Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector

This study investigates whether general-purpose Vision-Language Models (VLMs)—specifically the Gemma 4 2B and 4B models—can identify Fast Radio Bursts (FRBs) in radio dynamic spectra without any task-specific training. Traditional FRB detection relies on specialized deep learning models that require large, labeled datasets and are computationally expensive to retrain. The authors test a zero-shot, prompt-only regime where the model is given an image of a dynamic spectrum and a natural language instruction to classify the signal as an FRB, Radio Frequency Interference (RFI), or noise.

The researchers evaluated the models against the state-of-the-art specialized detector, SwinYNet, using a benchmark of 2,000 simulated spectra. The Gemma 4 2B model achieved 93.65% accuracy, which is statistically comparable to SwinYNet's 92.90%. Notably, the VLM demonstrated a superior ability to reject structured RFI, with a false-positive rate of 6.4% compared to 25.0% for the specialized model. While SwinYNet remains superior in probabilistic ranking (ROC-AUC of 1.0000 vs. 0.9482), the VLM's performance is remarkable given that it was never trained on astronomical data. Furthermore, the models could be reconfigured for three-class classification (FRB/RFI/noise) simply by changing the prompt, achieving up to 86% accuracy without false FRB detections.

This work demonstrates that generalist foundation models possess latent capabilities for complex scientific pattern recognition. By using small, locally executable VLMs, researchers could potentially bypass the need for massive, task-specific training pipelines. This approach offers a flexible, low-cost alternative for real-time transient detection, providing not just a classification, but also natural-language justifications for each decision, which could aid in human verification and pipeline transparency.