Abstract

Generative Foundation Models (GenFMs) have become widely used across various domains, but concerns remain about their trustworthiness, including truthfulness, safety, fairness, robustness, and privacy. This paper presents a framework to address these issues through three key contributions. First, we review global AI governance policies and industry standards and propose standardized guidelines for assessing and improving GenFM trustworthiness. Second, we introduce TrustGen, a benchmarking platform designed to evaluate models across different dimensions, including text-to-image, large language, and vision-language models. Unlike traditional evaluation methods, TrustGen enables adaptive assessments through metadata curation, test case generation, and contextual variation. Using TrustGen, we analyze the trustworthiness of current GenFMs, highlighting progress while identifying challenges such as overly cautious safety measures and persistent vulnerabilities in open-source models. This work provides a foundation for developing safer and more responsible generative AI and includes an open-source evaluation toolkit for further research.

Method Overview

TrustGen evaluates GenFMs through three main components:

Standardized Guidelines for Trustworthy GenFMs

A set of guidelines developed through multidisciplinary collaboration, integrating technical, ethical, legal, and societal perspectives. These guidelines provide a structured approach to evaluating and improving model trustworthiness.

Dynamic Trustworthiness Evaluation

TrustGen moves beyond static evaluation benchmarks by introducing a modular framework with three key components:

1. Metadata Curation – Collects and organizes evaluation metadata dynamically.
2. Test Case Generation – Produces diverse evaluation cases for different trust dimensions.
3. Contextual Variation – Modifies test cases to ensure adaptability across models and scenarios.

This approach allows for real-time, flexible assessments that reduce biases from predefined test cases.

Trustworthiness Assessment of State-of-the-Art Models

We apply TrustGen to benchmark leading generative models, including text-to-image, large language, and vision-language models. Our results show that while models have made progress, key trade-offs remain between safety, usability, and robustness.

Key Takeaways

Our evaluation of state-of-the-art GenFMs using TrustGen reveals several insights:

1. Persistent trustworthiness challenges:
   • While leading models perform well in safety and fairness, issues remain in truthfulness and robustness.
   • Some models are overly cautious, leading to reduced usefulness in benign scenarios.
2. Open-source models are catching up:
   • Certain open-source models now match or outperform proprietary ones in areas like privacy and fairness.
   • Models such as CogView-3-Plus and Llama-3.2-70B show trustworthiness levels comparable to top commercial models.
3. Narrowing trustworthiness gap:
   • The differences in trust scores among top models have decreased, suggesting improvements across the industry.
   • Collaboration and shared best practices have contributed to more consistent trustworthiness enhancements.
4. Interconnected trustworthiness factors:
   • Improvements in one area (e.g., safety) often impact others (e.g., usability).
   • A balanced approach is needed to ensure models remain both useful and responsible.

Conclusion

TrustGen provides a new benchmark for evaluating the trustworthiness of GenFMs, enabling adaptive and iterative assessments that address limitations of static evaluation methods. While significant improvements have been made, challenges remain, particularly in balancing safety, robustness, and practical usability. Future research should focus on refining evaluation strategies and fostering interdisciplinary collaboration to ensure that generative AI systems are fair, reliable, and aligned with human needs.

To support further research, the TrustEval-toolkit is available at:
🔗 GitHub Repository