Computer Vision and Machine Learning Group
Undergraduate Research Assistant (Advisor: Prof. Yaoyao Liu), 2025.05 - 2026.08, Champaign, IL
- Proposed adaptive conditioning methods for 3D-aware synthetic data generation to enhance world model understanding and embodied agent performance in interactive simulations with geometry-conditioned diffusion approaches, reducing FID by 32.8% on ImageNet and improving pose accuracy by 4.2x on PASCAL3D+.
- Scaled foundation-model training through distributed methods across TB-level datasets on National Center for Supercomputing Applications (NCSA) HPC clusters to enforce multi-view consistency with 4-bit NF4 quantization and custom CUDA kernels, decreasing FID by 18.3% on AC3S-50 and reducing generation latency by 78.7% at p95.
- Implemented camera-controlled novel view synthesis for video generation pipelines to advance real-time perception for SLAM, visual odometry, and 3D reconstruction, decreasing LPIPS by 21.0% on GSO and FV4D by 52.0% on OmniObject3D compared to SV4D 2.0.
Links: Lab
Multi-modal Generative Models for Continual Learning
Undergraduate Research Assistant (Advisor: Prof. Yaoyao Liu), 2026.02 - 2026.08, Champaign, IL
- Selected for NVIDIA Academic Grant Program Award, AMD University Program, and Lambda Research Grant to streamline class-incremental learning in multimodal foundation models across audio, image, and text without catastrophic forgetting or cross-modal alignment drift, increasing audio-to-text R@1 by 27.6% on AudioSet and raising last-task accuracy by 40.9% on UrbanSound8K.
- Implemented post-hoc tensor-level weight interpolation for multimodal retrieval on National Artificial Intelligence Research Resource (NAIRR) HPC clusters, minimizing trainable parameters from 182M to 499 sigmoid-parameterized coefficients and decreasing training time by 90.2% compared to LwF.
- Optimized checkpoint fusion pipelines to compose frozen checkpoints into one deployable model without additional inference latency, retaining 99.6% of audio-to-text R@1 after shrinking the exemplar memory by 90.0% and enhancing last-task accuracy by 10.9% on CIFAR-100.
Links: NVIDIA Grant
Reusable Evidence View Aggregation for Context-Efficient RAG Serving
Undergraduate Research Assistant, 2026.02 - 2026.06, Champaign, IL
- Designed attention-based scoring for post-retrieval context compression in retrieval-augmented generation (RAG) to reuse generator attention traces across queries and update scores offline, achieving a 91.6% cache hit rate and boosting F1 by 12.9% on HotpotQA.
- Deployed reusable evidence views keyed to unique documents in order to scale compression cost with documents rather than query volume, improving F1 by 13.2% on Natural Questions and decreasing online compression overhead by 6.1x at under 40 ms per request.
- Accelerated online serving to keep context compression off the GPU critical path, cutting compression latency by 99.3% compared to EXIT and 99.8% compared to FaviComp.
Links: IEEE ICDM 2026