I am a master's student in Computer Science (AI) at the University of Southern California, graduating in May 2027.
I build and evaluate LLMs and agents. My interests include reliable agent systems, model alignment, diffusion language models, and physical reasoning.
Previously, I worked on machine learning at TikTok, General Motors, and Shanghai AI Laboratory, with research experience at Princeton and UC Berkeley.
email / scholar / linkedin / openreview / website
Seeking 2027 opportunities in Software Engineering (AI/ML) and Machine Learning Engineering.
research · experience · open-source projects
I am interested in how models learn, reason, and act reliably — from better training signals to agents that understand the world around them.
 |
PhysGraphNet: Physical-State Scene Graphs via Latent Graph Reasoning and Counterfactual Supervision Predicts physical-state scene graphs from an image and a goal, combining latent graph reasoning with counterfactual supervision to support physical reasoning. |
 |
VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation Uses geometry-derived preference signals to improve the 3D consistency of video diffusion models without human preference annotations. |
 |
PreDiff-LM: Pretrained Discrete Masked Diffusion Language Modeling with Hybrid Attention paper / openreview / code Studies hybrid attention as a way to adapt causal language models to bidirectional denoising, with controlled comparisons of generation quality and training efficiency. |
 |
Less Data, Better Alignment: Data-Centric Multi-Evaluator Agreement for Preference Optimization DMAPO selects 1,871 training examples from 54,236 candidates using multi-evaluator agreement and critique, examining data quality in preference optimization. |
 |
RoboDream: Compositional World Models for Scalable Robot Data Synthesis Composes robot motion, scene priors, and object priors to synthesize demonstrations in new environments and support data-efficient robot learning. |
 |
A Cookbook of 3D Vision: Data, Learning Paradigms, and Application A data-centric map of 3D representations, datasets, learning paradigms, and applications spanning reconstruction, generation, and world modeling. |
| When | Where & work |
|---|---|
| Jun–Sep 2026 | TikTok · ML Engineer Intern Post-training, multimodal moderation, and agent tooling for case replay and hard-example collection. |
| May–Sep 2026 | Princeton University · Research Assistant Causal evaluation and credit assignment for preventive actions in LLM agents. |
| Jan–Apr 2026 | General Motors · ML Engineer Intern Hybrid attention and confidence-aware decoding for pretrained diffusion language models. |
| Sep–Dec 2025 | UC Berkeley · Research Assistant Data-centric alignment through multi-evaluator agreement and preference optimization. |
| Jun–Sep 2025 | Shanghai AI Laboratory · ML Engineer Intern Agent orchestration, retrieval, and efficient vision-language model serving. |
- DMAPO — Multi-evaluator agreement, confidence filtering, and preference optimization.
- PreDiff-LM — Pretrained masked diffusion language modeling with hybrid attention.
- ACM-ICL — Calibrated multi-agent reasoning and vLLM multi-LoRA serving.
- latent-agent-team — Learned latent channels for budgeted communication between agents.
- paper_read — Research workflows and citation-verification tools.
More projects
- RAMTL — Shared-backbone, multi-role tool use.
- updr-reasoning — Uncertainty-guided debate and self-repair.
- DEAMS — Alignment across heterogeneous vision-language agents.
- PAGC — Grounded communication in cooperative multi-agent learning.
- KTM-WM — Kernel-based world models for agent planning.
Toolkit
| Area | Technologies |
|---|---|
| Software & services | Python · C++ · SQL · FastAPI · Docker · Kubernetes · AWS · Redis |
| Training & alignment | PyTorch · Transformers · TRL · PEFT · LoRA/QLoRA · DeepSpeed · FSDP |
| Agents, retrieval & inference | LangGraph · vLLM · FAISS · BM25 · Reranking |
| Evaluation & delivery | MLflow · Weights & Biases · GitHub Actions |