Google Research Releases Agile Deliberation Framework for Human-AI Concept Definition
Google Research has open-sourced Agile Deliberation, a framework that lets users define fuzzy 'gray area' concepts before an AI classifier learns them. The system splits a concept into in-scope and out-of-scope subconcepts, surfaces borderline samples for user labeling, and has Gemini rewrite the definition as a VLM prompt. Across 18 live sessions it beat zero-shot Gemini 2.5 Flash by 10.5 F1 points on average, with code released under Apache 2.0.
Related Entities
Full Report
▲ Marktechpost AI · 2026-11-05 05
▲ Google Research just released Agile Deliberation, a framework that helps people define fuzzy "gray area" concepts before an AI classifier learns them.
The key idea is simple. Users rarely know a subjective concept's boundary upfront. So the system first splits a concept like "healthy food" into in-scope and out-of-scope subconcepts. Then it surfaces images that are borderline in meaning, the user labels them, and Gemini rewrites the definition that serves as the VLM prompt.
In 18 live 90-minute sessions, F1 rose to 0.59 on "paid to play" clickbait and 0.58 on healthy food. That is 10.5 points above zero-shot Gemini 2.5 Flash on average, and ahead of Modeling Collaborator (0.53 and 0.50).
Code is open under Apache 2.0, with a Colab/Jupyter demo built on Gemini 2.5 Pro, Gemini 2.5 Flash and a FAISS index......
Full analysis:
