DeCoDe turns off-the-shelf MLLMs into powerful few-shot classifiers

A new paper from researchers introduces DeCoDe (Decompose, Compare, and Decide), a simple technique that repurposes off-the-shelf Multimodal Large Language Models (MLLMs) for few-shot image classification without any additional training. The method reframes few-shot classification as a set of pairwise binary decisions: given a query image and a support image from a candidate class, the MLLM decides whether they share the same class. The logit of an affirmative response is treated as a similarity score, and the query is assigned to the class with the highest score across all support images.

DeCoDe achieves strong results even with basic prompts, but adding high-level context—such as the data domain (e.g., medical, satellite)—further boosts performance. The authors test DeCoDe on twelve datasets: six standard benchmarks and six newly curated few-shot benchmarks spanning diverse domains. The results show that this simple decomposition technique turns off-the-shelf MLLMs into powerful few-shot learners, significantly outperforming current state-of-the-art few-shot methods on both standard and novel domains. The code is released publicly.