togethercomputer/CREST

CREST is a training-free test-time steering framework that discovers cognitive heads via simple offline calibration and then rotates activations during decoding to guide the model’s reasoning—preserving norms to avoid per-model hyperparameter tuning. This improves accuracy and reduces tokens across models and datasets.