Conservative Bias in Large Language Models: Measuring Relation Predictions

2025-06-09

AI Summary

Large language models (LLMs) exhibit a pronounced conservative bias in relation extraction—over-relying on the “No_Relation” label when encountering ambiguous or unseen relations, thereby discarding critical relational information. Method: We formally define and quantify the binary trade-off between this conservative bias and hallucination, revealing the former occurs at roughly twice the rate of the latter. Grounded in Hobson’s Choice theory, we introduce the “Hobson’s Choice” concept to characterize a safety-oriented yet information-devoid degenerate decision mechanism. Using SBERT-based semantic similarity modeling and comparative experiments across constrained, semi-constrained, and open-ended prompting paradigms, we systematically validate the bias pattern across multiple benchmark datasets and establish a reproducible bias measurement protocol. Contribution/Results: Conservative bias is identified as the primary cause of relation extraction failure. Our framework demonstrates both generality—applicable across diverse LLMs and relation schemas—and robustness—validated under varied prompting strategies and evaluation settings.

Technology Category

• Reasoning under Uncertainty: Relational Probabilistic Models
• Natural Language Processing: Safety and Robustness
• Knowledge Representation and Reasoning: Preferences

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large language models (LLMs) exhibit pronounced conservative bias in relation extraction tasks, frequently defaulting to No_Relation label when an appropriate option is unavailable. While this behavior helps prevent incorrect relation assignments, our analysis reveals that it also leads to significant information loss when reasoning is not explicitly included in the output. We systematically evaluate this trade-off across multiple prompts, datasets, and relation types, introducing the concept of Hobson's choice to capture scenarios where models opt for safe but uninformative labels over hallucinated ones. Our findings suggest that conservative bias occurs twice as often as hallucination. To quantify this effect, we use SBERT and LLM prompts to capture the semantic similarity between conservative bias behaviors in constrained prompts and labels generated from semi-constrained and open-ended prompts.

Problem

• Measure conservative bias in LLMs' relation predictions
• Analyze trade-off between information loss and incorrect assignments
• Quantify semantic similarity in bias behaviors across prompts

Innovation

• SBERT measures semantic similarity in bias
• LLM prompts evaluate conservative bias
• Hobson's choice concept captures safe labels