cv

M.S. in Robotics (MSR)

B.S. in Computer Science and Mathematics (double major)

• Overall GPA: 3.93 / 4.0
• Selected coursework
  • Graduate: Natural Language Processing with Representation Learning, Inference & Representation, Reinforcement Learning
  • Computer Science: Machine Learning, Parallel Computing, Algorithms, Operating Systems, Data Structures
  • Mathematics (Honors): Ordinary Differential Equations, Analysis, Numerical Analysis, Linear Algebra
  • Mathematics: Partial Differential Equations, Linear and Nonlinear Optimization, Probability & Statistics

Auxiliary Streams for Video-Diffusion World-Action Models

• Investigating whether training-time auxiliary structure can improve the policy of a frozen Wan2.2-5B video-diffusion world-action model, particularly its spatial out-of-distribution generalization.
• Added inference-free alignment teachers (VGGT, CoTracker, DINO/REPA) and co-denoised streams routed by a Mixture-of-LoRA-Experts, passing 3D structure from 2D video to the action head at test time.
• Earlier in the project, adapted pretrained video diffusion models with lightweight modifications so that generated video obeys physical laws, conditioning on physical constraints through a modified ControlNeXt and generating supervised training data with 3D physics engines (Genesis).

Physics-Informed Vision-Language-Action Models

• Integrating trajectory planning into the action-prediction objective of VLA models such as GR00T N1.5.
• Using a 4D encoder that fuses multiple views across multiple timesteps to propagate spatial and temporal information jointly, aiding action prediction via future latent alignment.
• Designing an implicit 2D-to-3D mapping whose past-to-current 3D alignment, guided by the injected past 2D trace, aids action prediction.
• Implementing a cascaded pipeline in which a predicted future trace conditions the downstream action predictor.

Atomic Environment Imaging for Efficient Machine-Learning Force Fields

• Using 3D visualizations of molecular structure to improve force-field prediction models.
• Built a general pipeline that renders predefined multi-view images of the local environment centered on each atom.
• Added image encoders for models with direct force prediction, and coordinate encoders that reconstruct those images for gradient-based models that take only coordinates as input.

TraceGen: World Modeling in 3D Trace-Space

• Addressed the small-data problem in manipulation with a compact 3D trace-space of scene-level trajectories, letting a world model predict future motion geometrically rather than in pixel space and learn from cross-embodiment, cross-environment, and cross-task video.
• Contributed to TraceForge, the data pipeline that converts heterogeneous human and robot video into consistent 3D traces, yielding a 123K-video, 1.8M observation–trace–language triplet pretraining corpus.
• Pretraining on that corpus reaches 80% success across four tasks from five target-robot videos, at 50–600× the inference speed of video-based world models, and 67.5% real-robot success from five uncalibrated handheld human videos.

Efficient Training for Small-Molecule Force Fields

• Generated large pools of low-precision molecular dynamics samples, then used submodular selection to pick diverse conformations for recomputation at high precision.
• Showed that training GemNet on points chosen by similarity-based submodular functions outperforms both random and equal-timestep sampling.
• Implemented graph VAEs to learn geometry-aware features for submodular selection, and adapted EGNN to predict a coarse force field, yielding force-field-aware selection features.

Cost-Aware Finetuning of Language Models for Chemical Reaction Prediction

• Finetuned pretrained BART-based models for chemical reaction prediction under a compute budget.
• Constructed USPTO-50K_γ, a dataset pairing LLM-generated predictions with experimental data.
• Compared finetuning strategies, including a from-scratch LoRA implementation.

Popularity Prediction of YouTube Videos

• Applied the SlowFast action-recognition model to derive action labels for downstream prediction.
• Trained an MLP over visual features and the first seven days of view counts to predict total views at 30 days.

• Summa Cum Laude, B.S., NYU Shanghai

• Recognition Award, NYU Shanghai

• Dean's Honor List, NYU Shanghai

• Dean's Undergraduate Research Fund (DURF)