# Human-Calibrated Automated Testing and Validation of Generative Language Models

**Authors:** Agus Sudjianto, Aijun Zhang, Srinivas Neppalli, Tarun Joshi, Michal Malohlava  
**arXiv:** [2411.16391](https://arxiv.org/abs/2411.16391)  
**Date:** 2024-11-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.16391v1/

## Key points

- Authors introduce a framework for evaluating and validating generative language models (GLMs), focusing on Retrieval-Augmented Generation (RAG) systems in high-stakes domains like banking.
- Challenges in GLM evaluation include open-ended outputs and subjective quality assessments, addressed by the Human-Calibrated Automated Testing (HCAT) framework.
- HCAT framework includes automated test generation, embedding-based metrics for functionality assessment, risk, and safety attributes evaluation.
- Two-stage calibration aligns machine-generated evaluations with human judgments using probability calibration and conformal prediction techniques.
- Robustness testing evaluates model performance under adversarial conditions, out-of-distribution scenarios, and various input conditions.
- Targeted weakness identification through marginal and bivariate analysis helps pinpoint areas for improvement in GLM performance.
- The framework offers a scalable, transparent, and interpretable approach to assessing GLMs for accuracy, transparency, regulatory compliance in high-stakes environments like banking.

## Summary

In their paper titled "Human-Calibrated Automated Testing and Validation of Generative Language Models," authors Agus Sudjianto, Aijun Zhang, Srinivas Neppalli, Tarun Joshi, and Michal Malohlava introduce a comprehensive framework for evaluating and validating generative language models (GLMs). The focus is on Retrieval-Augmented Generation (RAG) systems used in high-stakes domains such as banking. The evaluation of GLMs presents challenges due to the open-ended nature of their outputs and the subjective nature of quality assessments. To address these challenges, the authors propose the Human-Calibrated Automated Testing (HCAT) framework. The HCAT framework leverages the structured design of RAG systems where generated responses are anchored in a predefined document collection. It integrates several key components: automated test generation through stratified sampling, embedding-based metrics for explainable assessment of functionality, risk, and safety attributes. Additionally, it uses a two-stage calibration approach that aligns machine-generated evaluations with human judgments using probability calibration and conformal prediction techniques. Furthermore, robustness testing is incorporated within the framework to evaluate model performance under adversarial conditions, out-of-distribution scenarios, and various input conditions. The authors also implement targeted weakness identification through marginal and bivariate analysis to pinpoint specific areas for improvement in GLM performance. Overall, this human-calibrated and multi-layered evaluation framework offers a scalable, transparent, and interpretable approach to assessing GLMs. It provides a practical and reliable solution for deploying these models in applications where accuracy, transparency, and regulatory compliance are crucial. The proposed framework not only enhances the evaluation process but also ensures that GLMs meet stringent requirements for reliability and effectiveness in high-stakes environments such as banking.