CLM: Contrastive Language Model
An 8B model from Stanford and NVIDIA that picks agent actions by vector similarity instead of generating text.
An 8B model from Stanford and NVIDIA that does not generate text. It puts the agent's current state and every possible action in the same vector space, then picks the closest action. Up to 9x faster than a normal model on the same decisions. Good fit for the small, repeated choices inside an agent loop: which tool to call, where to click, is this result correct.