrent continuum and high-resolution spectral line observations. Specialized modes, including phased-array beamforming, pulsar binning, and real-time RFI excision, are integral (Dougherty et al., 2010, Perley et al., 2011).

Parameter	VLA (Legacy)	EVLA (EyeVLA)	Improvement
Continuum rms	10 μJy/beam	1 μJy/beam	×10
Max. BW/pol	0.1 GHz	8 GHz	×80
Channels/Baseline	16	16,384	×1,024
Max Channels	512	4,194,304	×8,192
Freq. Coverage	22% (per band)	100%	×5

3. Sensitivity, Configuration, and Early Science

The rms point-source sensitivity is governed by the radiometer equation for interferometers:

$$\sigma = \frac{\mathrm{SEFD}}{\eta_s \sqrt{N(N-1)\Delta\nu\tau}}$$

where $\mathrm{SEFD}$ is the system equivalent flux density, $\eta_s$ is system efficiency, $N$ is the number of antennas, $\Delta\nu$ the bandwidth, and $\tau$ the integration time. By increasing processed bandwidth and reducing system noise, base EVLA sensitivity reaches 1 μJy/beam continuum rms in 9 hours, a tenfold improvement over the original VLA (Perley et al., 2011).

The array retains four reconfigurable baselines (A: 36 km – D: 1 km), with a diffraction-limited beam size as small as 40 mas at 50 GHz.

Operational access modes include:

• OSRO (Observatory Shared-Risk Observing): Remote access to two 128 MHz basebands, broadening as commissioning continued.
• RSRO (Resident Shared-Risk Observing): Full access to 64 sub-bands, 8 GHz per polarization, and the complete correlator feature set in exchange for on-site commissioning contributions.

Early science highlights include ultra-deep C-band continuum mosaics, wide-band spectral index mapping, high-fidelity ammonia imaging, pulsar gating, and polarimetric rotation measure synthesis, demonstrating the broadened discovery space enabled by the EVLA (Dougherty et al., 2010, Perley et al., 2011).

4. EyeVLA in Embodied Vision-Language-Action (VLA) Systems

In the context of robotic active perception, EyeVLA designates an integrated system wherein a robotic eyeball (pan-tilt gimbal with zoomable camera) is controlled by a unified Vision-Language-Action model (Yang et al., 19 Nov 2025). The architecture features:

• Robotic Eyeball: Two-axis pan-tilt and motorized zoom.
• Vision-Language Backbone: A frozen Vision Transformer (ViT) encoder paired with a transformer LLM (e.g., Qwen2.5-VL).
• Action Tokenization: Discretization of pan/tilt/zoom into hierarchical tokens using a canonical $\{5,2,1\}$ coin basis for efficient sequence modeling.
• Reinforcement Learning: Group Relative Policy Optimization (GRPO) to refine viewpoint selection, driven by composite rewards (localization IoU, action fidelity).

The model jointly predicts visual, linguistic, and camera-control tokens in a single autoregressive sequence, integrating 2D bounding box feedback into both reasoning and reward shaping. This enables the agent to actively select optimal viewpoints for task-driven perception under spatial and pixel constraints. Experimental results demonstrate that RL-enhanced EyeVLA achieves mean absolute errors of 2.04° (pan), 1.68° (tilt), and 65.4 zoom units, with a 96% task completion rate—substantially outperforming baseline or fixed-camera approaches under realistic embodied settings (Yang et al., 19 Nov 2025).

5. EyeVLA as a Vision-LLM for Radio Astronomy

Within the domain of radio astronomical source analysis, EyeVLA denotes a fine-tuned small-scale VLM based on the LLaVA-OneVision 7B model, employing a frozen SigLIP-so400m-patch14-384 ViT encoder and a 7B-parameter Qwen2 LLM (Riggi et al., 31 Mar 2025). Two training regimes are supported:

• Full Model Fine-Tuning: Updates both the LLM and its MLP adapter.
• Low-Rank Adaptation (LoRA): Only small, rank-64 matrices are learned per layer ($\Delta W = AB$ with scaling $\alpha=128$), preserving upstream knowledge and efficiency.

The fine-tuning data comprises 59,331 radio images (standardized, augmented) and 38,545 arXiv-derived figure-caption pairs (filtered for quality). Instruction tuning is performed using synthesized QA dialogs and multi-task loss combining classification, alignment, and instruction compliance: $$L_\mathrm{total} = \lambda_1 L_\mathrm{class} + \lambda_2 L_\mathrm{align} + \lambda_3 L_\mathrm{instr}$$, with relative weights $(1,0.5,0.5)$.

Task	F1 Gain (full FT)	F1 Recovery (LoRA+caps)	General Benchmarks Drop
Extended source detection	+30 pts	+10 pts (relative)	−20 pts
Morphology/artifact/peculiarity	+10 pts	+10 pts (relative)	−20 pts

On six radio-specific benchmarks (extended/diffuse source detection, morphology classification, radio-galaxy detection, artifact detection, peculiarity classification, FR-I/FR-II), full fine-tuning yields up to 30 percentage-point F1 improvements over the base model but incurs a 20-point accuracy drop on standard multimodal tasks, mitigated by LoRA and caption data integration. Noted limitations are due to misalignment between frozen ViT visual features and language decoding, and catastrophic forgetting under deep fine-tuning. Recommendations include contrastive pre-training, more consistent annotations, hybrid/hierarchical updating, and further dataset scaling (Riggi et al., 31 Mar 2025).

6. Technical and Operational Limitations

For the EVLA (“EyeVLA”), bottlenecks and error modes include the massive data output rate (necessitating efficient on-the-fly processing and data pipelining), calibration instabilities mitigated by digitization and fiber transport, and the challenge of full exploitation of spectral flexibility.

In active perception EyeVLA systems, limitations stem from the granularity of tokenized action space, the dependency on robust bounding-box supervision, and data scarcity for long-tailed embodied reasoning. For the vision-language EyeVLA models in radio astronomy, limitations include visual-textual misalignments and catastrophic forgetting, particularly when heavily fine-tuning all layers.

7. Implications, Extensions, and Future Directions

For the EVLA (“EyeVLA”), anticipated future developments focus on leveraging the dramatic improvements in sensitivity, bandwidth, and spectral agility for next-generation transient surveys, polarimetric tomography, and deep interferometric mapping across multiple astronomical domains (Dougherty et al., 2010, Perley et al., 2011).

Vision-Language-Action EyeVLA points toward more deeply integrated, resource-constrained embodied AI agents capable of instruction-following and efficient environmental perception in open-world domains (Yang et al., 19 Nov 2025). For vision-language assistants in radio analysis, future progress hinges on improved cross-modal alignment, higher-quality and more consistent annotation, and methodological innovation in adaptation strategies (contrastive pretraining, model merging, hybrid LoRA). A plausible implication is that combining these approaches enables robust AI assistants bridging task-critical, domain-specific visual-linguistic analyses with broader multimodal reasoning capabilities (Riggi et al., 31 Mar 2025).