Curriculum Vitae

Education

Long Video Understanding: Retrieval-augmented generation for long-video semantic understanding, with an emphasis on moving beyond shallow text similarity toward structured, multi-modal reasoning.
Video Large Models (VLMs): Architecture design for efficient image-text interaction, streaming video inference, and long-context video modeling.
Training Infrastructure: Large-scale distributed VLM training, parallel strategy design, activation recomputation, FlashAttention optimization, and cluster-level debugging.

ProEchoMem: Enhancing Long Video Understanding via Multi-Trace Probe-Echo Memory
- Co-first Author SIGIR 2026
- Proposed the PN-RAG framework to reveal implicit knowledge connections in long videos through probe generation, structured annotation, and fine-grained reranking.

VLM Architecture Exploration
- Designed cross-attention-based image-text interaction modules to reduce the computational load of full self-attention in video models.
- Added cross_gate mechanisms to improve training stability for cross-modal interaction.
Training Infrastructure Optimization
- Adapted Megatron-style parallel strategies and activation recomputation for cross-attention VLM training.
- Modified FlashAttention-3 support for cross_attention_mask, improving training efficiency and reducing memory pressure for long-video inputs.
Large-Scale MOSS-VL Training
- Core contributor to MOSS-VL, released in April.
- Responsible for the four-stage pretraining process and the full SFT workflow, including ablation experiments and formal training runs.
- Experienced with thousand-GPU cluster training and debugging training-framework, data, NCCL communication, and deadlock issues.

MOSS-VL Core Contributor
- Contributed to the training and open-source release of MOSS-VL, leading its four-stage pretraining, full SFT workflow, and large-scale training validation.
LlamaFactory: MOSS-VL Integration Merged to main
- Integrated MOSS-VL into LlamaFactory with image, video, and mixed-modal training and inference, plus full-parameter, freeze, and LoRA fine-tuning; validated 30 targeted tests and end-to-end training, resume, merge, and standalone inference workflows. Merged into upstream main.
ms-swift: MOSS-VL Integration Merged to main
- Added MOSS-VL as a first-class multimodal model on the Transformers backend, supporting LoRA/full-parameter SFT, mixed-media batches, and complete checkpoint workflows; validated with 16 targeted tests, H200 smoke runs, and 8×H200 ZeRO-3 full-parameter training. Merged into upstream main.

AI Algorithms: Large Language Models (LLMs), Video Large Models (VLMs), Long Video Understanding, RAG, SFT/RL Post-training, Linear Attention.
Training Infrastructure: Large-scale distributed training, Megatron-LM, TransformerEngine, DeepSpeed, FlashAttention/FlashAttention-3, TP/PP/DP parallelism, activation recomputation.
Engineering: Python, C++ (STL), PyTorch, Linux/Bash, cluster debugging, NCCL communication bottlenecks, deadlock resolution.

Efficient VLM Training: Further explore hybrid parallelism and activation recomputation to extend the trainable context length and frame limit of long-video VLMs.
VLM Post-training: Study SFT and RL training for long-video tasks, with a focus on better alignment with practical user needs.
Model Architecture: Explore low-precision training and hybrid architectures that combine linear attention with existing VLM backbones.