ReFT / LoReFT (Representation Fine-Tuning) & CAA (Contrastive Activation Addition)

freezes 100% of model weights and intervenes directly on hidden residual stream representations $h \in \mathbb{R}^d$ at a subset of layers:

• LoReFT: Learns a low-rank projection intervention $\Phi(h) = h + R^\top (W h + b - R h)$ with orthonormal rows $R \in \mathbb{R}^{r \times d}$ using 10–50× fewer parameters than LoRA.
• CAA / Activation Steering: Computes the mean difference vector $v_L = \mathbb{E}[h_L(x, y^+)] - \mathbb{E}[h_L(x, y^-)]$ between contrastive positive and negative behavior examples at layer $L$ and adds $\alpha v_L$ at inference time ("training-free alignment").