Research
Papers, ongoing work, and course research projects.
Hybrid Algorithms Consistently Beat Offline RLHF Methods
Benchmarked DPO, IPO, AOT, GRPO, DrGRPO, GSPO for online and offline baselines. Then created a hybrid algorithm with a scheduled AOT to DPO loss, confidence-weighted pairs, and reward-model reranking. 0.865 win rate vs. 0.763 for the best baseline.
WaveletFlow: Spatially-Localized Frequency Mixing for Flow Matching in PDE Turbulence Modeling
FourierFlow fights spectral bias by boosting high frequencies, but a Fourier mode lives everywhere in the domain. Replacing its Fourier mixing branch with a wavelet branch lets the model sharpen detail where the vortices are, cutting RMSE by 8.4% on PDEBench compressible Navier–Stokes.
Visual–Temporal Retrieval-Augmented Generation for Lecture Question Answering
Answers to lecture questions live in speech, slides, charts, and gestures. We bundle OCR, ASR, and VLM frame descriptions into time-aligned chunks and add a carry-forward rule for slides shown just before a window. On a new 156-question benchmark, lecture-filtered retrieval reaches 82% accuracy on DATA 100, up from 20% with no context, in about 3 s per question.
