FACTS Grounding

Measures: The model answers a request using only a supplied document of up to 32k tokens. Responses that do not fulfill the request are disqualified. The rest are scored on whether every claim is supported by the document. v1 used a judge panel of Gemini 1.5 Pro, GPT-4o and Claude 3.5 Sonnet; v2 uses Gemini 2.5 Flash and GPT-5.

Maintainer: Google DeepMind / Google Research; Kaggle runs the leaderboard
Introduced: 2024
Status: Active
Saturation: Headroom
Typical provenance: Independent
In scoring set: No

Not scored: Grounding is already a component of facts-benchmark-suite; scoring both would count it twice.

Paper Official leaderboard

Leaderboard

Show independent results only

Every result we hold, whoever ran it. Only independent and third-party rows count towards a score.

No results to show.