ML//Training//fine-tuning//ReFT

- Representation Fine-Tuning: instead of adapting weights (LoRA), intervene on internal representations (activations)

Representation Fine-Tuning: instead of adapting weights (LoRA), intervene on internal representations (activations)

Modify a learned linear subspace of hidden activations at specific layers and positions.

Far fewer parameters than LoRA — same or better performance with 10–50x fewer trainable params.

Built on representation engineering: steer behavior via internal representations, not weight edits.