Five papers on learning under imperfect supervision — selecting useful samples, adapting with scarce data, synthesizing missing examples, stabilizing preference optimization, and verifying temporal predictions in video. First or co-first author on all five: GAPO accepted to NeurIPS 2026, work published at ICML 2023 and AAAI 2024, and two preprints, including SUTURE on video temporal grounding.
Proposed a batch-conditioned adversarial anchor and anchor-gap reweighting for brittle preferences; improved AlpacaEval 2.0 LC win rate by 3.6 percentage points over SimPO (Mistral-7B).
Developed Bayesian prompt adaptation with data-dependent priors and Wasserstein gradient flow for posterior inference, improving few-shot generalization under distribution shift.
Read the paper → Code →Co-developed sharpness-aware acquisition through pseudo-label-based perturbed-loss estimation, with a theoretical optimization bound; evaluated classification, detection, and domain-adaptive segmentation.
Read the paper → Code →Introduced a group-conditioned verifier with an exact decomposition into IoU and a covariance correction; improved temporal grounding at every reported IoU threshold across five benchmarks.
Read the paper →When real defects are too rare to train on, generate them — diffusion synthesis that models the relationship between foreground defect and background surface explicitly, and argues the disentanglement theoretically rather than only showing it. The method behind the production system on the product page.
Read the paper →2026 · NVIDIA Nemotron Hackathon — Winner, Track B
Team winner in Track B (Domain-Specialized Model). Implemented the preference-optimization post-training (SimPO on Megatron-Bridge) behind the team's winning video content-safety VLM.
Safe diffusion remains in progress, extending the safety side of the moderation stack. Findings are withheld until the work is public.
Keeping a diffusion model from generating what it shouldn't.