RPAM: A Principled Metric for Evaluating Associations in Language Models with High Predictive Validity in Downstream Outputs

2026-07-06

AI Summary

Existing methods for evaluating bias in language models struggle to balance generalizability with the ability to predict actual bias manifested in downstream generated content. This work proposes an upstream evaluation metric—Relative Probability Association Measure (RPAM)—which assesses cross-model, generalizable bias by analyzing the relative strength of associations between concepts in model continuation probabilities. Both theoretical analysis and empirical experiments demonstrate that RPAM is the first upstream association metric to establish strong correlations with human implicit and explicit associations as well as bias observed in downstream tasks. Validation on Mistral-7B-Instruct, Mistral-7B, and GPT-2 shows that RPAM significantly outperforms existing approaches in predictive validity.

Abstract

Language models (LMs) exhibit problematic biases, such as stereotypes. Effectively analyzing and mitigating such biases requires accurate and generalizable evaluation methods of the underlying associations. Some existing approaches focus on downstream metrics that analyze associations in generated text. Since generated text content can vary drastically across LMs, such metrics often require specialized evaluation datasets, which limits the generalization of such downstream metrics. In contrast, upstream metrics examine LMs at the fundamental level of embeddings or continuation probabilities, enabling principled association analyses across LMs. Yet, to date, no upstream metric for generative LMs has uncovered a strong relationship with real-world associations, including those measured in generated text. To address this gap, we introduce the Relative Probability Association Metric (RPAM), an association evaluation metric for generative LMs. For three LMs of different quality of language generation and purpose (Mistral-7B-Instruct, Mistral-7B, and GPT-2) and well-studied evaluation datasets (WEAT-WS, Bellezza, WS-353, and SST2), we find a strong relationship between upstream RPAM measurements and corresponding implicit and explicit associations observed in humans, as well as biases measured downstream with LM-specific tasks, outperforming prior record values where applicable.

Problem

• bias evaluation
• language models
• upstream metrics
• downstream outputs
• association measurement

Innovation

• Relative Probability Association Metric
• upstream metric
• language model bias
• predictive validity
• association evaluation