Controlled Decoding from Language Models

TL;DR

Controlled decoding (CD) is a modular, tokenwise RL-based method that uses a prefix scorer to steer language model outputs towards desired rewards, enabling effective control, multi-objective handling, and transferability without retraining.

Contribution

Introduces a novel modular controlled decoding approach that employs a prefix scorer for tokenwise RL, allowing flexible, multi-objective, and transferable control of language models without additional training.

Findings

Effective control on benchmark tasks

Combines multiple rewards at inference time

Transfers control to unseen models without tuning

Abstract

KL-regularized reinforcement learning (RL) is a popular alignment framework to control the language model responses towards high reward outcomes. We pose a tokenwise RL objective and propose a modular solver for it, called controlled decoding (CD). CD exerts control through a separate prefix scorer module, which is trained to learn a value function for the reward. The prefix scorer is used at inference time to control the generation from a frozen base model, provably sampling from a solution to the RL objective. We empirically demonstrate that CD is effective as a control mechanism on popular benchmarks. We also show that prefix scorers for multiple rewards may be combined at inference time, effectively solving a multi-objective RL problem with no additional training. We show that the benefits of applying CD transfer to an unseen base model with no further tuning as well. Finally, we show that CD can be applied in a blockwise decoding fashion at inference-time, essentially bridging the gap between the popular best-of-K strategy and tokenwise control through reinforcement learning. This makes CD a promising approach for alignment of language models.