Yihao HU

Yihao HU

CS Undergraduate @ Hainan University | Research Intern @ Meituan & Alibaba Amap

About Me

Hello! I’m Yihao HU (胡益豪), a third-year undergraduate student majoring in Computer Science at Hainan University (GPA: 3.93/4.0, Top 1%, Rank 2/207), expected to graduate in 2026. My research focuses on Multimodal AI, Reasoning LLM/VLM, AI Agent, and Computer Vision.

I am currently an Algorithm Intern at Meituan (Core R&D Platform, Native Multimodal LLM - LongCat) and an AI Agent & LLM Alignment Intern at Alibaba Group, Amap (高德). Previously, I was a Research Assistant at the Digital Media Computing & Design Lab, Zhejiang University (Prof. Juncheng LI).

I have published/submitted papers at NeurIPS 2025, ACL 2026, ICML 2026, and top journals (Q1), and have received 7+ national/international competition awards including the National Scholarship.

🎓 Education

  • 2022.09 - Present, B.Eng. in Computer Science, Hainan University (海南大学), Hainan, China.
    • GPA: 3.93/4.0 (Top 1%, Rank 2/207)
    • Key Courses: Data Structures (97), Data Mining (97), Machine Learning (95), Introduction to AI (95), Operating Systems (95), OOP (98), Database Design (96)

💼 Research & Internship Experience

  • 2025.10 - Present, Algorithm Intern (Native Multimodal LLM - LongCat), Meituan (美团), Core R&D Platform, Shenzhen, China.
  • Ongoing, AI Agent & LLM Alignment Intern, Alibaba Group, Amap (高德), Shenzhen, China.
  • 2025.02 - 2025.06, Research Assistant, Digital Media Computing & Design Lab, Zhejiang University (Prof. Juncheng LI).
  • 2023.06 - Present, Undergraduate Research Team Member (Group Leader), Prof. Xiaodong BAI’s Lab, Hainan University.
  • 2024.10 - 2025.10, Group Leader, National-Level College Student Innovation & Entrepreneurship Training Program.

📜 Publications

2026 (Submitted)

  • Y. Hu et al., “VulThinker: Deep Reasoning for Vulnerability Detection”, ACL 2026, Submitted. (1st author, CCF-A)
  • Y. Hu et al., “OmniVideo-R1: Intention-driven Deep Fusion for Audio-Visual Reasoning”, ICML 2026, Submitted. (2nd author, CCF-A)
  • X. Yu et al., “Dual Latent Memory for Visual Multi-agent System”, ICML 2026, Submitted. [arXiv] (CCF-A)
  • “HoloRoom: Holistic and Compositional 3D Scene Generation via Global-to-Local Assembly”, CVPR 2026, Submitted. (CCF-A)

2025

  • M. Gao et al., “Counterfactual Evolution of Multimodal Datasets via Visual Programming”, NeurIPS 2025. (CCF-A)
  • Y. Hu et al., “SDE-DET: A Precision Network for Shatian Pomelo Detection in Complex Orchard Environments”, Smart Agriculture Technology, Minor Revision, 2025. (1st author, Q1, IF: 5.7)
  • Y. Hu et al., “A Multi-Strategy Framework for Enhancing Shatian Pomelo Detection in Real-World Orchards”, Engineering Applications of Artificial Intelligence, Second Review, 2025. (Co-1st author, Q1, IF: 7.7)

🔬 Research Projects

Project 1: VulThinker — Deep Reasoning for Vulnerability Detection (2025.06 - 2025.10)

  • Reformulated black-box binary detection into a verifiable causal reasoning framework via multi-view SFT with forward/backward counterfactual reasoning and generative repair.
  • Introduced two-stage curriculum RL with GRPO and reward design based on lexical verification and soft evidence matching to suppress logical hallucination.
  • Built VulReason-Bench (1850 expert-validated samples); achieved 83.5%/67.6% accuracy on BenchVul and VulReason-Bench, discovering 18 real-world 0-day vulnerabilities.

Project 2: DeepOmni-R1 — Audio-Visual Reasoning (2025.10 - Present)

  • Proposed “DeepOmni-R1”, a Zero-RL framework enabling reasoning without process annotation via intention-chain self-supervision and two-stage GSPO-based alignment.
  • Designed DeepFusion contrastive reward to maximize multimodal gain over unimodal reasoning.
  • Built an 88K audio-visual instruction dataset; achieved +7.8% on OmniVideoBench (SOTA), outperforming Gemini-3-Pro on Daily-Omni.

Project 3: SCOPE — Counterfactual Evolution via Visual Programming (2025.02 - 2025.06, ZJU)

  • Proposed SCOPE, a symbolic visual programming framework that converts implicit reasoning into executable Python code via an “Abduct-Intervene-Predict” loop.
  • Constructed SCOPE-Train and SCOPE-Test benchmarks; designed “Memory and Attention Path Learning” for structured difficulty progression.
  • Published at NeurIPS 2025 (CCF-A).

Project 4: SDE-DET — Pomelo Detection (National Innovation Program, 2024.10 - 2025.10)

  • Proposed a lightweight SDE-DET with tar Block for high-dimensional feature mapping and fine-grained texture preservation.
  • Integrated Deformable Attention and Efficient Multi-Scale Attention for discriminative representation.
  • Constructed STP-AgriData dataset; achieved 83.8% accuracy with 3.29M parameters (32.4 GFLOPs), outperforming YOLOv10 and RT-DETR.

Project 5: IMF-ITD — Interpretable Image-Text Description (2023.06 - Present)

  • Designed an interpretable vision-language generation framework integrating CLIP, BLIP2, and ConvNeXt-XXLarge.
  • Applied LoRA-based parameter-efficient fine-tuning and attention weight re-calibration for improved modality interaction.

🌟 Honors & Awards

🏅 Scholarships & Honors

  • National Scholarship (国家奖学金, 50th among all undergraduates in the whole school), 2025.10
  • “WuXu” Scholarship (1/207), 2024.10
  • First-Class Academic Scholarship, Hainan University (Top 10 in Department), 2023.10
  • Merit Student, Hainan University (80th among all undergraduates), 2023.12

🥇 Competition Awards

  • International Finals – First Prize, 5th Global Campus AI Algorithm Elite Competition (Stable Diffusion Prompt Optimization Track), 2023.12
  • National First Prize, 15th National College Mathematics Competition, 2023.12
  • National Second Prize, 2023 China Collegiate Computer Programming Competition, 2024.01
  • International Finals – Second Prize, 6th Global Campus AI Algorithm Elite Competition (AI + New Discipline Track), 2024.12
  • International Finals – Third Prize, 6th Global Campus AI Algorithm Elite Competition (AI + New Medical Track), 2024.12
  • National Finals – Second Prize, 12th National College Digital Media Technology & Creativity Competition, 2024.12
  • National Finals – Second Prize, 27th China Robot & AI Competition, 2025.08
  • National Finals – Third Prize, 27th China Robot & AI Competition, 2025.08

🛠️ Skills

  • Programming: C/C++ (Proficient), Python (Proficient), Matlab (Familiar)
  • Tools: Tableau, SPSS, AutoCAD, CATIA, Solidworks, Hyperworks
  • Writing: LaTeX (Proficient), Microsoft Office (Familiar)
  • Research Areas: Multimodal LLM, Reasoning VLM, AI Agent, Computer Vision, RLHF/DPO