What is Grounding?
Tying a model's claims to real evidence so its output can be checked against the facts.
Formal Definition
Grounding is the degree to which a model's statements are supported by verifiable evidence, such as the source data provided to it, rather than invented. Well-grounded output can be traced back to facts. AI Stock Challenge scores evidence grounding directly: a model is rewarded when its stock claims are supported by the market data it was given.
In Simple Terms
It is whether an AI is actually backing up what it says. A grounded answer points to real data you can check; an ungrounded one just sounds confident. The benchmark rewards models that cite the evidence in front of them.
Example
If a model says a stock is cheap, grounding asks whether the valuation numbers it was handed actually support that, or whether it made the claim up.