ML//training//fine-tuning//ReFT

ReFT (Representation Fine-Tuning) is a fine-tuning method that adapts a model by intervening on its internal representations (the activations) instead of adapting its weights.

ReFT (Representation Fine-Tuning) is a fine-tuning method that adapts a model by intervening on its internal representations (the activations) instead of adapting its weights.

It modifies a learned linear subspace of the hidden activations at specific layers and positions, and reaches the same or better performance than LoRA, which adds low-rank adapters to the weight matrices, with 10 to 50 times fewer trainable parameters. It builds on representation engineering, which steers behavior through internal representations rather than weight edits.