Shuai Chen 陈帅

Algorithm Engineer · Xiaomi Corporation · XiaoAI Algorithm Strategy

📧 shuai_chen@stu.xidian.edu.cn · 📍 Beijing, China
🔗 Google Scholar · GitHub · CV

About

I am Shuai Chen, an M.S. student in Computer Science & Technology at Xidian University, currently working as an Algorithm Engineer on Xiaomi’s XiaoAI team (full-time since July 2026). From 2025, I interned on the same team, working on GUI multimodal content understanding for XiaoAI.

My research and engineering focus includes:

• Multimodal LLMs (M-LLM): fine-tuning, distillation, and reprogramming (GPT-2 / LLaMA-3.2 / BERT / Qwen2.5-VL)
• Reinforcement Learning (RL): GRPO and off-policy improvements; reward and loss design (ScreenR1-SFT / ScreenR1-COT)
• Agent / System Prompt Learning: agent self-evolution algorithms, training-free prompt optimization
• Cognitive Radio Signal processing (RSC): signal classification, time–frequency feature fusion, and specific emitter identification

Academic metrics: 6 SCI papers (several as student first author) · Google Scholar citations 78 · h-index 6 · i10-index 3 · 1 invention patent

Experience

Xiaomi · XiaoAI

Algorithm Engineer · Jul 2026 – present (full-time)
Algorithm Intern · Feb 2025 – Sep 2025

Project: GUI Multimodal Content Understanding for XiaoAI

In on-screen scenarios (smartphone, tablet, smart cockpit, etc.), extract 13 types of information — address, phone number, email, flight number, delivery tracking, calendar events, and more — from arbitrary UI screens, empowering Super XiaoAI with multimodal large language models.

• Problem analysis: online error cases follow a long-tail distribution (structured-knowledge errors, semantic-topic drift, etc.); simply scaling SFT data yields limited gains
• Scenario expansion: extended from address/phone to 13 scenarios; built training data via OCR + multimodal model joint labeling
• Model optimization: shifted from SFT → RL; distilled Qwen2.5-VL into ScreenR1-SFT, then used RL to obtain ScreenR1-COT with multimodal slow thinking; replaced on-policy GRPO with off-policy updates and continuous recovery of hard online samples
• Business results:
  • Phone call: 89% → 96.27%
  • Navigation: 88% → 95.28%
  • Pre-recognition: 90% → 95%
• Related results written up and submitted to ICLR 2026

China Electronics Technology Group (CETC) Intelligent Technology Research Institute

Project: Construction Technology for Electromagnetic Foundation Models · Jan 2024 – Dec 2024

• Signal reprogramming: encoded radio signals as tokens understandable by LLMs; used LLMs as backbones for downstream fine-tuning
• Pretraining task design, test-time scaling laws, data cleaning against contamination, and progressive scaling strategies
• Deployment: LoRA fine-tuning; 2×A100 half-precision pretraining → full-precision fine-tuning

AVIC Xi’an Flight Automatic Control Research Institute

Project: High-Precision Small-Target Detection System on UAV Platforms · Jun 2023 – Sep 2023

• Hybrid classical CV + deep learning detection: coarse localization to reduce compute, INS-aided dynamic correction, then heatmap-based deep refinement
• CPU–GPU operator migration, pruning, and knowledge distillation
• Achieved industrial metrics of 28.4 ms per frame and localization error ≤ 5 pixels

Education

Honors: Top 10% undergraduate GPA · National Scholarship · Nominee, Xidian University President’s Award · First Prize, University Mathematical Modeling Contest · Huawei Industry–Education Scholarship · National First Prize (Physical Category), Beidou Cup National Youth Science & Technology Innovation Competition

Campus experience: academic exchange with the University of Tokyo (Mar 2023 – Jun 2026) · mathematical modeling (university / national / MCM) · Xidian Stand-up Comedy Club · National College Student Innovation & Entrepreneurship Training Program (smart agriculture + hyperspectral analysis)

Selected Publications

Full list: Publications ｜ Google Scholar

1. Learning Temporal–Spectral Feature Fusion Representation for Radio Signal Classification — IEEE Transactions on Industrial Electronics (TII) (IF 11.8, SCI TOP 1), student first author, 2024
2. A Generative Self-supervised Framework for Cognitive Radio Leveraging Time-Frequency Features and Attention-based Fusion — IEEE Transactions on Wireless Communications (TWC) (IF 8.79, SCI TOP 1), student first author, 2024
3. RadioLLM: Introducing Large Language Model into Cognitive Radio via Hybrid Prompt and Token Reprogrammings — under review at IEEE JSAC (CCF-A) / arXiv:2501.17888, student first author
4. Enhanced Prediction of Partial Nitrification-Anammox Process… — Journal of Environmental Management (IF 7.9, SCI TOP 2), student first author, 2025
5. ACPO: Adaptive Curriculum Policy Optimization for Aligning Vision-Language Models in Complex Reasoning — arXiv:2510.00690, 2025 (in collaboration with Xiaomi XiaoAI)

Patent

• A Semi-supervised Building Change Detection Method and System Based on CutMix-ResNet — student first author, CN114494873 (invention patent)

Awards

Skills

• Deep learning frameworks: PyTorch · TensorFlow · MindSpore · HuggingFace
• Programming languages: Python · C++ · Matlab
• LLM stack: fine-tuning & reprogramming (GPT-2 / LLaMA-3.2 / BERT / Qwen2.5-VL) · data cleaning & curation · LoRA · RLHF / GRPO
• Other: academic writing & mathematical analysis
• Languages: English (CET-4 / CET-6; exempted in graduate school)
• Interests: fitness, hiking, swimming, stand-up comedy