Projects
8 projects
Video Real2Sim (VR2S)
6.8300 final project: uses a video model to hallucinate orbital observations from one populated-scene photo, then reconstructs the scene with a pose-free 3D stack.
Fast Humanoid Loco-Manipulation via Flow Matching
Flow Matching achieves 820 vs. 280 survival steps over DDPM at 5-NFE. Zero-shot loco-manipulation from walking-only data via classifier guidance.
RL vs. SFT for Mathematical Reasoning in LLMs
GMPO achieves 74.2% on GSM8K (vs. SFT 76.7%), demonstrating RL can match SFT without step-by-step supervision.
Enhancing Diffusion Models with RL & Adversarial Rewards
21.7% FID reduction via RL fine-tuning with adversarial reward signals. Plug-and-play for existing models.
PaperPlay: Hand-drawn Sketches to Playable Games
HackMIT 2025: 2nd Place, Modal Prize. Turn hand-drawn sketches into playable physics games with real-time AI commentary.
Consistent Local Video Editing via Attention Manipulation
Training-free framework for local video editing using BrushNet inpainting, DDIM inversion, and PerVFI for temporal coherence.
Daily Papers: Personalized ArXiv Research Digest
Agentic LLM pipeline for autonomous paper discovery with multi-step relevance filtering, ranking, and summarization.
Algorithm Design for the Metric k-Center Problem
Survey and evaluation framework. Proposed algorithms achieving empirical approximation ratio 1.049 (vs. SCR 1.064).