Behavior
What models do, and the mechanisms behind it.
Tools
ProbeOne prompt, with and without an intervention.VectorsDirections in a model's activations that stand for a concept.Training setsIs a training set sound? Its pairs side by side, with what is wrong flagged.TrainingFine-tuning runs that change what a model does: SFT, DPO, GRPO.CircuitsWhich internal features produce an answer.