Zhang Zhuocheng’s Homepage
I am a fifth-year Ph.D. student in the Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS), co-supervised by Prof. Yang Feng, and Prof. Min Zhang.
My research focuses on Retrieval-Augmented Generation (RAG) and Document-Level Machine Translation (DNMT), with a strong focus on leveraging Large Language Models (LLMs) to build advanced, real-world systems. I have published multiple papers in top-tier conferences such as ACL and EMNLP, and am an active contributor to the open-source community.
📣📣📣 I am currently seeking full-time opportunities in the areas of RAG, LLMs, and related fields, where I can apply my research experience to solve practical challenges in natural language processing.
📄 Research Experience
I am broadly interested in the field of natural language processing (NLP), with a focus on machine translation (MT) and large language models (LLMs). My research interests include: Document-Level Machine Translation (DNMT) and Retrieval-Augmented Generation (RAG).
Retrieval-Augmented Generation (RAG)
LevelRAG: A Hierarchical Multi-hop Hybrid Retrieval Framework This method comprises a high-level planner and multiple low-level retrievers, which respectively handle retrieval planning for complex queries and adaptive query optimization for individual retrievers. To further leverage the advantages of sparse retrieval in fine-grained search, we propose a Lucene-syntax-based rewriting strategy and apply it to the sparse retrievers. Our approach achieves significant performance improvements across multiple datasets.
FlexRAG: A Flexible and Efficient Retrieval-Augmented Generation System Led the development of the FlexRAG framework, which features high reproducibility, ease of use, and strong performance. It supports various RAG scenarios including text, multimodal, and web-based retrieval. The framework offers a complete pipeline and evaluation process, supports asynchronous processing and persistent caching, and is compatible with the Hugging Face ecosystem for rapid RAG system construction. This work has been accepted to ACL 2025 Demo Track.
Document-Level Machine Translation (DNMT)
Scaling Laws in Document-Level Machine Translation Through large-scale experiments, we found that excessively long contexts do not continuously improve translation quality, and that the optimal context length scales logarithmically with model size. Further analysis reveals that error accumulation is a key limiting factor in document-level translation performance. This research was published in Findings of EMNLP 2023.
Length Bias in Document-Level Machine Translation To address instability in traditional models when processing inputs of varying lengths, we propose methods including sentence-level sliding windows, attention length scaling, and dynamic training length sampling. These effectively mitigate length bias and enable models to handle sequences of arbitrary length in a stable manner. This research was published in Findings of EMNLP 2023.