Hanning Wang

theunknownthing@sjtu.edu.cn acm.sjtu.edu.cn/~theunknownthing/

Profile

Undergraduate in the ACM Honors Class (2024–Present) at Shanghai Jiao Tong University, working on data and systems for large-scale LLM pretraining, including data acquisition, processing, and synthesis, and efficient long-context training.

Education

  1. Shanghai Jiao Tong University, B.S. Computer Science (ACM Class) 2024 - Present

    Overall GPA: 4.0 / 4.3

Research Experience

  1. Shanghai Innovation Institute, Research Intern (Co-advised by Dr. Pengfei Liu and Dr. Siyuan Feng) Summer 2026–Present

    Building pretraining data pipelines and training infrastructure for thousands of GPUs, with contributions to 801-30B-A3B-Base, a 30B-parameter MoE model trained from scratch on 11.8T tokens.

    • Attention-balanced sampling: Implemented a sampler that balances attention workloads across data-parallel ranks, achieving 2.26× training throughput with 256K context; authored §7.1 of the technical report.
    • Data acquisition and retrieval: Crawled hundreds of terabytes of raw PDF data and helped scale the midtraining data retrieval pipeline.
    • Data processing and synthesis: Built PDF and code data processing and retrieval pipelines, and capability-oriented data synthesis pipelines.
    • Training infrastructure: Optimized context parallelism and developed faster training kernels. Worked closely with the pretraining team to investigate training issues and implement features.
    • Vision-language training: Contributed dataloader and model support and reviewed related pull requests.

  2. GAIR Lab, Research Intern (Advisor: Dr. Pengfei Liu) Summer 2025

    Researching Large Language Models with interests in mid-training and post-training stages. Working on providing efficient training data which incentivizes model's agentic capabilities.

Publications & Technical Reports

  1. No Second Run: Does RSI Stop at Frontier Pre-training?

    SII & Xinxu AI (Hanning Wang, contributor).

    Technical report, 2026. Project.

  2. davinci-dev: Agent-native mid-training for software engineering

    Ji Zeng, Dayuan Fu, Tiantian Mi, Yumin Zhuang, Yaxing Huang, Xuefeng Li, Lyumanshan Ye, Muhang Xie, Qishuo Hua, Zhen Huang, Mohan Jiang, Hanning Wang, Jifan Lin, Yang Xiao, Jie Sun, Yunze Wu, Pengfei Liu.

    ICML'26 Spotlight. Code.

Teaching Experience

  1. PPCA Networking, Track Lead (course repository) Summer 2026

    • Led the PPCA networking track and coordinated its course delivery.
    • Designed and scaffolded a more diverse set of lab projects, including QUIC congestion control and MiniCaddy.
    • Led a hands-on session on vibe coding, introducing students to AI-assisted programming workflows and fundamentals.

  2. Teaching Assistant, Data Structure (C++) Spring 2026

    • Prepared and delivered a lecture on segment trees and lowest common ancestors (LCA).
    • Co-designed the new lab project deque with Yuetian Wang, including the project specification, implementation requirements, and grading tests for students to practice C++ container design.

  3. Teaching Assistant, Programming (C++) Fall 2025

    • Gave lectures on shell usage and Git workflows, and supported students in hands-on sessions.
    • Provided an interactive learning platform (adapted from pwn.college) with challenge-based exercises to reinforce shell and Git concepts. Open-sourced our platform on: acmdojo and shell-dojo-core.

Selected Projects

  1. R10K-CPU Fall 2025

    An RV32I CPU with MIPS R10K-like out-of-order execution, implemented with assassyn. Course project for Computer Systems (Fall 2025) at Shanghai Jiao Tong University, developed by Zhaoyuan Wan and Hanning Wang.

  2. RCompiler Fall 2025 & Spring 2026

    A C++ implementation of a compiler for a subset of the Rust programming language, with advanced IR optimizations like mem2reg, SCCP, ADCE, strength reduction, and function inlining.

  3. Networking Systems Summer 2025

    • Transparent QUIC Proxy: Implemented a proxy layer to handle QUIC session forwarding and address translation.
    • NeoFRP: Built a lightweight reverse proxy (Fast Reverse Proxy style) for tunneling internal services. 70% faster than original FRP in throughput under same setting.

  4. vscode-acmoj Spring 2025

    A simple integration of ACMOJ into VSCode, which allows you to browse problemsets (contests/homework), view problems, submit your code, and check submission results without leaving your editor.


Contributed

  1. ACM Class Online Judge, programmer and operations Aug 2025–Present

    • Maintainer of online judge service, used by 9000+ students throughout the campus.
    • Contributed to judge reliability, web UI maintenance, deployment documentation, and semester course-operation workflows.

  2. volcengine/verl

    • Fixed a bug in multi-turn tool-calling message handling.
    • Opened issue and flagged a Qwen3 chat-template bug during rollout.

Skills

Selected Coursework

All courses below: 4.3 / 4.3