Google Research Launches Agile Deliberation After 18 Live Sessions Showed Stronger Subjective Image Classifiers
Updated
Updated · Google Research · Nov 4
Google Research Launches Agile Deliberation After 18 Live Sessions Showed Stronger Subjective Image Classifiers
1 articles · Updated · Google Research · Nov 4
Summary
Google Research unveiled Agile Deliberation, a framework that helps users build image classifiers for subjective categories by iteratively defining boundaries and resolving ambiguous edge cases.
18 live sessions lasting 1.5 hours each found the system outperformed standard baselines and improved labeling consistency, especially on hard borderline images where generic classifiers often fail.
The workflow splits the task into concept scoping and concept iteration, using vision-language models to surface semantically borderline images and an LLM-driven loop to refine prompts automatically.
Participants reported lower effort, stress and mental fatigue than with manual prompt engineering, and every user preferred Agile Deliberation over manual methods.
Google also open-sourced the code and algorithms on GitHub, positioning the tool for uses such as content moderation, curation and other human-in-the-loop vision tasks.