Probing the Limits of Stylistic Alignment in Vision-Language Models
This work addresses the problem of expensive data acquisition for stylistic alignment in vision-language models, but it is incremental as it focuses on data efficiency for specific styles.
The paper tackled the challenge of aligning small vision-language models to specific styles like humor or romantic with limited preference data, finding that minimal data can achieve stylistic saturation and benchmarking their performance limits.
Vision-language models are increasingly used to generate image captions in specific styles, such as humor or romantic. However, these transformer-based models often struggle with this subjective task in a zero-shot setting. While preference data can be used to align them toward a desired style, such data is expensive to acquire, limiting the ability to explore the models' full capabilities. This work addresses this by studying the data efficiency of aligning small vision-language models to humor and romantic styles. This approach helps to define the performance limits of these models and determine how little preference data is needed to achieve stylistic saturation, benchmarking their capabilities and limitations.