CV

Education

University of Illinois Urbana-Champaign (UIUC), Urbana, IL, USA
B.S. in Math & Computer Science, Expected 06/2026
- Relevant Courses: AI Efficiency: Sys. & Algor. (CS598), Machine Learning Systems (CS498), Distributed Systems (CS425), Parallel Programming (CS483), System Programming / Operating System

Wei, T., et al. RecScale: System-Aware Scaling Laws for Deep Learning Recommendation Models.
Under submission to the 21st ACM European Conference on Computer Systems (EuroSys 2026).
Fan, J., Wei, T., et al. Adaptive Divergence Regularized Policy Optimization for Fine-tuning Generative Models.
Accepted at the Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025).
Venkatraman, A., Cao, H., Wei, T., et al. MSAFlow: A Unified Approach for MSA Representation, Augmentation, and Family-based Protein Design.
Accepted at the NeurIPS AI4Science Workshop 2025; Under submission to the International Conference on Learning Representations (ICLR 2026).

RecScale: System-Aware Scaling Laws for Deep Learning Recommendation Models (Mar 2024 – Aug 2025)
- Developed a system-aware framework addressing memory/communication bottlenecks in scaling DLRMs.
- Achieved up to 16× memory reduction and 3.3× speedup on 64 GPUs, with preserved accuracy.
Adaptive Divergence Regularized Policy Optimization (ADRPO) (Mar 2025 – May 2025)
- Introduced adaptive regularization for RL fine-tuning of generative models.
- Enabled a 2B parameter SD3 model to outperform 4.8B–12B baselines in alignment, semantic control, and human preference metrics.
MSAFlow: Unified Approach for MSA Representation, Compression, and Augmentation (Mar 2025 – Aug 2025)
- Built a 130M-parameter generative autoencoder for multiple-sequence alignments.
- Reduced storage to 6.5% while improving AF3 zero-shot structure prediction (TM-score 0.62 vs. 0.55 baseline).

No experience yet.

To be updated.