Research portfolio · 2026
Yuchen (Ian) Wang
I build reliable learning and systems infrastructure for agentic language models, long-horizon tool use, efficient inference, and generative media.
- Current
- Research Scientist Intern Alibaba · Sunnyvale
- Graduate study
- M.S. Computer Science University of Illinois Urbana-Champaign
- Foundation
- Zhi Class · Elite Program Peking University
Research interests: LLMs, AI agents, reinforcement learning, ML systems, diffusion models, and video generation.
Selected Research
Paper and report previews link to the complete PDF.
CineFlow: Parallel Video Generation through Semantic Dependency Scheduling
Research paper · 2026
Evaluated on Wan2.2-5B, CogVideoX-5B, and HunyuanVideo across 8×H100s, CineFlow achieved 1.7–5.5× end-to-end speedups, 5.4–17.3% higher VBench overall scores, and 1.30–2.02× lower P90 latency in scheduler comparisons.
- 5.5×
- peak speedup
- +17.3%
- VBench overall
- 2.02×
- lower P90 latency
Dynamic Prefill Optimization for LLM Inference via Adaptive Online Packing
Technical report · UIUC · 2025–2026
A workload-aware queue layer for continuous-batching inference engines. The design combines an AIMD controller driven by p95 time-to-first-token feedback with length-aware greedy and dynamic-programming prompt packing, reducing TTFT by up to 20% on production-style DynamoLLM traces.
- −20%
- TTFT
- p95
- control signal
- Greedy + DP
- online packing
Open-Source Contributions
Selected upstream work in large-scale training and inference systems.
NVIDIA-NeMo / Emerging-Optimizers 269 stars · Python
Keep Newton–Schulz scale-invariant for small-norm Muon inputs
Numerically stable optimizer behavior without changing the large-norm path.
vLLM / Vime 442 stars · Python
Forward recompute flags and fix hybrid-model runtime paths
Recompute propagation plus targeted fixes for Mamba-attention-MoE training.
NVIDIA / Megatron-LM 17,647 stars · Python
Fuse GatedDeltaNet Q/K normalization into the kernel
Long-context performance work for 128K-token supervised fine-tuning.
NVIDIA / Megatron-LM 17,647 stars · Python
Add selective Mamba recomputation
Memory-saving recompute controls validated on an 8×B200 training setup.
NVIDIA-NeMo / RL 1,966 stars · Python
Sanitize non-finite rollout log probabilities
Fail-safe handling for unstable values in the asynchronous vLLM worker.
sgl-project / SGLang 32,603 stars · Python
Make hybrid-model weight reload checks safer
Skip intentionally excluded tensors while resetting the weight-checker state.