AI Researcher & Engineer at BAAI · Previously ByteDance and Meituan

I build dependable foundation models for specialized domains through post-training, retrieval agents, and multimodal reasoning. Explore the selected projects below for papers, code, datasets, and models.

Website · Google Scholar · Hugging Face · Email

• MechVQA / MechVL ICML 2026 — A benchmark and domain-specialized models for understanding mechanical drawings. Code · Data & models
• IAR 2026 preprint — Staged post-training to internalize document knowledge, align question answering, and recover general capabilities without retrieval.
• SPAR / SPARBench 2025 preprint — Multi-agent scholarly retrieval with an evaluation dataset. Code · Data
• SciSage / SurveyScope 2025 preprint — Multi-agent scientific survey generation and evaluation. Code · Data

More work on post-training: Wnuan · RAFT · SFTKey · MoSLD (COLING 2025).

Multimodal retrieval and domain models: ChartWalker · CareBot (AAAI 2025) · Aquila-Med.

Full research index · Research collaboration · Published patent applications

• CCI3.0-HQ — high-quality Chinese pre-training data · Paper
• IndustryCorpus2 — multilingual multi-industry pre-training corpus
• IndustryCorpus — the original multilingual multi-industry corpus
• Industry Instruction — multilingual multi-industry instruction data

These are contributions to projects maintained by their respective organizations and authors:

• FlagAI — Added the Aquila-SQL training, inference, and evaluation example. Contribution
• patent-disclosure-skill — Added inventor-based patent portfolio retrieval from CNIPA's publication system, with parsing, tests, and documentation. Contribution

• paper-reading-skill — Agent skill for visual, source-grounded paper reading: one standalone HTML atlas of methods, equations, and experiments.
• prompt-thinking-toolkit — A thinking-mode router for agents: 12 bounded prompt workflows (Socratic diagnosis, first principles, bidirectional steelman) packaged as a cross-agent SKILL.md for Codex, Claude Code, and Kimi Code.
• ai-avatar-skill — Purpose-driven AI avatars and profile pictures: Codex skill, ChatGPT prompts, examples, and crop previews.

• CHINESE-OCR — Chinese scene-text detection and recognition; legacy project.
• Image2Katex — Image-to-LaTeX recognition for printed and handwritten formulas.
• DKT-TensorFlow — Deep knowledge tracing in TensorFlow.