From Lists to Emojis: How Format Bias Affects Model Alignment

• URL: http://arxiv.org/abs/2409.11704v1
• Date: Wed, 18 Sep 2024 05:13:18 GMT
• Title: From Lists to Emojis: How Format Bias Affects Model Alignment
• Authors: Xuanchang Zhang, Wei Xiong, Lichang Chen, Tianyi Zhou, Heng Huang, Tong Zhang,
• Abstract summary: We study format biases in reinforcement learning from human feedback. Many widely-used preference models, including human evaluators, exhibit strong biases towards specific format patterns. We show that with a small amount of biased data, we can inject significant bias into the reward model.
• Score: 67.08430328350327
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• 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.

Related papers

• GPTBIAS: A Comprehensive Framework for Evaluating Bias in Large Language Models [83.30078426829627]
  Large language models (LLMs) have gained popularity and are being widely adopted by a large user community. The existing evaluation methods have many constraints, and their results exhibit a limited degree of interpretability. We propose a bias evaluation framework named GPTBIAS that leverages the high performance of LLMs to assess bias in models.
  arXiv  Detail & Related papers  (2023-12-11T12:02:14Z)
• Current Topological and Machine Learning Applications for Bias Detection in Text [4.799066966918178]
  This study utilizes the RedditBias database to analyze textual biases. Four transformer models, including BERT and RoBERTa variants, were explored. Findings suggest BERT, particularly mini BERT, excels in bias classification, while multilingual models lag.
  arXiv  Detail & Related papers  (2023-11-22T16:12:42Z)
• It's All Relative: Interpretable Models for Scoring Bias in Documents [10.678219157857946]
  We propose an interpretable model to score the bias present in web documents, based only on their textual content. Our model incorporates assumptions reminiscent of the Bradley-Terry axioms and is trained on pairs of revisions of the same Wikipedia article. We show that we can interpret the parameters of the trained model to discover the words most indicative of bias.
  arXiv  Detail & Related papers  (2023-07-16T19:35:38Z)
• Soft-prompt Tuning for Large Language Models to Evaluate Bias [0.03141085922386211]
  Using soft-prompts to evaluate bias gives us the extra advantage of avoiding the human-bias injection. We check the model biases on different sensitive attributes using the group fairness (bias) and find interesting bias patterns.
  arXiv  Detail & Related papers  (2023-06-07T19:11:25Z)
• Debiasing Vision-Language Models via Biased Prompts [79.04467131711775]
  We propose a general approach for debiasing vision-language foundation models by projecting out biased directions in the text embedding. We show that debiasing only the text embedding with a calibrated projection matrix suffices to yield robust classifiers and fair generative models.
  arXiv  Detail & Related papers  (2023-01-31T20:09:33Z)
• Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases [55.45617404586874]
  We propose a few-shot instruction-based method for prompting pre-trained language models (LMs) We show that large LMs can detect different types of fine-grained biases with similar and sometimes superior accuracy to fine-tuned models.
  arXiv  Detail & Related papers  (2021-12-15T04:19:52Z)
• Improving Robustness by Augmenting Training Sentences with Predicate-Argument Structures [62.562760228942054]
  Existing approaches to improve robustness against dataset biases mostly focus on changing the training objective. We propose to augment the input sentences in the training data with their corresponding predicate-argument structures. We show that without targeting a specific bias, our sentence augmentation improves the robustness of transformer models against multiple biases.
  arXiv  Detail & Related papers  (2020-10-23T16:22:05Z)
• Towards Robustifying NLI Models Against Lexical Dataset Biases [94.79704960296108]
  This paper explores both data-level and model-level debiasing methods to robustify models against lexical dataset biases. First, we debias the dataset through data augmentation and enhancement, but show that the model bias cannot be fully removed via this method. The second approach employs a bag-of-words sub-model to capture the features that are likely to exploit the bias and prevents the original model from learning these biased features.
  arXiv  Detail & Related papers  (2020-05-10T17:56:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.