A Long Way to Go: Investigating Length Correlations in RLHF

TL;DR

This paper reveals that reinforcement learning from human feedback (RLHF) often improves model responses primarily by increasing length, with reward models being a key source of this bias, affecting alignment efforts.

Contribution

It uncovers the significant influence of response length in RLHF improvements and identifies reward models as the main source of length bias in training dynamics.

Findings

Length optimization is a major factor in RLHF success.

Purely length-based rewards can replicate RLHF improvements.

Reward models are non-robust and influenced by length biases.

Abstract

Great success has been reported using Reinforcement Learning from Human Feedback (RLHF) to align large language models, with open preference datasets enabling wider experimentation, particularly for "helpfulness" in tasks like dialogue and web question answering. Alongside these improvements, however, RLHF also often drives models to produce longer outputs. This paper demonstrates, on three diverse settings, that optimizing for response length is, much more than previously thought, a significant factor behind RLHF. Studying the strategies RL optimization uses to maximize reward, we find improvements in reward to largely be driven by increasing response length, instead of other features. Indeed, we find that even a purely length-based reward reproduces most downstream RLHF improvements over supervised fine-tuned models. Testing a comprehensive set of length-countering interventions, we identify the dominant source of these biases to be reward models, which, by studying training dynamics, we find are non-robust and easily influenced by length biases in preference data.