This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Whose View of Safety? A Deep DIVE Dataset for Pluralistic Alignment of Text-to-Image Models
Created by
Haebom
Author
Charvi Rastogi, Tian Huey Teh, Pushkar Mishra, Roma Patel, Ding Wang, Mark D iaz, Alicia Parrish, Aida Mostafazadeh Davani, Zoe Ashwood, Michela Paganini, Vinodkumar Prabhakaran, Verena Rieser, Lora Aroyo
Outline
This paper addresses the limitations of existing text-to-image (T2I) models that fail to account for diverse human experiences, and proposes a 'pluralistic alignment' that allows for understanding and reconciling diverse and often conflicting human values. To this end, we provide three major contributions. First, we introduce a novel multimodal dataset for Diverse Intersectional Visual Evaluation (DIVE), which enables deep alignment across diverse safety perspectives through a large number of demographically cross-sectional raters who provided extensive feedback on 1,000 prompts. Second, we empirically confirm that demographic characteristics are important proxies for diverse perspectives in this domain, and reveal significant context-dependent differences in harm perception that are different from existing assessments. Third, we discuss Takeaways for building aligned T2I models, including efficient data collection strategies, LLM judgment functions, and model reconciliation possibilities for diverse perspectives. This study provides a foundational tool for more fair and aligned T2I systems.
Takeaways, Limitations
•
Takeaways:
◦
Presentation of the concept of pluralistic alignment that takes into account various human values and emphasis on its importance
◦
Providing a new multimodal dataset for diverse cross-sectional visual evaluation (DIVE)
◦
Empirically confirming that demographic characteristics are important surrogate variables in safety assessment of the T2I model
◦
Suggesting a direction for building an improved T2I model by presenting an efficient data collection strategy, LLM judgment function, and model adjustment possibility
◦
Providing a foundational tool for building a more fair and aligned T2I system
•
Limitations:
◦
As mentioned in the paper, it contains sensitive content and thus has the potential to cause harm.
◦
Additional validation of the scale and generalizability of the DIVE dataset is needed.
◦
Additional research is needed on the application and effectiveness of the proposed methodology to real T2I models.
◦
Lack of specific technical details on LLM judgment functions and model tunability.
