Projects
Current and selected research projects.
Current Projects
ARC — Aligned Riemannian Concept Erasure
Geometry-aware concept and feature unlearning for diffusion models. I model text representations on a hyperspherical manifold and study geodesic projection operators for localized, training-free removal while preserving neighboring semantics.
Feature and Concept Unlearning in LLMs
Analyzing where semantic features, concepts, and task-specific knowledge emerge and disentangle across transformer layers, with the goal of enabling localized and sequential unlearning without disrupting unrelated capabilities.
Semantic Reconstruction & Parallel Reasoning
Developing position-relaxed reconstruction objectives for masked diffusion language models to improve deep-mask semantic reconstruction and make parallel decoding more effective on language and reasoning tasks.
Model- and Task-Aware Data Curation
Selecting data jointly with respect to the target model, task, dataset composition, and modality gap to improve cross-modal alignment, coverage, and data efficiency.
Selected Published Projects
HEART
Training-free fine-grained subject and attribute control using the intrinsic hyperspherical geometry of text-conditioning representations and Kent-representation traversal.
Paper →CURE
Fast, training-free concept erasure through closed-form cross-attention weight editing with orthogonal projection and spectral representation geometry.
Paper →SlimDiff
Activation-guided, timestep-aware, operator-aware low-rank compression for diffusion models, with approximately 35% faster inference and about 100M parameters removed.
Paper →Backpropagation-Free Learning
Direct Feedback Alignment combined with structured low-rank geometry and orthogonality-preserving updates for scalable local learning.
WACV Paper → WiCV / CVPRW →AlphaBlend
Mixed-alphabet-set multipliers and quantization for efficient DNN workloads, connecting hardware efficiency directly to model design.
Paper →