Yufei Luo

About

I study how to build reliable LLM agents for long-horizon tasks.

My research develops external mechanisms—including structured memory, execution-state modeling, and experience reuse—that help agents preserve task-relevant information and act consistently across extended interactions. My current work, MemSIF, transforms long-term interaction histories into structured representations and query-adaptive fact memories.

Previously, I developed robust learning methods for linguistic steganalysis under distribution shift and worked on multimodal and multi-turn intent recognition at Lenovo Research Institute. I completed an M.S. in Cyberspace Security at Beijing University of Posts and Telecommunications, after earning a B.S. in Mathematics and Applied Mathematics from Nanchang University.

Selected Publications

Google Scholar

arXiv
MemSIF: From Structured Interactions to Dual-Track Fact Memory for LLM Agents

Yufei Luo, Xiucheng Xu, and Zhen Yang

arXiv preprint arXiv:2608.01742, 2026

Introduces a structured interaction-to-fact memory framework that achieves the highest Total ACC across all evaluated settings on LoCoMo and LongMemEval-S.

Abs DOI arXiv Bib PDF Code

Long-term memory is critical for LLM agents operating over long-horizon interactions, yet existing memory systems exhibit Temporal–Structural Misalignment and Delayed Utility Manifestation. MemSIF combines Structured Interaction Memory with Dual-Track Fact Memory: CoreFact consolidates stable, schema-guided information at write time, while ActiveFact forms and promotes reusable facts when their utility emerges through later queries. Across five backbone LLMs, MemSIF achieves the highest Total ACC in all evaluated settings on LoCoMo and LongMemEval-S.
@article{luo2026memsif, title = {{MemSIF}: From Structured Interactions to Dual-Track Fact Memory for {LLM} Agents}, author = {Luo, Yufei and Xu, Xiucheng and Yang, Zhen}, journal = {arXiv preprint arXiv:2608.01742}, year = {2026}, doi = {10.48550/arXiv.2608.01742}, note = {<span class="publication-impact">Introduces a structured interaction-to-fact memory framework that achieves the highest Total ACC across all evaluated settings on LoCoMo and LongMemEval-S.</span>} }

TIFS

class-aware-adversarial-domain-adaptation-framework.png

Class-Aware Adversarial Unsupervised Domain Adaptation for Linguistic Steganalysis

Zhen Yang, Yufei Luo, Jinshuai Yang, and 3 more authors

IEEE Transactions on Information Forensics and Security, 2025

Develops class-aware adversarial adaptation to transfer linguistic steganalysis across domains while preserving class-discriminative structure.

DOI Bib HTML Code

@article{yang2025classaware,
  title = {Class-Aware Adversarial Unsupervised Domain Adaptation for Linguistic Steganalysis},
  author = {Yang, Zhen and Luo, Yufei and Yang, Jinshuai and Xu, Xin and Zhang, Ru and Huang, Yongfeng},
  journal = {IEEE Transactions on Information Forensics and Security},
  volume = {20},
  pages = {5181--5194},
  year = {2025},
  doi = {10.1109/TIFS.2025.3569409},
  note = {<span class="publication-impact">Develops class-aware adversarial adaptation to transfer linguistic steganalysis across domains while preserving class-discriminative structure.</span>}
}

ICASSP

Contrastive Hypersphere for One-Class Linguistic Steganalysis

Yufei Luo, Zhen Yang, Xin Xu, and 2 more authors

In ICASSP 2026 – 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

Develops contrastive hypersphere learning for one-class linguistic steganalysis when only cover-text examples are available during training.

DOI Bib HTML PDF Code

@inproceedings{luo2026contrastive,
  title = {Contrastive Hypersphere for One-Class Linguistic Steganalysis},
  author = {Luo, Yufei and Yang, Zhen and Xu, Xin and Wang, Yelei and Zhang, Ru},
  booktitle = {ICASSP 2026 -- 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages = {13602--13606},
  year = {2026},
  doi = {10.1109/ICASSP55912.2026.11462507},
  note = {<span class="publication-impact">Develops contrastive hypersphere learning for one-class linguistic steganalysis when only cover-text examples are available during training.</span>}
}

IH&MMSec

Clustering-Driven Pseudo-Labeling in Source-Free Domain Adaptation for Linguistic Steganalysis

Yufei Luo, Zhen Yang, Xin Xu, and 3 more authors

In Proceedings of the ACM Workshop on Information Hiding and Multimedia Security, 2025

Introduces clustering-driven pseudo-labeling for source-free adaptation, enabling linguistic steganalysis without access to the original source-domain data.

DOI Bib HTML PDF Code

@inproceedings{luo2025clustering,
  title = {Clustering-Driven Pseudo-Labeling in Source-Free Domain Adaptation for Linguistic Steganalysis},
  author = {Luo, Yufei and Yang, Zhen and Xu, Xin and Wang, Yelei and Zhang, Ru and Huang, Yongfeng},
  booktitle = {Proceedings of the ACM Workshop on Information Hiding and Multimedia Security},
  pages = {153--163},
  year = {2025},
  doi = {10.1145/3733102.3733114},
  note = {<span class="publication-impact">Introduces clustering-driven pseudo-labeling for source-free adaptation, enabling linguistic steganalysis without access to the original source-domain data.</span>}
}

Research Agenda

Research directions for reliable agents

I investigate how structured execution state, external memory, and reusable experience can improve agent reliability across long-horizon tasks.

Reliable long-horizon agents

Maintaining coherent decisions and behavior across extended tasks and interactions.

Structured memory

Organizing task-relevant facts, context, and execution progress into explicit state representations.

Experience reuse

Turning prior interactions into reusable knowledge, strategies, and capabilities.

Selected Honors

2026

Beijing Outstanding Graduate (Top 5%)

Beijing Municipal Education Commission
2025

National Scholarship for Graduate Students

Ministry of Education of China
2025

Open Source Security Award (Third Prize)

China Cybersecurity Association

Beyond Research

More about me

Outside research, I enjoy motorcycle travel, reading, basketball, and badminton. I also keep notes on books, journeys, and everyday experiences that shape how I think.

Contact

Interested in reliable long-horizon agents?

I am seeking Ph.D. opportunities beginning in Fall 2027.

luoyf@bupt.edu.cn