Google Research introduces Agile Deliberation for subjective image classification

Google Research has released Agile Deliberation, a human-in-the-loop framework for defining visual categories whose boundaries are subjective, such as “healthy food,” “inspiring artwork” or clickbait. Rather than requiring a complete definition upfront, the system helps users refine one by examining borderline images. The resulting text serves both as a readable policy and as the prompt for a vision-language classifier.

The framework uses Gemini 2.5 Pro and Gemini 2.5 Flash through the API, without fine-tuning. It breaks a concept into up to three units, retrieves related images using CLIP embeddings and FAISS, and presents groups that share a semantic ambiguity. Users label images and explain disagreements; a language model converts that feedback into candidate definitions, while the system retains the version with the best F1 score.

Google evaluated the approach in 18 live sessions involving nine participants and two concepts: “paid to play” clickbait and “healthy food.” Compared with zero-shot Gemini 2.5 Flash, average F1 increased by 10.5 points. Scores rose from 0.48 to 0.59 for paid to play and from 0.48 to 0.58 for healthy food. A demo notebook and code are available under Apache 2.0, with a Gemini API key required.