AI Explorer · Software Engineering Undergraduate at Tongji University
I study how AI systems listen, look, reason, and revise — and I build tools that make those decisions easier to inspect.
- Multimodal perception: grounding objects and events across audio, vision, motion, and language.
- Reasoning behavior: understanding when longer reasoning helps a model and when it only creates drift.
- AI for research: inspectable agent workflows for reading, experimentation, and scientific communication.
PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs
Haojie Hu, Chenhao Dang, Yaojia Liu, Hengrui Kang, Conghui He, Weijia Li arXiv:2608.02218 · Paper · Code · Project page
Turns scientific papers into editable PowerPoint posters through capacity-aware slots, explicit design controls, and bounded visual-quality repair. Evaluated on 621 papers with an 81.3% print-ready rate at roughly $0.38 per poster.
Listening to the Motion: Audio-Conditioned Kinematic Verification for Robust Audio-Visual Segmentation
Under review · Code
Audio-visual segmentation leans on static visual saliency, so a silent guitar on a poster can outvote the one actually being played. KEVA makes audio interrogate the motion field before it is allowed to drive segmentation — spectral–kinematic alignment filters the flow, verified motion initializes the object queries, and a zero-initialized gate keeps the appearance prior when a source barely moves.
Under review · Code
Longer chain-of-thought is not uniformly better — forcing it on an already-clear query is an overthinking trap. The hidden state a model holds just before its first reasoning token turns out to encode whether reasoning will help, so a lightweight controller can pick a Zero / Short / Long budget before any reasoning is generated, without an extra forward pass.
An auditable NBA draft prediction agent that fuses talent, expert mocks, and market signals across 30 GM personas and 1,500 Monte Carlo scenarios. Built for the AWS Summit Shanghai 2026 hackathon; placed third and advanced to the Macau round.
Turns any voice into a playable instrument with Web Audio, an ESP32-S3, physical keys, and reactive LEDs. Excellence Award at Xiaohongshu AI Builder.
A Cocos2d-x systems project covering map interaction, character control, collision detection, inventory, and farming simulation mechanics.
| Award | Awarded by | Year | |
|---|---|---|---|
| 🏅 | National Scholarship | Ministry of Education of the People's Republic of China | 2025 |
| 🏅 | Qidi Scholarship | Tongji University | 2026 |
| 🏅 | First-Class Outstanding Student Scholarship | Tongji University | 2024 |
| 🏅 | Social Activity Scholarship | Tongji University | 2024, 2025 |
| ⭐ | Outstanding Student | Tongji University | 2024, 2025 |
| ⭐ | Computer Science Youth Pioneer | School of Computer Science and Technology, Tongji University | 2026 |
Looking for AI research and engineering internships around multimodal learning, LLM reasoning, research agents, and evaluation-heavy systems.
Python · PyTorch · C++ · TypeScript · Computer Vision · Multimodal Learning · Multi-Agent Systems · Research Tooling
I also enjoy playful systems, game mechanics, and projects that make difficult ideas visible.