A Multifaceted Analysis of Negative Bias in Large Language Models through the Lens of Parametric Knowledge

2025-11-14

IEEE Transactions on Audio, Speech, and Language Processing

AI Summary

Large language models (LLMs) exhibit a systematic negative bias in binary decision tasks—overproducing “no” responses. This work identifies prompt formatting—not semantic content—as the primary driver of this bias. To investigate, we propose a fine-grained evaluation pipeline grounded in parametric knowledge categorization, enabling the first empirical demonstration that models resort to a “default negation” heuristic due to knowledge gaps. Through controlled ablation studies, we systematically assess the impact of prompt engineering, context injection, explicit “I don’t know” options, and chain-of-thought (CoT) prompting: the former two significantly mitigate the bias, whereas CoT exacerbates it. Our core contributions are threefold: (1) a mechanistic deconstruction of the bias’s triggers, (2) empirical validation of prompt structure—not just content—as decisive in bias induction, and (3) a reproducible evaluation framework with empirically validated mitigation strategies.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Preferences

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Negative bias refers to the tendency of large language models (LLMs) to excessively generate negative responses in binary decision tasks (e.g., yes-no question answering). Previous research has focused on detecting and addressing negative attention heads that induce negative bias. However, the underlying detailed factors influencing negative bias remain underexplored. In this paper, we demonstrate that LLMs exhibit format-level negative bias, meaning the prompt format more influences their responses than the semantics of the negative response. For the fine-grained study of the negative bias, we introduce a pipeline for constructing the evaluation set, which systematically categorizes the dataset into three subsets based on the model's parametric knowledge: correct, incorrect, and insufficient relevant knowledge. Through analysis of this evaluation set, we identify a shortcut behavior in which models tend to generate negative responses when they lack sufficient knowledge to answer a yes-no question, leading to negative bias. We further examine how negative bias changes under various prompting scenarios related to parametric knowledge. We observe that providing relevant context and offering an"I don't know"option generally reduces negative bias, whereas chain-of-thought prompting tends to amplify the bias. Finally, we demonstrate that the degree of negative bias can vary depending on the type of prompt, which influences the direction of the response. Our work reveals the various factors that influence negative bias, providing critical insights for mitigating it in LLMs.

Problem

• Analyzes negative bias in LLMs during binary decision tasks
• Investigates how prompt formats override semantic meaning
• Explores knowledge-dependent shortcut behaviors causing negative responses

Innovation

• Pipeline categorizes dataset by parametric knowledge levels
• Identifies shortcut behavior causing bias in knowledge gaps
• Analyzes bias changes under different prompting scenarios