Understanding the performance gap between online and offline alignment algorithms

TL;DR

This paper investigates why online alignment algorithms outperform offline ones in reinforcement learning from human feedback, revealing the importance of on-policy sampling and the limitations of offline methods.

Contribution

The study provides empirical evidence on the performance gap between online and offline alignment algorithms and explores underlying causes beyond data coverage and quality.

Findings

Online algorithms excel at generation tasks, offline at pairwise classification.

Offline data coverage and quality do not fully explain performance differences.

Sampling process critically impacts the discriminative and generative capabilities of policies.

Abstract

Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and data quality by itself cannot convincingly explain the performance difference. We also find that while offline algorithms train policy to become good at pairwise classification, it is worse at generations; in the meantime the policies trained by online algorithms are good at generations while worse at pairwise classification. This hints at a unique interplay between discriminative and generative capabilities, which is greatly impacted by the sampling process. Lastly, we observe that the performance discrepancy persists for both contrastive and non-contrastive loss functions, and appears not to be addressed by simply scaling up policy networks. Taken together, our study sheds light on the pivotal role of on-policy sampling in AI alignment, and hints at certain fundamental challenges of offline alignment algorithms.