Sohini Chintala

Safe Inference-Time Alignment via Lagrangian Reward Augmentation

Yaswanth Chittepu, Ativ Joshi, Sohini Chintala, Scott Niekum

Inference-time alignment steers a frozen language model during decoding using auxiliary reward signals, avoiding the cost of repeated weight updates. However, existing inference-time alignment methods typically optimize a single scalar score, so explicit safety constraints must e…Read more