Understand this story
On November 4, 2026, Google Research proposed the Agile Deliberation framework, which transforms the construction of subjective image classifiers into an iterative dialogue between humans and large language models: users first break down subcategories and exclusions, then the vision-language model proactively selects semantically boundary-case images, and the feedback enters an automatic prompt optimization loop. According to Google Research, the team validated this process in 18 real-human experiments of 1.5 hours each, stating that it reduced user fatigue and anxiety and that the resulting classifiers outperformed standard baselines; the code and algorithm have been open-sourced on GitHub. Separately, according to a Unite.AI report, the framework prototype uses Gemini 2.5 Pro for concept decomposition and Gemini 2.5 Flash for classification and boundary-sample mining, without additional training or fine-tuning; in a study where 9 participants each completed two sessions of about 90 minutes, F1 improved by 11% over the zero-shot baseline in the pay-to-play scenario and by 10% in the healthy food scenario.
AI-generated from public reports
Latest development · Report
Google Research proposes Agile Deliberation, using human deliberation to build subjective image classifiersDevelopments
1 confirmed developments
- Source:Google Research
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 5
Unite.AI
Google Research proposes Agile Deliberation, using human deliberation to build subjective image classifiers
On November 4, 2026, Google Research released Agile Deliberation, an interactive framework for building image classifiers around subjective categories, accompanied by open-source code and a notebook demo. The framework simultaneously converts users' gradually formed definitions into readable specifications and prompts that drive a vision-language model classifier; the prototype uses Gemini 2.5 Pro for concept decomposition and Gemini 2.5 Flash for classification and boundary sample mining, with no additional training or fine-tuning. In a study where 9 participants each completed two sessions of about 90 minutes, Agile Deliberation improved F1 by 11% over the zero-shot baseline in the pay-to-play scenario, while human deliberation improved it by 8%; in the healthy food scenario, it improved by 10% and decreased by 3%, respectively, and participants reported lower engagement and lower unease than with the human approach.
Google ResearchFirst-party
Agile Deliberation: How Google Research Uses AI to Help Humans Clarify the Gray Areas of Subjective Classification
Google Research proposes the Agile Deliberation framework, which turns the construction of subjective image classifiers into an iterative dialogue between humans and LLMs: users first break down subcategories and exclusions, then the VLM proactively selects semantically borderline images, and the feedback enters an automatic prompt optimization loop. The team validated this in 18 real-human experiments of 1.5 hours each; the process reduced user fatigue and anxiety, and the resulting classifiers outperformed standard baselines. The code and algorithms are open-sourced on GitHub.
Google ResearchRelated discussion
Google Research joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.