Shallow Preference Signals: Large Language Model Aligns Even Better with Truncated Data?

TL;DR

This paper reveals that large language models primarily rely on early tokens for preference signals, and training on truncated data focusing on these tokens can achieve comparable or better alignment performance, suggesting a shift in alignment strategies.

Contribution

The study uncovers the prevalence of shallow preference signals in LLMs and demonstrates that training on truncated datasets can enhance alignment efficiency and effectiveness.

Findings

Models trained on truncated data perform as well or better than full data models.

Shallow preference signals are concentrated in the early tokens of responses.

Decoding strategies leveraging shallow signals improve alignment and efficiency.

Abstract

Aligning large language models (LLMs) with human preferences remains a key challenge in AI. Preference-based optimization methods, such as Reinforcement Learning with Human Feedback (RLHF) and Direct Preference Optimization (DPO), rely on human-annotated datasets to improve alignment. In this work, we identify a crucial property of the existing learning method: the distinguishing signal obtained in preferred responses is often concentrated in the early tokens. We refer to this as shallow preference signals. To explore this property, we systematically truncate preference datasets at various points and train both reward models and DPO models on the truncated data. Surprisingly, models trained on truncated datasets, retaining only the first half or fewer tokens, achieve comparable or even superior performance to those trained on full datasets. For example, a reward model trained on the Skywork-Reward-Preference-80K-v0.2 dataset outperforms the full dataset when trained on a 40\% truncated dataset. This pattern is consistent across multiple datasets, suggesting the widespread presence of shallow preference signals. We further investigate the distribution of the reward signal through decoding strategies. We consider two simple decoding strategies motivated by the shallow reward signal observation, namely Length Control Decoding and KL Threshold Control Decoding, which leverage shallow preference signals to optimize the trade-off between alignment and computational efficiency. The performance is even better, which again validates our hypothesis. The phenomenon of shallow preference signals highlights potential issues in LLM alignment: existing alignment methods often focus on aligning only the initial tokens of responses, rather than considering the full response. This could lead to discrepancies with real-world human preferences, resulting in suboptimal alignment performance.