A Long Way to Go: Investigating Length Correlations in RLHF

By Prasann Singhal, Tanya Goyal, Jiacheng Xu, Greg Durrett. 2023

284 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Information Systems, Alignment & RLHF

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

  1. Percy Liang
  2. Mohit Bansal
  3. Luke S. Zettlemoyer
  4. Jianfeng Gao
  5. Yejin Choi
  6. Quoc V. Le
  7. Danqi Chen
  8. Yoshua Bengio

In Information Systems

  1. Percy Liang
  2. Danqi Chen
  3. Yu Meng
  4. Tianyi Zhou
  5. Tom Goldstein
  6. Tatsunori B. Hashimoto
  7. Bryan Catanzaro
  8. Stefano Ermon

In Alignment & RLHF

  1. Rafael Rafailov
  2. Yu Meng
  3. Percy Liang
  4. Tatsunori B. Hashimoto
  5. Wei Xiong
  6. Chelsea Finn
  7. Ryan Park
  8. Stefano Ermon

Citations in the corpus