Contact Information
| Name | Yuchen Wang |
| yuchenwang0303@gmail.com | |
| Website | https://yuchenwang3.github.io |
Experience
-
2026 - present Sunnyvale, CA, USA
Research Intern — Agentic LLM Post-Training
Alibaba Group (U.S.)
- Architected and delivered Occamy-1.0’s end-to-end post-training stack, advancing a 35B-A3B co-work model from Qwen3.6-35B-A3B through full-parameter SFT, uniform model soup, and GRPO/SAO reinforcement learning.
- Built a verifier-gated, multi-harness pipeline with immutable provenance, token-exact replay, state reconstruction, episode credit across context rewrites, and quarantine gates.
- Raised Combined ClawEval T/C Strict Pass@1/3 from 65.16/73.87% to 77.39/85.93%; reduced tokens per trajectory by 38.6% and trace wall time by 36.1% on the same frozen harness.
-
2025 - 2025 Beijing, China
Research Intern — 3D Vision and Generative AI
Freedo Technology
- Built an end-to-end depth-conditioned ControlNet pipeline for 3D reconstruction from noisy point clouds, improving geometric fidelity from 32.7% to 85.3%.
Education
Open Source Projects
-
Open-Source LLM Systems Engineering
Contributions across NVIDIA NeMo, Megatron-LM, vLLM/Vime, SGLang, and ModelScope.
- Muon optimizer stability, recompute correctness, sequence packing, and NCCL startup reliability.
- Long-context GatedDeltaNet kernel fusion, selective Mamba recompute, and RL rollout hardening.
Projects
-
CineFlow — Dependency-Driven Parallel Video Generation
Semantic dependency decomposition and scheduling for scalable, consistent video diffusion.
- Achieved 1.7–5.5× speedup and 17.3% higher visual quality across selected evaluations.
-
Dynamic Prefill Optimization via Adaptive Online Packing
AIMD control with live p95 TTFT feedback and greedy/DP prompt packing.
- Achieved up to 20% lower TTFT on production-style DynamoLLM traces.
-
FlashAttention-style CUDA Optimization for GPT-2
Tiled online softmax, memory-efficient attention, and kernel fusion.
- Reduced HBM traffic by roughly 10× and achieved up to 9% end-to-end speedup.
-
Reinforcement Learning for Legal Reasoning
Zero-RL → distilled-CoT SFT → GRPO for structured legal multiple-choice reasoning.
- Reached 57.6% accuracy, outperforming SFT-only and RL-only baselines.
Skills
Python, C++, CUDA, PyTorch, Hugging Face Transformers, reinforcement learning, diffusion transformers, 3D vision, Megatron-LM, NeMo RL, vLLM, SGLang, distributed training, and profiling