Implicit Bias-Like Patterns in Reasoning Models

2025-03-14

AI Summary

This paper investigates whether reasoning-capable large language models (LLMs) exhibit human-like implicit biases during step-by-step reasoning. Method: We propose the Reasoning Model Implicit Association Test (RM-IAT), the first method to extend implicit bias detection from model outputs to the reasoning process itself. RM-IAT quantifies token consumption differences between compatible and incompatible semantic associations along reasoning paths, integrating IAT paradigm adaptation, token-level reasoning trajectory analysis, controlled prompt engineering, and cross-task consistency validation. Contribution/Results: Empirical evaluation across mathematical, commonsense, and ethical reasoning tasks reveals that models consume on average 12.7% more tokens when processing incompatible associations—a robust, task-invariant pattern. This demonstrates systematic, implicit-bias-like behavior embedded within LLMs’ internal reasoning dynamics. The work introduces a novel paradigm and a quantifiable tool for probing latent cognitive mechanisms in LLMs.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

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

Abstract

Implicit bias refers to automatic or spontaneous mental processes that shape perceptions, judgments, and behaviors. Previous research examining `implicit bias' in large language models (LLMs) has often approached the phenomenon differently than how it is studied in humans by focusing primarily on model outputs rather than on model processing. To examine model processing, we present a method called the Reasoning Model Implicit Association Test (RM-IAT) for studying implicit bias-like patterns in reasoning models: LLMs that employ step-by-step reasoning to solve complex tasks. Using this method, we find that reasoning models require more tokens when processing association-incompatible information compared to association-compatible information. These findings suggest AI systems harbor patterns in processing information that are analogous to human implicit bias. We consider the implications of these implicit bias-like patterns for their deployment in real-world applications.

Problem

• Examines implicit bias-like patterns in reasoning models.
• Develops RM-IAT to study bias in model processing.
• Finds AI systems process information with bias-like patterns.

Innovation

• Developed RM-IAT for bias analysis
• Analyzed token usage in reasoning models
• Identified implicit bias-like processing patterns