Data for: Implicit Bias-Like Patterns in Reasoning Models

This deposit contains the raw and processed data for the Reasoning Model Implicit Association Test (RM-IAT), a method for measuring implicit bias-like processing in reasoning models — LLMs that generate step-by-step reasoning before producing a final answer.Contents:- rmiat/ — RM-IAT outputs for five reasoning models (o3-mini, DeepSeek-R1, gpt-oss-20b, Qwen3-8B, Claude 3.7 Sonnet), 10 IAT variants each. Each CSV contains the target word, category, model-assigned attribute, full chain-of-thought reasoning text, reasoning token count, and association-compatible/incompatible condition label. .RData files contain fitted mixed-effects models used in the paper's analyses.- downstream/ — Results from two downstream behavioral tasks (Word Association Task and Resume Decision Task) used to establish convergent validity between RM-IAT asymmetries and behavioral bias.Research context: Implicit biases are automatic mental processes that shape perceptions, judgments, and behavior. Prior work on bias in LLMs has focused on model outputs; the RM-IAT instead measures bias-like processing by comparing reasoning-token usage across association-compatible vs. association-incompatible task conditions. Across four models (o3-mini, DeepSeek-R1, gpt-oss-20b, Qwen3-8B), association-incompatible tasks consistently required greater computational effort. Claude 3.7 Sonnet showed a reversed pattern, linked via thematic analysis to its internal focus on reasoning about bias and stereotypes. Larger RM-IAT asymmetries predicted stronger behavioral bias in both downstream tasks.