ModelsMistral Large 4 makes seeing the evidence part of the reasoning problem
The new multimodal preview puts visual grounding under scrutiny: can a model identify the right evidence before reasoning about it?
Model-building research and analysis covering Mistral.
ModelsThe new multimodal preview puts visual grounding under scrutiny: can a model identify the right evidence before reasoning about it?
TrainingReinforcement learning must keep challenging a model as it improves. The engineering problem is producing useful attempts, trustworthy rewards and a curriculum that does not stand still.