- Jasaxion一只大雄

主页

📫 Email

GitHub

Twitter/X

🙋‍ Zhan Shaoxiong (詹少雄)
📧 zhansx24@mails.tsinghua.edu.cn / jasaxion@gmail.com
📍 Shenzhen, China · Tsinghua SIGS

I am an M.S. student at the Knowledge Engineering Lab at Tsinghua Shenzhen International Graduate School, supervised by Prof. Haitao Zheng. My research interests lie in large language models, vision-language reasoning, information retrieval, and training data construction. I received my B.Eng. in Computer Science from Huazhong Agricultural University.

Recent Updates

• 2026.05 – 3ViewSense was accepted to ICML 2026 as an Oral / Spotlight presentation.

• 2025.12 – Joined JD Explore Academy as an LLM research intern.

• 2025.11 – MathSmith was accepted to AAAI 2026.

• 2025.11 – RoutingGen was accepted to AAAI 2026.

• 2025.10 – HSPIM was accepted to Information Sciences.

• 2025.07 – LexSemBridge was accepted to ECAI 2025.

• 2025.06 – IRSC Benchmark was accepted to NLPCC 2025.

• 2025.02 – QAEA-DR was accepted to IEEE TKDE.

Selected Publications

ICML 2026 Oral / Spotlight First Author Spatial Reasoning / VLM

3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models
Shaoxiong Zhan, Yanlin Lai, Zheng Liu, Zijian Lin, Hai Lin, Xiaodong Cai, Shen Li, Hai-Tao Zheng
Orthographic-view-based framework for spatial and mental perspective reasoning in vision-language models.

arXiv:2603.07751

AAAI 2026 First Author Data Synthesis / RL

MathSmith: Towards Extremely Hard Mathematical Reasoning by Forging Synthetic Problems with a Reinforced Policy
Shaoxiong Zhan, Yanlin Lai, Ziyu Lu, Dahua Lin, Ziqing Yang, Fei Tan
RL-based framework to forge olympiad-level math problems from concept-explanation pairs. 9.8%–18.1% relative gains on AIME/Olympiad benchmarks.

arXiv:2508.05592

ECAI 2025 Spotlight First Author Information Retrieval / Embedding

LexSemBridge: Fine-Grained Dense Representation Enhancement through Token-Aware Embedding Augmentation
Shaoxiong Zhan, Hai Lin, Hongming Tan, Xiaodong Cai, Hai-Tao Zheng
Lexical-semantic bridging to enhance fine-grained matching in dense retrieval without modifying backbone encoders.

arXiv:2508.17858

IEEE TKDE 2025 Co-first Author RAG / Dense Retrieval

QAEA-DR: A Unified Text Augmentation Framework for Dense Retrieval
Hongming Tan*, Shaoxiong Zhan*, Hai Lin, Hai-Tao Zheng, Wai Kin Chan
Unified text augmentation framework for dense retrieval via LLM-generated QA pairs and event structures.

DOI: 10.1109/TKDE.2025.3543203

Information Sciences 2025 Co-first Author Scientific Evaluation / LLM

A Hierarchical Framework for Measuring Scientific Paper Innovation via Large Language Models
Hongming Tan*, Shaoxiong Zhan*, Fengwei Jia, Hai-Tao Zheng, Wai Kin Chan
Hierarchical paper-section-QA decomposition framework for measuring research innovation with confidence-weighted aggregation.

arXiv:2504.14620

NLPCC 2025 Co-first Author RAG / Benchmark

IRSC: A Zero-shot Evaluation Benchmark for Information Retrieval through Semantic Comprehension in RAG Scenarios
Hai Lin*, Shaoxiong Zhan*, Junyou Su, Hai-Tao Zheng, Hui Wang
Zero-shot retrieval benchmark with five tasks and cross-lingual evaluation, introducing SSCI and RCCI metrics.

arXiv:2409.15763

AAAI 2026 Co-author Code Generation

IntentionChain-of-Thought Prompting with Dynamic Routing for Code Generation
Shen Li, Li Huang, Shaoxiong Zhan, Weifeng Sun, Tao Yin, Zhongxin Liu, Meng Yan
Difficulty-aware dynamic routing to avoid overthinking in code generation; 46% token reduction with SOTA performance.

arXiv:2512.14048

Honors and Awards

• National Scholarship in China (×3): 2021, 2022, 2023

• Gratitude Scholarship for Chinese Scientists, 2023

• MCM Finalist (Top 1%), 2023

• China University Big Data Challenge National Second Prize, 2022

• MathorCup National Second Prize, 2022

Research Experience

JD Explore Academy – LLM Algorithm Research, Foundation Model / Knowledge & Cognitive Data, Beijing

Dec 2025 – Present

Building pretraining data pipelines for the code domain, including high-quality code data collection for pretraining and mid-training stages.
Restructured GitHub commit data processing pipelines and studied how data distribution impacts code generation capability.
Proposed a multimodal issue localization benchmark, extending software engineering fault localization to scenarios involving UI screenshots and complex error logs.

SenseTime Research – LLM Algorithm Research, Model Foundation Group, Shenzhen

Mar – Oct 2025

Focused on math reasoning enhancement for LLM foundation models via data-centric approaches.
Designed RL-based hard-problem synthesis strategies, producing the MathSmith dataset and leading to a first-author paper at AAAI 2026.
Contributed to VLM foundation model iteration for the "photo-solve" product line. Identified VLM spatial reasoning gaps, leading to the first-author paper 3ViewSense.

Skills

Python C/C++ PyTorch Transformers vLLM LLM Training Reinforcement Learning Dense Retrieval Git Linux

English: CET-6 · Proficient in academic reading and writing
Tools: Hugging Face Transformers, vLLM, DeepSpeed, LLaMA-Factory, verl, FAISS, LangChain, Claude Code, Cursor, Codex

Teaching, Service and Leadership

• Teaching Assistant, Artificial Intelligence, Tsinghua University

• Head of the Youth Media Center in HZAU, 2022–2023

• Deputy Party Branch Secretary in HZAU, 2023–2024

• Class President in HZAU, 2023–2024

Miscellaneous

👋I'm a hands-on tech enthusiast who enjoys tinkering with gadgets and experimenting with new ideas💡–even if things break along the way. Outside of work, you'll find me playing badminton🏸, swimming🏊‍♀️, or dancing💃 (hiphop/kpop). Always happy to connect–feel free to add me on WeChat: Jasaxion_Taurus0405 🤝