STEERING LLAMA 2 WITH CONTRASTIVE ACTIVATIONS

Summary from the TLDR AI newsletter

There are lots of ways to align a language model, for example, SFT, LoRa, RLHF, DPO, Prompting, and Threatening. This work suggests using negative embeddings added to the bias terms to push model generations towards a desired outcome.