Google Research 发布 Agile Deliberation 框架,人机协作定义模糊概念
Google Research 开源 Agile Deliberation 框架,让用户在 AI 分类器学习之前先界定模糊的“灰色地带”概念。系统将概念拆分为范围内与范围外子概念,再展示边界样本供用户标注,由 Gemini 重写定义作为 VLM 提示词。18 场实测中 F1 平均较零样本 Gemini 2.5 Flash 高 10.5 分,代码以 Apache 2.0 开源。
报道正文
▲ 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:
