DATG++: Efficient Multimodal Control via Dynamic Attribute Graphs with Fairness-Safety Alignment

Rong Fu, GUO XIAOYU, Simon James Fong

Keywords: Dynamic Attribute Graphs, Fairness-Safety Alignment, Efficient Multimodal Control, Attribute Boundary Controlling, Plug-and-Play Framework

TL;DR: DATG++ pioneers dynamic attribute graphs for efficient multimodal control, enabling precise semantic steering with integrated fairness-safety alignment while preserving generative quality.

Abstract:

We present DATG++, a flexible framework for controlled text generation that seamlessly integrates multimodal inputs and ethical considerations. By leveraging dynamic attribute graphs, DATG++ enables fine-grained control over semantic attributes such as sentiment or toxicity, while maintaining fluency and generality across diverse contexts. Unlike traditional approaches that rely on retraining or rigid templates, DATG++ operates as a plug-and-play module compatible with large language models, offering interpretable and adaptable control through graph-based semantic manipulation.

The framework combines probabilistic adjustment and prompt-based guidance to steer generation toward desired attributes. It further extends to multimodal scenarios by aligning attribute representations across text, image, and audio modalities. A built-in fairness-safety mechanism mitigates biased outputs and supports privacy-preserving generation. DATG++ promotes transparency, efficiency, and ethical alignment, paving the way for responsible deployment in domains such as human-robot interaction, assistive technologies, and scientific communication.

Primary Area: generative models

Code Of Ethics: true

Submission Guidelines: true

Anonymous Url: true

No Acknowledgement Section: true

Submission Number: 1703