Multimodal LLM Post-Training · Reinforcement Learning · Retrieval
Software Engineering student focused on multimodal model post-training and reliable evaluation.
Open to research and algorithm internship opportunities.
- AutoConcept — First-author, training-free concept-guided reranking; accepted at PRICAI 2026 (CCF-C). [arXiv] [Paper]
- AMEND — Multi-agent, constraint-aware text detoxification through specialized critics and minimal repair. [arXiv] [Paper]
- Ant Group — Multimodal LLM post-training intern; built
ms-swift-based training workflows and evaluated UI grounding models. - R&L Lab, Nanjing University — Integrated 127 algorithms and built a processing pipeline for 70,000 road-scene images.
- Incoming M.Eng. student at Southeast University; B.Eng. candidate at Soochow University.
- Python · C/C++ · PyTorch · ms-swift · SFT · GRPO · VLM · GUI grounding · distributed training