Towards Cost-Effective Reward Guided Text Generation

Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a \emph{single call} to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.

Uncertainty-Guided Optimization on Large Language Model Search Trees

Beam search is a standard tree search algorithm when it comes to finding sequences of maximum likelihood, for example, in the decoding processes of large language models. However, it is myopic since it does not take the whole path from the root to a leaf into account. Moreover, it is agnostic to prior knowledge available about the process: For example, it does not consider that the objective being maximized is a likelihood and thereby has specific properties, like being bound in the unit interval. Taking a probabilistic approach, we define a prior belief over the LLMs' transition probabilities and obtain a posterior belief over the most promising paths in each iteration. These beliefs are helpful to define a non-myopic Bayesian-optimization-like acquisition function that allows for a more data-efficient exploration scheme than standard beam search. We discuss how to select the prior and demonstrate in on- and off-model experiments with recent large language models, including Llama-2-7b, that our method achieves higher efficiency than beam search: Our method achieves the same or a higher likelihood while expanding fewer nodes than beam search.

A Critical Look At Tokenwise Reward-Guided Text Generation

Large language models (LLMs) can significantly be improved by aligning to human preferences -- the so-called reinforcement learning from human feedback (RLHF). However, the cost of fine-tuning an LLM is prohibitive for many users. Due to their ability to bypass LLM finetuning, tokenwise reward-guided text generation (RGTG) methods have recently been proposed. They use a reward model trained on full sequences to score partial sequences during a tokenwise decoding, in a bid to steer the generation towards sequences with high rewards. However, these methods have so far been only heuristically motivated and poorly analyzed. In this work, we show that reward models trained on full sequences are not compatible with scoring partial sequences. To alleviate this issue, we propose to explicitly train a Bradley-Terry reward model on partial sequences, and autoregressively sample from the implied tokenwise policy during decoding time. We study the property of this reward model and the implied policy. In particular, we show that this policy is proportional to the ratio of two distinct RLHF policies. We show that our simple approach outperforms previous RGTG methods and achieves similar performance as strong offline baselines but without large-scale LLM finetuning.