Yuchen Wang

CV

Research experience, systems work, and education.

Download PDF

Contact Information

Name Yuchen Wang
Email yuchenwang0303@gmail.com
Website https://yuchenwang3.github.io

Experience

  • 2026 - 2026

    Sunnyvale, CA, USA

    Research Intern — LLM Training Algorithms & Infrastructure
    Alibaba U.S.
    • Core contributor to Occamy-1.0 (35B-A3B); co-designed staged SFT/HDPO, model soup, and SAO with trainable context compaction and cross-segment credit assignment. 500K+ downloads across Occamy variants.
    • Architected environment-/capability-first data synthesis and multi-harness training infrastructure with state-aware graders and asynchronous actor–critic learning; the deduplicated SFT union spans 15K trajectories and 403M tokens.
    • Shipped Occamy with Accio; versus its Qwen3.6 base, the released model raised AutomationBench strict pass from 7.5% to 27.6% and Claw-Eval average from 69.5 to 82.2, while reducing Claw-Eval T/C full-trial time by 36.6%.
  • 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

  • 2025 - 2027

    Champaign, IL, USA

    M.S.
    University of Illinois Urbana-Champaign
    Computer Science
  • 2021 - 2025

    Beijing, China

    B.S.
    Peking University
    Intelligent Science and Technology (Artificial Intelligence)
    • Zhi Class (Elite Program), Top 10%

Open Source Projects

  • Open-Source LLM Systems Engineering

    RL infrastructure, long-context training, inference optimization, optimizer correctness, and training-data pipelines across NVIDIA NeMo, Megatron-LM, vLLM, SGLang, FlashInfer, and Hugging Face.

    • RL infrastructure: bypassed the Ray driver for NeMo RL teacher top-k transfers, achieving 4.44–5.31× speedup in controlled 8×B200 transfer measurements. Corrected async GRPO/distillation checkpoint-resume bookkeeping in TRL. Focused CPU checks passed, with end-to-end training-resume validation pending.
    • Long-context training: developing Qwen4-Exp integration in Megatron Bridge for QSA, per-layer embeddings, gated residuals, and HF conversion; GPU integration remains pending. Added fused GDN Q/K normalization and selective Mamba recompute in Megatron-LM. Made packed QSA selection document-local in mcore-bridge, 2.70× faster on an 8K synthetic selector case.
    • Inference memory and correctness: in-place BF16 MoE conversion in vLLM halved 120B TP2 conversion peak allocation from 7.88 to 3.94 GiB; fixed calibrated FP8 KV prefill in FlashInfer. Fixed Mamba cache-publication ordering in SGLang.
    • Optimizer correctness: fixed small-norm normalization in Muon and implemented checkpoint-safe Adam counters in Megatron-LM. CPU/DCP round-trip checks passed; full CUDA training validation remains pending.
    • Training data: preserved source sharding in HF Datasets streaming shuffle, restoring all four DataLoader workers on 19 source shards, with deterministic epoch and checkpoint-restoration coverage.

Projects

  • RareDx — Graph-Grounded RL for Rare-Disease Diagnosis

    Co-first author; advised by Prof. Tianyu Liu at Tsinghua University. Graph-grounded rewards for partial diagnostic matches.

    • Co-designed rewards using disease-ontology distance, biomedical similarity, and phenotype–disease links. With SFT and retrieval, 27B diagnosis Hit@10 increased from 28.53% to 40.76%.
  • CineFlow — Dependency-Driven Parallel Video Generation

    Co-first author; advised by Prof. Fan Lai at UIUC. Dependency-driven video diffusion using semantic DAGs, runtime trajectory fusion, online edge pruning, and critical-path scheduling without retraining.

    • Analyzed 1K+ MovieGenBench prompts; more than 68% exhibit non-linear dependencies.
    • Across Wan2.2-5B, CogVideoX-5B, and HunyuanVideo on 8×H100, achieved 1.7–5.5× end-to-end speedup, 1.30–2.02× lower P90 latency, and 5.4–17.3% higher VBench overall.
  • 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

    Advised by Prof. Yansong Feng at Peking University; Zero-RL → distilled-CoT SFT → GRPO for structured legal multiple-choice reasoning.

    • Reached 57.6% accuracy, outperforming SFT-only and RL-only baselines.

Academic Service

Awards

  •
    Merit Student
  •
    Zhi Class Scholarship
  •
    CMO (Provincial First Prize), 2020

Skills

Python, C++, CUDA, PyTorch, Ray, FSDP, Megatron-LM, NeMo RL, vLLM, SGLang, Nsight, RL, distributed training, diffusion transformers, ControlNet, 3D vision, and point clouds