AstroVLBench: A Benchmark for Vision-Language Models in Observational Astronomy

AstroVLBench is a comprehensive benchmark for evaluating vision-language models (VLMs) on authentic astronomical observation tasks. It comprises over 4,100 expert-verified evaluation instances spanning five observational modalities central to extragalactic and time-domain research: optical imaging, radio interferometric imaging, multi-wavelength photometry, time-domain light curves, and optical spectroscopy.

The benchmark is uniquely centered around Active Galactic Nuclei (AGN) classification — a fundamental task in galaxy astronomy that naturally manifests across all electromagnetic wavebands — enabling clean, controlled comparison of AI capabilities across different modalities while minimizing confounding effects of varying domain knowledge requirements. All ground truth labels are rigorously established through spectroscopic confirmation and expert identification.

Paper: A systematic evaluation of vision-language models for observational astronomical reasoning tasks Anonymous, Anonymous, Anonymous, Anonymous

Tasks Overview

Task	Domain	Modality	Classes	Samples	Source Surveys
Task 1	QSO Host Galaxy	Optical imaging	2 (AGN, Galaxy)	557	HSC-SSP
Task 2	Radio Morphology	Radio imaging	2 (FRI, FRII)	833 × 2 surveys	MiraBest (FIRST + NVSS)
Task 3	SED Classification	Multi-wavelength photometry	3 (Type-1 AGN, Type-2 AGN, Galaxy)	168	AKARI NEP + HSC + Euclid + WISE
Task 4	Light Curve	Time-domain photometry	5 (AGN, SNIa, TDE, RRL, Mira)	142	PLAsTiCC
Task 5	Spectral Interpretation	Optical spectroscopy	3 questions	700	DESI DR1

Task Details

Task 1: QSO Host Galaxy Classification

Tests the ability to distinguish between unresolved point sources (quasars) and extended galaxy bulges. The key discriminant is whether the central light profile is dominated by a seeing-limited PSF or resolved stellar emission from the host galaxy.

• Data: 557 RGB composite images from HSC-SSP PDR3 (g, r, i bands, 10″×10″ cutouts)
• Labels: 300 spectroscopically confirmed Type-1 AGN, 257 inactive galaxies
• Sources: HSC-SSP (Aihara et al. 2017), AGN sample (Li et al. 2021), Galaxy sample (Simard et al. 2011)

Task 2: Radio Galaxy Morphology Classification

Evaluates morphological classification of radio galaxies into Fanaroff-Riley Type I (FRI, core-brightened) and Type II (FRII, edge-brightened), testing how spatial resolution affects classification.

• Data: 833 sources with matched images from two surveys:
  • MiraBest_F (FIRST survey, higher resolution): 606 images
  • MiraBest_N (NVSS survey, lower resolution): 833 images
• Labels: 397 FRI, 436 FRII (expert-verified, excluding hybrids)
• Sources: MiraBest (Porter et al. 2023), FIRST (Becker et al. 1995), NVSS (Condon et al. 1998)

Task 3: SED Classification

Tests the ability to diagnose source types from multi-wavelength photometry, requiring inference of relative contributions of stellar emission, unobscured accretion, and dust-reprocessed infrared emission.

• Data: 168 sources with SED plots (ν f_ν vs. wavelength, log-log scale) and 21-band photometry (HSC g/r/i/z/y, Euclid Y/J/H, AKARI N2–L24, WISE W1–W4)
• Labels: 44 Type-1 AGN, 40 Type-2 AGN, 84 Galaxy (redshift range 0.04 < z < 0.45)
• Sources: AKARI NEP (Kim et al. 2020), cross-matched with DESI DR1 AGN/QSO catalog; BPT classifications from Baldwin et al. 1981, Veilleux & Osterbrock 1987

Task 4: Light Curve Classification

Evaluates classification of five physically distinct classes spanning persistent variables and transient phenomena from multi-band light curves.

• Data: 142 objects with light curve plots and flux time-series across 6 LSST-like bands (u, g, r, i, z, y)
• Labels: 30 AGN, 30 SNIa, 30 TDE, 30 RRL, 22 Mira
• Sources: PLAsTiCC (The PLAsTiCC team 2019)

Task 5: Spectral Interpretation (ASIB)

Evaluates hierarchical spectral characterization mirroring the logic of astronomical spectral analysis: emission line detection → broadening assessment → line-ratio diagnostics.

• Data: 700 optical spectra organized into 7 sub-groups (A, B, C1–C4, D), 100 per group
• Questions:
  • Q1 (700 samples, all groups): Both H-alpha and H-beta emission lines present?
  • Q2 (500 samples, groups C1–C4, D): Is this a Broad-Line AGN (BLAGN)?
  • Q3 (400 samples, groups C1–C4): BPT classification (Star-Forming, Composite, Seyfert, LINER)
• Sources: DESI DR1 (DESI Collaboration 2025), emission-line measurements from FastSpecFit (Moustakas et al. 2023)

Evaluation Protocol

The benchmark supports three levels of evaluation:

1. Baseline performance across five modalities with guided prompts
2. Mechanistic ablations:
   • Perceptual vs. conceptual guidance: unguided / phenomenological / physical prompts
   • Visual vs. numerical input: image plots vs. structured magnitude/flux tables
   • Few-shot instruction: zero-shot vs. few-shot with visual exemplars
3. Reasoning quality analysis: evaluating whether correct predictions are accompanied by physically valid justifications

Directory Structure

AstroVLBench/
├── data/
│   ├── Task1_QSOHost/
│   │   ├── image_labels.csv              # Labels: image path, AGN/Galaxy
│   │   └── images/                       # 557 RGB galaxy images (JPG)
│   ├── Task2_RadioMorph/
│   │   ├── MiraBest_F/
│   │   │   ├── metadata.jsonl            # Labels: filename, FRI/FRII
│   │   │   └── images/{FRI,FRII}/        # 606 FIRST radio images (PNG)
│   │   └── MiraBest_N/
│   │       ├── metadata.jsonl
│   │       └── images/{FRI,FRII}/        # 833 NVSS radio images (PNG)
│   ├── Task3_SED/
│   │   ├── nirsed_v2_catalog.csv         # Catalog with photometry + labels
│   │   └── images/                       # 168 SED plot images (PNG)
│   ├── Task4_LightCurve/
│   │   ├── manifest.csv
│   │   ├── figures/{AGN,SNIa,TDE,RRL,Mira}/  # 142 light curve plots (PNG)
│   │   └── csv/{AGN,SNIa,TDE,RRL,Mira}/      # 142 flux time-series (CSV)
│   └── Task5_SpecType/
│       ├── ASIB_v1_selection_with_snr.csv     # Catalog with group labels
│       ├── figures/{Group_A,...,Group_D}/      # 700 spectrum plots (PNG)
│       └── spectra/                           # Spectrum CSV files
└── code/
    ├── task1/llm.py
    ├── task2/llm.py
    ├── task3/llm.py
    ├── task4/llm.py
    └── task5/llm.py

Inference Code

Each task includes a self-contained llm.py script supporting:

• Multiple VLMs via OpenAI-compatible API (GPT, Claude, Gemini, Grok, Qwen, InternVL)
• Multiple prompt strategies (--prompt-type guided/woguide/phenomenological)
• Multiple modalities where applicable (--modality image/text)
• Redshift ablation for Task 3 (--redshift-mode with/without)
• Resume capability (--resume)

Example Usage

# Task 1: Classify galaxy images with physical prompt
python code/task1/llm.py --model gpt-4o --prompt-type guided

# Task 2: Classify radio galaxies (FIRST survey)
python code/task2/llm.py --dataset MiraBest_F --model gemini-3-pro-preview

# Task 3: SED classification — image vs. text modality
python code/task3/llm.py --model gpt-4o --modality image --redshift-mode with
python code/task3/llm.py --model gpt-4o --modality text --redshift-mode with

# Task 4: Light curve classification
python code/task4/llm.py --model gpt-4o --modality image

# Task 5: Spectral interpretation (three questions)
python code/task5/llm.py --question Q1 --model gemini-3-pro-preview
python code/task5/llm.py --question Q2 --model gemini-3-pro-preview
python code/task5/llm.py --question Q3 --model gemini-3-pro-preview

Environment Variables

All API credentials are configured via environment variables (no hardcoded URLs):

OPENAI_API_KEY=...        # OpenAI API key
OPENAI_BASE_URL=...       # OpenAI-compatible base URL
CLAUDE_API_KEY=...        # Anthropic Claude API key
GROK_API_KEY=...          # xAI Grok API key
QWEN_API_KEY=...          # Alibaba Qwen API key
QWEN_BASE_URL=...         # Qwen API base URL
INTERN_API_KEY=...        # InternVL API key
INTERN_BASE_URL=...       # InternVL API base URL

Citation

@article{anonymous2025astrovlbench,
  title={A systematic evaluation of vision-language models for observational astronomical reasoning tasks},
  author={Anonymous and Anonymous and Anonymous and Anonymous},
  year={2025},
  url={https://huggingface.co/datasets/anonymous4ai/AstroVLBench}
}

License

This dataset is released under the CC BY 4.0 license. The underlying astronomical data are derived from publicly available surveys: HSC-SSP, MiraBest, AKARI NEP, PLAsTiCC, and DESI DR1.