Saved in:
Bibliographic Details
Main Authors: Xu, Wenda, Zhu, Guanglei, Zhao, Xuandong, Pan, Liangming, Li, Lei, Wang, William Yang
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2402.11436
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909225689546752
author Xu, Wenda
Zhu, Guanglei
Zhao, Xuandong
Pan, Liangming
Li, Lei
Wang, William Yang
author_facet Xu, Wenda
Zhu, Guanglei
Zhao, Xuandong
Pan, Liangming
Li, Lei
Wang, William Yang
contents Recent studies show that large language models (LLMs) improve their performance through self-feedback on certain tasks while degrade on others. We discovered that such a contrary is due to LLM's bias in evaluating their own output. In this paper, we formally define LLM's self-bias - the tendency to favor its own generation - using two statistics. We analyze six LLMs (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral and DeepSeek) on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks. The code and data are released at https://github.com/xu1998hz/llm_self_bias.
format Preprint
id arxiv_https___arxiv_org_abs_2402_11436
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement
Xu, Wenda
Zhu, Guanglei
Zhao, Xuandong
Pan, Liangming
Li, Lei
Wang, William Yang
Computation and Language
Artificial Intelligence
Recent studies show that large language models (LLMs) improve their performance through self-feedback on certain tasks while degrade on others. We discovered that such a contrary is due to LLM's bias in evaluating their own output. In this paper, we formally define LLM's self-bias - the tendency to favor its own generation - using two statistics. We analyze six LLMs (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral and DeepSeek) on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks. The code and data are released at https://github.com/xu1998hz/llm_self_bias.
title Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2402.11436