Lightweight steering method fixes language model drift mid-generation

Original: Gradient-Guided Reward Optimization for Inference-time Alignment

Source: arxiv.org ↗