My name is Haotian Fang (WncFht), an undergraduate student at the UM-SJTU Joint Institute, Shanghai Jiao Tong University (SJTU), majoring in Electrical and Computer Engineering (ECE).
I'm passionate about Large Language Models — the full training stack from pre-training to post-training and reinforcement learning — and Embodied AI.
Currently, I'm a research assistant at the APEX Data & Knowledge Management Lab, SJTU. Over the past year (2025.09–2026.09) I was a research intern at Meituan, working on generative recommendation; before that, I was a research assistant at X-Lance Lab, SJTU, and an Agent Development Engineer at Shanghai Jiangran Technology, building a universal memory & context middleware for LLM agents.
I'm a strong believer in open-source culture. Check out my selected projects below!
I'm writing blogs on my website — notes and ideas from learning and research.
- HCGRec: Hint-Conditioned Generative Recommendation with Semantic IDs — CIKM 2026 (paper · code)
- Empowering LLM Tool Invocation with a Tool-Call Reward Model — ICLR 2026 Poster (paper)
| Area | Topics |
|---|---|
| LLM | Pre-training, Post-training, RL & Reward Models |
| Embodied AI | Embodied Agents, VLA, World Models |
| Agent Systems | Memory & Context, Tool Calling, RAG |
| Project | Role | Description |
|---|---|---|
| HCGRec | First Author | CIKM 2026 codebase — hint-aware GRPO + constrained decoding for semantic-ID generative recommendation |
| devin2api | Owner | Devin upstream → OpenAI Responses / Anthropic Messages compatible API proxy (Go) |
| agent-basic-skill | Owner | Reusable Codex/Claude skills and local agent configs |
| attention-calculator | Owner | Attention calculator reproduction + exact-arithmetic extension — machine-checkable definite-integral identities |
| zcode-teardown | Owner | Version-by-version teardown of ZCode IDE (55 versions) — signature snapshots & upload-mechanism analysis |
| Tencent-GRec | Owner | Tencent Ad Algorithm Competition 2025 — sequential recommendation + multimodal Transformer (HSTU) |
