Efficient Controlled Language Generation with Low-Rank Autoregressive Reward Models

Language models trained on large amounts of data are known to produce inappropriate content in some cases and require careful tuning to be used in the real world. We revisit the reward augmented decoding (RAD) approach to control the generation from a language model using the scores from a task-specific reward model. We investigate the training objective of RAD, and reformulate it as a task of learning a reward matrix. We show that RAD is designed to support high flexibility when representing the reward matrices, which leads to a higher computational costs during decoding. However, we demonstrate that RAD does not use its full flexibility. Motivated by this, we propose a simpler but more efficient low-rank parametrization of the reward model enabling fast and effective guided decoding. For the detoxification and sentiment control tasks, we show that our low-rank reward model performs on par with the more flexible RAD parametrization, while requiring only a single reward model call per generated token.