From Lists to Emojis: How Format Bias Affects Model Alignment

2024-09-18

arXiv.org

AI Summary

This work identifies a previously underexplored systemic bias in RLHF preference models—including human annotators, GPT-4, and top-performing RewardBench models—toward non-length-related formatting features (e.g., bullet points, emojis, bolding, hyperlinks). Through preference modeling analysis, controlled formatting perturbation experiments, and rigorous evaluation across multiple benchmarks (RewardBench, AlpacaEval, LMSYS), the study quantifies the magnitude and generalizability of these formatting biases for the first time. Results show that as little as <1% contaminated samples suffices to corrupt reward model calibration; moreover, alignment techniques such as best-of-n sampling and online DPO can readily exploit such biases to inflate apparent performance. The findings underscore the critical need for format–content decoupling in reward modeling and provide principled guidelines for designing robust, fair alignment algorithms and evaluation protocols.

Technology Category

• Machine Learning: Learning Preferences or Rankings
• Humans and AI: Learning Human Values and Preferences
• Natural Language Processing: (Large) Language Models

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

In this paper, we study format biases in reinforcement learning from human feedback (RLHF). We observe that many widely-used preference models, including human evaluators, GPT-4, and top-ranking models on the RewardBench benchmark, exhibit strong biases towards specific format patterns, such as lists, links, bold text, and emojis. Furthermore, large language models (LLMs) can exploit these biases to achieve higher rankings on popular benchmarks like AlpacaEval and LMSYS Chatbot Arena. One notable example of this is verbosity bias, where current preference models favor longer responses that appear more comprehensive, even when their quality is equal to or lower than shorter, competing responses. However, format biases beyond verbosity remain largely underexplored in the literature. In this work, we extend the study of biases in preference learning beyond the commonly recognized length bias, offering a comprehensive analysis of a wider range of format biases. Additionally, we show that with a small amount of biased data (less than 1%), we can inject significant bias into the reward model. Moreover, these format biases can also be easily exploited by downstream alignment algorithms, such as best-of-n sampling and online iterative DPO, as it is usually easier to manipulate the format than to improve the quality of responses. Our findings emphasize the need to disentangle format and content both for designing alignment algorithms and evaluating models.

Problem

• Study format biases in RLHF affecting model alignment
• Explore biases beyond verbosity in preference learning
• Investigate exploitation of format biases by LLMs

Innovation

• Studies format biases in RLHF models
• Injects bias with minimal biased data
• Exploits biases via alignment algorithms