---
type: "post"
id: "google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6"
slug: "google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6"
title: "Google Research introduced Agile Deliberation, a VLM framework for building classifiers on subjective concepts."
description: "Google Research has introduced Agile Deliberation, a framework for creating image classifiers for subjective categories through iterative human feedback."
retrieval_nugget: "Google Research has introduced Agile Deliberation, a framework for creating image classifiers for subjective categories through iterative human feedback. Rather than treating labels as fixed, it helps users define positive scope, exclusions, and edge cases while refining a classifier—a useful research direction for teams whose policies cannot be reduced to an obvious label."
published_at: "2026-11-05"
updated_at: "2026-11-07"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["google"]
entities: ["Google Research","Agile Deliberation","Otilia Stretcu","Leijie Wang"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Google Research’s Agile Deliberation framework for classifying subjective concepts is not covered by the archive documents."
prior_refs: []
source_url: "https://research.google/blog/agile-deliberation-how-ai-helps-humans-define-the-gray-areas"
source_urls: ["https://research.google/blog/agile-deliberation-how-ai-helps-humans-define-the-gray-areas"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6/","markdown":"https://newruntime.com/posts/google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6.md","json":"https://newruntime.com/posts/google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6.json"}
markdown_url: "https://newruntime.com/posts/google-research-introduced-agile-deliberation-a-vlm-framework-18d5e6.md"
---

# Google Research introduced Agile Deliberation, a VLM framework for building classifiers on subjective concepts.

## Retrieval answer

Google Research has introduced Agile Deliberation, a framework for creating image classifiers for subjective categories through iterative human feedback. Rather than treating labels as fixed, it helps users define positive scope, exclusions, and edge cases while refining a classifier—a useful research direction for teams whose policies cannot be reduced to an obvious label.

## Summary
Google Research has introduced Agile Deliberation, a framework for creating image classifiers for subjective categories through iterative human feedback. Rather than treating labels as fixed, it helps users define positive scope, exclusions, and edge cases while refining a classifier—a useful research direction for teams whose policies cannot be reduced to an obvious label.

## What happened
Google Research researchers Otilia Stretcu and Leijie Wang introduced Agile Deliberation in a November 4, 2026 post. The framework addresses image-classification tasks built around subjective concepts: categories whose boundaries are not universal, stable, or fully specified before work begins. Google gives examples including “healthy food,” “inspiring artwork,” and content moderation. In these cases, cultural context, personal philosophy, or evolving operating policy can change what should count as a match.

The authors contrast these categories with comparatively objective ones such as a stop sign or a golden retriever. Conventional training workflows assume that a person applying labels already has a clear, stable definition of the category. Google argues that this assumption breaks down for subjective concepts: people may begin with an intuition and discover exceptions, boundaries, and nuances only after seeing real examples. A rigid labelling workflow can therefore force a false certainty on the person defining the task.

Agile Deliberation turns classifier construction into a two-stage workflow. In the first stage, concept scoping, a user breaks a broad concept into a hierarchy of positive sub-concepts and explicit exclusions. For “healthy food,” the post proposes salads, grain bowls, and lean proteins as possible positive categories, while processed foods and raw, unprepared ingredients could be exclusions. The result is an initial structured definition rather than a single broad label.

The second stage, concept iteration, tests that definition against difficult examples. The system retrieves borderline images, exposing cases where the intended classification is unclear. The user responds directly to those cases, and an LLM-driven optimization loop turns that feedback into prompt refinements and a ready-to-use image classifier. Google describes this as an alternative to manual prompt engineering: the user spends effort clarifying the cases that create ambiguity instead of attempting to articulate every rule in advance.

Google evaluated the framework in 18 live sessions with human participants, each lasting 1.5 hours. The post says static benchmarks cannot capture changing intent for subjective concepts, so the researchers observed the interaction directly. Google reports that the workflow reduced user fatigue and anxiety, produced more consistent and higher-performing models, and outperformed standard baselines. It has open-sourced the underlying code and algorithms on GitHub and links to a paper. The supplied material does not provide the repository address, license, numerical effect sizes, baseline details, hardware requirements, pricing, or a hosted service.

## What the archive adds
This is a new research event on a known classification theme. [Parallel Decisions API: web-grounded classification](https://newruntime.com/posts/parallel-decisions-api-web-grounded-classification-dc4d7c/) covered a commercial API for web-grounded classification with batch requests and per-call pricing. Agile Deliberation concerns a different problem: helping a human define a subjective visual category before and during classifier construction.

## Trend signal
No supplied pattern is a sufficiently direct fit. The source is evidence about a research workflow for resolving ambiguous image labels, not evidence of a broader production or organizational shift.

## Why it matters
Many AI product categories sound simple until a team must operationalize them. “Healthy,” “inspiring,” “harmful,” “high quality,” or “acceptable” can conceal policy choices rather than describe naturally settled facts. A classifier can appear inaccurate when the deeper failure is that reviewers, product owners, and users never agreed on what the classifier should recognize or reject.

The useful contribution here is procedural. Agile Deliberation makes the definition work explicit: identify positive sub-categories, state exclusions, inspect borderline cases, and revise the classifier from direct feedback. That gives a team a possible way to distinguish model mistakes from unresolved policy disagreements. It can also create a more reviewable record of why a category has the boundaries it does.

The practical implication is not that teams should replace a production workflow with this research framework. It is that model tuning should not be the first response to a fuzzy task. Before optimizing prompts or collecting a larger label set, teams can ask whether the target concept has documented inclusions, exclusions, and owners able to decide edge cases. Google’s reported live-session results make that hypothesis worth testing, but they do not quantify the operational benefit for deployed systems.

## Limits and uncertainty
The source is a Google Research post, not an independent replication or a production case study. It reports 18 sessions of 1.5 hours but does not provide numerical performance changes, participant details, task distributions, baseline definitions, or evidence about durability after deployment. Claims of reduced fatigue, reduced anxiety, greater consistency, and better performance are Google’s reported findings in the supplied material.

The post also does not establish access conditions beyond saying the code and algorithms are open-sourced. The supplied excerpt has no repository URL, license terms, model requirements, operating costs, privacy model, or instructions for integrating the workflow into a moderation or classification stack. It remains unclear how stable a refined definition stays when policy, users, or incoming image distributions change.

## Skill to build now
Run a short edge-case review for one existing classifier or rubric. Ask the policy owner to write three positive sub-categories and three explicit exclusions, then collect a small set of borderline examples and record decisions with a reason. Separate disagreements about the definition from errors against an agreed definition. This exercise can reveal whether another prompt iteration is needed—or whether the actual missing artifact is a decision policy.

## Reader takeaway
Subjective classifiers cannot be made reliable by pretending their labels are objective. Google Research’s Agile Deliberation frames classifier building as an iterative definition process: scope the concept, state exclusions, inspect edge cases, and revise from feedback. The approach is promising research, but its production performance, access details, and generalizability remain unproven.

## Open question
How well do definitions refined through live deliberation hold up when a team’s policy, users, or image distribution changes after deployment?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)