Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Quantifying and Mitigating Default Behaviors in Text-to-Image Models

AbstractText-to-image (T2I) models consistently collapse to stereotypical attribute values when prompts leave concepts underspecified, a phenomenon we term default behaviors. We conduct a large-scale empirical study using , an LLM-driven measurement approach that quantifies diversity at the level of individual concept attributes. We find pervasive default behaviors across all evaluated T2I models, with stronger models tending to exhibit less diversity. We trace these defaults to reporting bias in training captions: when captions omit common attributes, the accompanying images are dominated by a narrow set of values, and model outputs closely mirror this skew. We confirm this link causally with controlled fine-tuning interventions. As a proof of concept, we show that making sampled attribute values explicit broadens outputs, with some loss of faithfulness to secondary prompt details. Together, our results recast low diversity not as an inherent limitation of T2I models, but as a measurable and steerable consequence of what their training captions leave unsaid. Code and data available at: https://royira.github.io/GRADE.

More from EMNLP 2026

All EMNLP 2026 papers →