Few-shot audio classification requires audio-language models to recognize acoustic events from very limited labeled data. Real-world sounds vary widely, so models must generalize beyond the exact examples they have seen.

DF-PALM (Dynamic Fusion Prompt Learning) tackles this by building on PALM's formulation, which adapts encoded class representations without changing the pretrained encoders. The key idea is to keep the large audio and language encoders fixed and instead steer the class representations that connect them, concentrating the learning burden on the prompt side.

The approach is notable for its efficiency: because the pretrained encoders are untouched, DF-PALM avoids the cost and risk of full fine-tuning. The arXiv abstract does not include experimental numbers, so the method's practical performance relative to other few-shot audio classifiers is not yet reported in this source.