I am Yanxin Chen (陈衍鑫), a Ph.D. student in Artificial Intelligence at Fudan University. I work on long-video understanding and video-language models.

My research spans model architecture and training systems. I am particularly interested in making video models reason over long contexts while remaining efficient and reliable at large training scale.

Research

Long-video reasoning

I study how models retrieve, connect, and reason over evidence distributed across long videos, moving beyond surface-level text similarity toward structured multimodal understanding.

Streaming video-language models

I explore cross-attention architectures and efficient image-text interaction for models that process long or continuously arriving video.

Large-scale training systems

I work on distributed VLM training, activation recomputation, FlashAttention, and system-level debugging across communication, memory, data, and framework boundaries.

Selected Work

MOSS-VL

OpenMOSS · Core Contributor

I contributed to the model's open-source release and was responsible for its four-stage pretraining, full SFT workflow, ablation studies, and large-scale training validation.

ProEchoMem

SIGIR 2026 · Co-first Author

A probe-echo memory framework for long-video understanding that constructs structured episodic traces, activates relevant evidence, and synthesizes it for downstream reasoning.

Open-source framework integrations

LlamaFactory · ms-swift

I added first-class MOSS-VL training and inference support to both frameworks, covering multimodal data processing, LoRA and full-parameter training, checkpoint workflows, and targeted regression tests.

Background

Before my Ph.D., I conducted research with the CityU AML Lab and the Fudan NLP Group on multimodal learning, retrieval-augmented generation, long-video understanding, and video model training infrastructure.

Across these projects, I have worked from problem formulation and model design through implementation and large-scale training validation.

AI Infra study notes