This paper proposes a single-pass, depth-selective reading framework for multi-aspect sentiment analysis, designed to reduce computational redundancy by separating shared sentence encoding from aspect-specific querying. Our work managed to reduce end-to-end computation by up to 60% in multi-aspect settings (M > 2).

Rather than treating generation and editing as separate problems, this paper - OneHOI models both as a unified conditional denoising process grounded in structured ⟨person, action, object⟩ relations, enabling (i) Single-interaction generation; (ii) Identity-preserving editing, and (iii) Multi-HOI editing. All within one coherent framework.

We address label unlearning in Vertical Federated Learning (VFL), where labels are essential for training yet also privacy-sensitive. Our approach uses representation-level manifold mixup to generate synthetic embeddings for both forgotten and retained samples, enabling more effective and efficient gradient-based forgetting while preserving model performance

Gorgeous is a diffusion-based makeup generator that turns any image prompt (like fire or moonlight) into creative, narrative-driven makeup on a face, rather than simply copying an existing style. From inspiration to personalized beauty in seconds to unlock scalable, AI-powered creativity for brands, artists, and consumers.

We propose Maverick, the first Collaboration-free Federated Unlearning framework, enabling unlearning based on a single client’s request for medical applications. That is, as soon as a client requests data removal, our method can act independently, eliminating the need for global collaboration from other clients