Google's borderline-image labelling lifts Gemini classifier F1 by 10.5 points

single source· 1 articles · confidence: low · first seen 2026-11-05 05:42 UTC

What this means for you

Nothing to act on. No code, weights or dataset is named, and the 10.5-point gain is reported without an evaluation date or test set, so it cannot be lined up against published classifier numbers yet.

Google Research has released a method for training image classifiers on visual categories where the boundary is a judgement call. The framework, Agile Deliberation, keeps a person in the loop: users label borderline images, the classifier is adjusted, and the cycle repeats. Across 18 live sessions it raised the F1 score (a combined measure of precision and recall) of a Gemini-based classifier by 10.5 points against a zero-shot baseline. The report gives no evaluation date, no test-set details and no code or weights.

Key facts

  • ·Google Research released a framework called Agile Deliberation for defining subjective visual concepts with users in the loop. source
  • ·The method works by having users label borderline images rather than writing a definition. source
  • ·It was tested in 18 live sessions. source
  • ·F1 for a Gemini-based classifier rose 10.5 points over a zero-shot baseline; no evaluation date or test set is given. source

What the sources say

  • MarkTechPost — Short release note on the framework; supplies the session count and the F1 gain, no method detail.

Sources

The original reporting. Follow these — they did the work.

← the wire