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 that combines a frozen CLIP ViT-L/14 vision encoder, a two-layer MLP connector, and Qwen2.5-1.5B-Instruct. Stage 1 trains only the connector, while Stage 2 warm-starts the connector and adds LoRA adapters to the language model.

The paper compares the two stages on an untouched image-disjoint internal test set and evaluates their external behavior through zero-shot prompted closed-set classification on AstroVLBench Tasks 1–2 and single-reference solar-image captioning on DeepSDO. Stage 2 improves automatic reference alignment on the internal test set and the selected AstroVLBench aggregate, but its external behavior is task-dependent: it regresses on FIRST and does not improve the DeepSDO overlap measures. All reported outcomes are automatic measurements and do not establish scientific correctness or factual reliability.