Hi, I'm Yuan Shan.

Publications

Working Papers

1. May 2025 – Oct 2025

   LLM Algorithm Engineer Intern

   Wenge Tech

   Work accepted at ICASSP 2026

   • Designed and built MEOW — the first metadata-driven, end-to-end LLM framework for academic survey outline generation, combining SFT with reinforcement learning (GRPO) under RLHF.
   • Constructed a complete evaluation stack: a public benchmark plus multi-dimensional metrics covering structure, content, and pragmatics for outline quality assessment.
   • Curated and distilled ~20k+ surveys from arXiv, bioRxiv, and medRxiv; built a Chain-of-Thought (CoT) annotation layer to guide structured taxonomy reasoning and improve survey coherence.
   • Optimized Qwen3-8B (full fine-tuning and LoRA) end-to-end; designed custom reward functions on structural distance and format compliance to align generation with human writing preferences.
   • Leveraged vLLM to accelerate both inference and RL training, achieving state-of-the-art performance — surpassing baselines such as SurveyX and outperforming strong LLMs (GPT-5 Nano, DeepSeek-R1) on the benchmark.
2. Jul 2024 – Aug 2024

   Data Engineering Intern

   Lakala Payment

   • Assisted in the development and maintenance of the company's internal data platform using Hadoop ecosystem tools.
   • Designed and developed the internal data platform's data quality monitoring system.
3. Jun 2023 – Oct 2023

   Research Intern

   Beijing Academy of Artificial Intelligence (BAAI)

   • Contributed to BAAI's 3D electron-microscopy (EM) processing toolchain, with a focus on advanced neuron-segmentation methodologies including PyTorch Connectomics (PyTC) and Local Shape Descriptors (LSD).
   • Built reproducible training pipelines for large-volume connectomics data, covering preprocessing, augmentation, and patch-based inference for cubic-micron EM volumes.
   • Optimized segmentation models via systematic hyperparameter tuning combining Grid Search and Bayesian Optimization, materially improving downstream segmentation quality on benchmark volumes.
4. Jun 2022 – Aug 2022

   Machine Learning Engineering Intern

   Tencent · Tencent Video

   • Independently owned an end-to-end ML project to surface high-potential video categories on the Tencent Video platform — running the full pipeline from data extraction and cleaning, through algorithm design and evaluation, all the way to written reporting for product stakeholders.
   • Built scalable data pipelines on Tencent's internal task-scheduling platform, handling dependency resolution across heterogeneous data sources for a reproducible workflow.
   • Engineered an XGBoost-based scoring model to produce quantitative indicators of video-category "potential", surfacing under-served verticals with strong growth signal; tuned the model with cross-validated hyperparameter search.
   • Wrote production SQL and Python against large-scale internal data warehouses to extract, join, and transform features at scale.
   • Deployed Spark for distributed feature engineering and model scoring, keeping the pipeline tractable on volumes of up to 7,500,000 records.
   • Synthesized findings into two project reports that fed directly into downstream content-strategy and operational decisions on Tencent Video.