Curriculum Vitae
Research in omni-modal modeling across language, vision, audio, and video, and full-duplex human–AI interaction: systems that listen while responding and handle interruptions. Experience spans method design, model pre-training and post-training, and inference evaluation.
Education
Xiamen University
MAC Lab · Artificial Intelligence · M.S. student · 2025.09 - Present
Advisor: Prof. Xiawu Zheng
Sichuan Agricultural University
College of Information Engineering · Data Science and Big Data Technology · B.S. · 2021.09 - 2025.06
Selected Research
ICML 2026 · CCF-A · First Author. Training-free omni-modal orchestration with LLM routing, full-duplex streaming speech interaction, and cross-modal memory.
arXiv 2026 · Under review · First author. Offline evaluation of speaker identification, turn-entry decisions, and response generation using query-time-bounded evidence; 2,000 perception items, 200 interaction-generation items, and 11 evaluated omni models.
CVPR 2026 · CCF-A · Fourth Author. Training-free long-video frame selection with wavelet-based semantic-boundary detection and adaptive frame allocation.
ICML 2026 · CCF-A · Fourth Author. Automatic generation, augmentation, verification, and analysis for formally grounded abstract reasoning benchmarks.
SCI · JCR Q1 · Second Author. CBAM-enhanced YOLOv7 for dense hemp duck counting and intelligent livestock monitoring.
arXiv 2026 · Co-author. Continuous-audio evaluation of when an assistant should answer, stay silent, or stop speaking in multi-party full-duplex dialogue.
arXiv 2026 · Co-author. Diagnoses whether decisions depend on multimodal evidence and rebalances modality contributions during inference without additional training.
arXiv 2026 · Co-author. Training-free video token compression using wavelets to separate temporal redundancy from spatial detail.
arXiv 2026 · Co-author. A unified model for matching identities across images, infrared, sketches, and text, with knowledge distillation for edge deployment.
Honors and Awards
- 1China International College Students' Innovation Competition · National Bronze Award
- 2Sichuan Gold Award
- 3Challenge Cup · University Champion
- 4Challenge Cup · Sichuan Special Prize
- 5Challenge Cup · National Bronze Award
- 1National Special Prize
- 2C-problem top-score track
- 3First responsible student
- 1Challenge Cup Entrepreneurship Plan Competition · National Bronze Award
- 2Sichuan Internet+ Competition · Gold Award
- 3Second responsible student
- 1National Entrepreneurship Training Program · First Author · 2023
- 2National Innovation Training Program · Second Author · 2022
- 1China-US Young Maker Competition · National Third Prize
- 2iCAN College Student Innovation and Entrepreneurship Competition · National Third Prize
Academic Service
International Conference on Learning Representations (ICLR 2027)
国际学习表征会议（ICLR 2027）
33rd International Conference on MultiMedia Modeling (MMM 2027)
第 33 届国际多媒体建模会议（MMM 2027）
Conference on Neural Information Processing Systems (NeurIPS)
神经信息处理系统会议（NeurIPS）
Projects
Founder / first responsible student. Built a cloud-edge visual perception system for intelligent duck farming, disease monitoring, density estimation, and production deployment with an industry partner.
Second responsible student. Developed a multimodal pig-farming system with abnormal-sound recognition and a breeding resource case library.
Applied DBSCAN clustering, nonlinear programming, and particle swarm optimization to optimize discharge-point deployment.
Technical lead for Qwen1.5-32B based domain LLM applications, covering instruction data construction, LoRA SFT, prompt design, LangChain deployment, and scenario delivery for water-management and railway-maintenance workflows.
Open-source Contributions
Submitted over 4,000 lines of code to SGLang-Omni and vLLM-Omni, supporting multimodal inference and reproducible evaluation. Experience includes integrating specialized audio and vision models with language-model controllers, debugging semantic alignment between models, and validating changes through code review, regression checks, and GPU execution.
I use Cursor, Claude Code, and Codex to assist development, with code review and testing on target systems before delivery.