About Me

Hi! I am Xinyi Huang (黄欣怡), a sophomore undergraduate student majoring in Information Security (Experimental Class) at the College of Computer Science & Fan-Gongxiu Honors College, Beijing University of Technology (BJUT).

I am currently a Research Assistant at the Beijing Institute of Artificial Intelligence, advised by Associate Researcher Jinduo Liu. My research focuses on Multimodal LLM Safety, Machine Unlearning, LLM Red-Teaming, and Multi-Agent Systems.

In the summer of 2025, I attended the Peking University Summer School (School of Computer Science) and participated in Nankai University's Xinya Program on vision-language model acceleration and efficient fine-tuning.

🔬 Research Interests

🛡️
MLLM Safety & Unlearning Connector-level intervention, cross-modal risk suppression, selective unlearning to reduce over-refusal in multimodal LLMs.
🔴
LLM Red-Teaming Budget-efficient safety evaluation, diverse policy violation discovery, cognitive-guided adversarial testing of frontier models.
🤝
Multi-Agent Systems Distributed coordination benchmarking, information-silo evaluation, scalable multi-agent collaboration under resource constraints.
👁️
Vision-Language Models Privacy inference risks in VLMs, multimodal agentic frameworks, cross-modal compositional threat analysis.

🧠 Selected Projects

Cross-Modal Safety Unlearning in Multimodal LLMs 2024 – Present
Lead Researcher · Advisor: Associate Researcher Jinduo Liu, BIAI
Designed a connector-localized safety unlearning framework that estimates a low-rank risk subspace from evidence-weighted activations and suppresses only risk-aligned directions, enabling selective forgetting without degrading benign utility. Evaluated across three benchmarks (SafeEraser, VLGuard, SIUO) using LLaVA-1.5-7B and InternVL2.5-8B.
  • Proposed Risk-Structured Fusion (RSF) for per-sample risk scoring via confidence-gated fusion of OCR, object tags, and connector activations.
  • Proposed Connector-Localized Selective Unlearning (CLSU) with iterative feedback correction for progressive risk boundary refinement.
  • Submitted to ACM MM 2026 (CCF-A) as first author.
Budget-Efficient LLM Safety Evaluation 2024 – Present
First Author · Advisor: Associate Researcher Jinduo Liu, BIAI
Reframed LLM safety testing as a budget-constrained failure discovery problem. Developed a framework combining cognitive profiling, Monte Carlo Tree Search, and reinforcement learning to discover diverse policy violations earlier and more efficiently than existing red-teaming baselines.
  • Introduced efficiency-oriented metrics (k-FDQ, NDA, CCR) to measure discovery timing and harm category coverage beyond traditional Attack Success Rate.
  • Evaluated on six frontier LLMs (GPT, Gemini, Claude, DeepSeek, Qwen, LLaMA) across HarmBench and StrongREJECT benchmarks.
  • Submitted to IJCAI ECAI 2026 as first author.
Private Attribute Profiling via Vision-Language Models 2024
3rd Author · Collaboration with Nanyang Technological University, Northwestern University
Constructed PAPI, the largest benchmark for multi-image private attribute profiling (2,510 images, 251 individuals, 12 privacy attributes). Co-developed HolmesEye, a hybrid VLM-LLM agentic framework that surpasses human-level accuracy in inferring abstract personal attributes from photo collections.
  • HolmesEye achieved 90.5% average accuracy, outperforming human analysts by 15.0% on abstract attributes.
  • Published at ACM MM 2025 (CCF-A).
AI Traffic Complaint Response System 2023 – 2024
Principal Investigator · Xinghuo Fund Project (BJUT)
Led a funded research project building an intelligent reply system for traffic domain complaints using large language models. Successfully concluded with full deliverables submitted.

📄 Publications

* equal contribution  ·  † corresponding author  ·  underline = myself  ·  Published papers listed first.

[1]
The Eye of Sherlock Holmes: Uncovering User Private Attribute Profiling via Vision-Language Model Agentic Framework Published
Proposes PAPI, a large-scale benchmark for multi-image privacy profiling, and HolmesEye, a VLM-LLM agentic framework that surpasses human-level accuracy in inferring abstract personal attributes.
Feiran Liu, Yuzhe Zhang, Xinyi Huang, Yinan Peng, Xinfeng Li†, Lixu Wang, Yutong Shen, Ranjie Duan, Simeng Qin, Xiaojun Jia, Qingsong Wen, Wei Dong
ACM International Conference on Multimedia (ACM MM 2025) · CCF-A
[2]
[Multimodal LLM Safety Unlearning] 1st Author Under Review
A connector-localized framework for cross-modal safety unlearning that estimates a low-rank risk subspace from evidence-weighted activations, enabling selective suppression of harmful directions while preserving benign utility.
Xinyi Huang, et al.
ACM International Conference on Multimedia (ACM MM 2026) · CCF-A
[3]
[LLM Safety Evaluation & Red-Teaming] 1st Author Under Review
Reframes LLM safety testing as budget-constrained failure discovery; introduces a cognitive-guided MCTS framework with efficiency-oriented metrics for diverse policy violation coverage under fixed query budgets.
Xinyi Huang, et al.
IJCAI ECAI 2026 · CCF-A (ranked CCF-A at submission; reclassified CCF-B in 2026 update)
[4]
[Multi-Agent Collaboration & Benchmarking] Under Review
Introduces a scalable, role-free benchmark for evaluating distributed coordination in LLM-based multi-agent systems under information silos, revealing a fundamental Communication-Reasoning Gap in current models.
[Authors], Xinyi Huang (4th author), et al.
ACL ARR 2026 · CCF-A
[5]
[LLM Refusal Robustness & Calibration] Under Review
Identifies history-induced over-refusal as a failure mode in safety-aligned LLMs and proposes a calibration framework using rank-1 LoRA to reduce false refusals under benign conversational histories while preserving unsafe refusal.
[Authors], Xinyi Huang (2nd author), et al.
ICIC 2026 · CCF-C

🏆 Selected Awards & Competitions

National Level
1st 睿抗机器人开发者大赛 (RobMaster Developer Competition), 2025 Lead PI
1st 全国大学生数字媒体科技作品及创意竞赛 (National Digital Media Innovation Competition), 2025 Lead PI
Excel. 中国机器人及人工智能大赛 National Excellence Award, 2025
Provincial / Ministerial Level
1st 全国大学生数学建模大赛 Beijing 1st Prize, 2025 Lead PI
1st 挑战杯"青聚AI"人工智能+专项赛 Provincial 1st Prize, 2025
Silver 中国国际大学生创新大赛 (Beijing Division) Provincial Silver, 2025
1st 北京大学生创新创业大赛（京彩大创）Provincial 1st Prize, 2025
3rd 中国机器人及人工智能大赛 (Beijing Division), 2025 Lead PI
3rd 第18届全国三维数字化创新设计大赛 (Beijing Division), 2025
2nd 全国大学生学术英语词汇竞赛, 2024
3rd 第九届"复旦社杯"大学生学术英语词汇赛, 2024

🌟 Honors

  • 三好学生 Outstanding Student (BJUT)
  • 优秀学生干部 Outstanding Student Cadre (BJUT)
  • 学习优秀奖 Academic Excellence Award (BJUT)
  • 创新创业奖 Innovation & Entrepreneurship Award (BJUT)
  • 优秀志愿者 Outstanding Volunteer (BJUT)
  • 五育突出表现奖 Five-Education Outstanding Performance · Fan-Gongxiu Honors College
  • 万集领导力奖（二等）Wanji Leadership Award (2nd Class) · Fan-Gongxiu Honors College
  • 服务突出贡献奖（二等）Outstanding Service Contribution (2nd Class) · Fan-Gongxiu Honors College
  • 马克思主义者培养工程班"新生英才计划"优秀学员 Outstanding Fellow, Marxist Training Program
  • 中国光大银行光合游学营 优秀实践生 & 优秀学员

📌 Service & Activities

Class President, Class 232501, BJUT (2023–2024)
Study Committee Member, Fan-Gongxiu Honors College, Class of 2023 (2023–Present)
Secretary, MoreFun Rubik's Cube Club, BJUT (2024–Present)
Student Assistant, One-Stop Student Community — organized 6 university-level events as lead organizer (2023–2024)
Academic Affairs Dept., University Science Association (2023–2024)
International Volunteer, Anan City, Tokushima, Japan
Student Delegate, 4th National Symposium on Top-Tier Undergraduate Cultivation, Qingdao
Student Representative, Undergraduate Education Assessment Expert Panel
Cumulative volunteer service: 100+ hours

🛠 Skills

Programming Python, C/C++, Java  ·  PyTorch, HuggingFace (research-level) English CET-4: 595  ·  CET-6: 544 Tools LaTeX, Git, Linux, CUDA
Last updated: April 2026  ·  Beijing, China