Per-behavior mean-difference steering vectors in the Llama 2 residual stream causally steer behavior and transfer across layers and from base to chat

Panickssery et al. build one residual-stream vector per behavior as the mean activation difference over contrastive A/B answer pairs (seven behaviors, 290-1000 pairs each) in Llama 2 7B and 13B Chat [panickssery-etal-2024-caa] Adding or subtracting the vector at layer 13 (7B optimal layer; 14 or 15 for 13B) shifts multiple-choice behavior in the steered direction for six of seven behaviors on 50 held-out questions, e.g. refusal 0.56 / 0.78 / 0.86 at multipliers -1 / 0 / +1 in 13B Chat; sycophancy is not monotone there [panickssery-etal-2024-caa] A vector from layer 13 steers when applied at other layers (steep drop-off near layer 17), and vectors from the Llama 2 7B base model transfer to the Chat model, most strongly between layers 10 and 15 [panickssery-etal-2024-caa] PCA of answer-token activations shows behavioral clustering emerging suddenly (refusal: layer 10 in 7B Chat) [panickssery-etal-2024-caa]