Minimally Invasive Steering Optimizes Language Model Outputs Without Retraining
Read this because Developers can improve model alignment and reward scores without expensive fine-tuning or parameter updates, preserving original model capabilities.
New MISVO method steers frozen language models via pre-logit vectors, balancing reward optimization with generation quality using local KL geometry.