CV

Education

Outstanding Graduate of Sun Yat-sen University (June 2026)
National Scholarship (2023-2024 Academic Year)
First-Class Excellent Student Scholarship (SYSU, 2022-2025, Every Academic Year)
China Robot Competition & RoboCup China Open: National Third Prize (Team Leader)
National College Students’ Mathematical Modeling Contest (Guangdong): Provincial Third Prize

Role: Research Intern Nov. 2025 – Jun. 2026
Summary: Worked on pre-training data construction and model training for LongCat-Next Unified Foundation Model, based on a unified discrete-token (Discrete Native Autoregressive) architecture. Contributed to multimodal understanding, generation, and interleaved image-text generation.
Contributions: Built scalable pipelines for Caption image-text data, Knowledge-rich Entity data, UI Agent Caption & Grounding data, and interleaved image-text trajectory data. Expanded the Caption dataset to 85M image-text pairs and implemented automated pipelines for large-scale data production and quality validation. Also contributed to model pre-training, trajectory filtering, and data quality control.
Publication: Co-authored LongCat-Next: Lexicalizing Modalities as Discrete Tokens.

Summary: Systematically studies how teacher update schedules affect long-horizon stability in Self On-Policy Distillation. Identifies reference collapse and teacher contamination under different update mechanisms, and proposes Consolidation-Gated Teacher Refresh (CGTR), a state-aware teacher refresh strategy that preserves isolation periods while preventing unstable student snapshots from being copied to the teacher. CGTR achieves zero collapse across four tasks with a single shared parameter set.
Status: First Author; Under Review at EMNLP 2026 (Long Paper).

Summary: Introduced MCPHallu, a benchmark for diagnosing hallucination failures in LLM agents operating under the Model Context Protocol (MCP). The benchmark covers 358 tasks across five domains and evaluates four failure types: Branch Collapse, Unreachable Goal, Tool Misuse, and Context Forgetting. Experiments on 14 frontier LLMs and nearly 5,000 execution trajectories reveal fine-grained reliability issues beyond overall task success rate.
Status: Co-first Author; Under Review at NeurIPS 2026 E&D Track.

Summary: Proposed Mamba Snake, a deep snake algorithm based on State Space Modeling (SSM) to address multi-scale structural heterogeneity in Unified Medical Image Segmentation (UMIS). Introduced a Mamba Evolution Block for spatiotemporal information aggregation and a dual-classification collaborative mechanism for improving micro-structure segmentation. Achieved a 3% mDice improvement over SOTA methods across five clinical datasets.
Status: Second Author; Accepted at ACM MM 2025 (CCF-A, Oral).

Summary: Introduced GAMED-Snake, a deep snake architecture that combines gradient-aware differential convolution, a Distance Energy Map Prior (DEMP), and cross-attention to model dynamic features across adjacent iterations. The proposed approach improves semantic segmentation by alleviating misclassification and mask-hole problems, achieving an approximately 2% improvement in mDice.
Status: Co-first Author; Accepted at ICME 2025 (CCF-B).

Summary: Proposed the first multimodal state-space deep snake framework for medical image segmentation. TEAMS integrates a Spatiotemporal Snake Evolution Strategy (SSES), Contour Morphology-Aware Module (CMAM), and Text-driven Collaborative Dual-Head Snake (TCDHS) mechanism to improve segmentation of complex anatomical structures and enhance cross-modal generalization.
Status: Third Author; Accepted at Medical Image Analysis (IF=14.0).