AstraQ-VL: Parameter-Efficient Astronomy Vision-Language Modeling with Connector Alignment and LoRA Tuning

AstraQ-VL is a compact LLaVA-style astronomy vision-language model built with a frozen CLIP ViT-L/14 vision encoder, a two-layer MLP connector, and Qwen2.5-1.5B-Instruct. The paper reports a two-stage parameter-efficient training recipe: Stage 1 trains only the connector, while Stage 2 warm-starts the connector and adds LoRA adapters in the language model.

The work evaluates the model on a per-image held-out split of astronomy caption and question-answer records using automatic reference-based, embedding, NLI, and astronomy-specific unsupported-detail proxies. The reported results are interpreted as improvements in in-domain reference alignment, not as a general factual-reliability claim.