2026-10-27 · ArXiv
arXiv:2604.24589v2 Announce Type: replace
Abstract: Vision-language models (VLMs) are increasingly proposed as general-purpose tools for scientific data interpretation, yet their reliability on real astronomical observations across diverse modalities remains untested. We present AstroVLBench, a benchmark of over 4,100 expert-verified instances across five tasks spanning optical imaging, radio interferometry, multi-wavelength photometry, time-domain light curves, and optical spectroscopy. We systematically evaluate six state-of-the-art frontier models and find that performance is strongly modality-dependent and uncorrelated with general benchmark rankings. Mechanistic cross-model ablations reveal three concrete bottlenecks. First, prompt grounding exhibits marked model and task dependence: physical guidance yields more balanced classifications on optical imaging, whereas phenomenological prompts better navigate undersampled time-domain data. Second, presenting one-dimensional measurements as numerical tables rather than rendered plots consistently improves light-curve classification across all models by up to 23 percentage points, demonstrating that plot rendering can obscure physically relevant signal. Third, an expert audit of optical spectroscopy reveals a notable "right-answer-wrong-reason" phenomenon: across evaluated models, an average of 26% of correct predictions are accompanied by physically inconsistent or invalid reasoning chains, showing that accuracy alone is insufficient for trustworthy scientific deployment. AstroVLBench provides the first systematic, multi-modal baselines for VLMs in observational astronomy and identifies the representation, grounding, and reasoning bottlenecks that must be addressed before VLMs can be trusted as autonomous scientific interpreters.
原文链接：https://arxiv.org/abs/2604.24589
查看更多资讯请见 AI 行业资讯。