Large Language Models Enable Textual Interpretation of Image-Based Astronomical Transient Classifications

Modern astronomical surveys deliver immense volumes of transient detections, yet distinguishing between real astrophysical signals (e.g., explosive events, variable stars) and bogus imaging artifacts remains challenging. Convolutional neural networks (CNNs) are effective for such real-bogus classification in optical imaging data; however, their reliance on latent representations makes it difficult to discern the underlying physical reasoning behind each classification. Here, we show that large language models (LLMs) achieve accuracy comparable to CNNs on three major optical transient survey datasets (Pan-STARRS, MeerLICHT, and ATLAS) while simultaneously providing direct, human-readable descriptions for every transient. Using only 15 examples and a concise set of instructions, Google's LLM, Gemini, achieves a 93\% average accuracy across these datasets which have quite diverse resolution and pixel scales. This is the first demonstration of a successful application of an LLM to imaging data from optical transient surveys and it eliminates the need for extensive and complex labeled sets. Furthermore, we demonstrate that a second LLM can evaluate the coherence of the first LLM classifications, thus guiding iterative improvements by indicating problematic examples. This opens up new possibilities: rather than laboriously training a network from scratch, one can simply define the desired output characteristics and rely on the LLM to deliver them. Furthermore, by generating textual descriptions of observed features, LLMs enable users to query classifications as if navigating an annotated catalog, rather than deciphering abstract latent spaces. As next-generation telescopes and surveys further increase data streams, LLM-based classification could help bridge the gap between automated detection and transparent, human-level understanding.