PhD Student, Shenzhen University

Tengjin Weng

I work on multimodal large language models, with a focus on visual reasoning, fine-grained visual perception, and reliable MLLM evaluation and training. My recent projects study visual discrepancy sensitivity, numerical reasoning, and reinforcement learning for stronger multimodal reasoning.

Publications Google Scholar OpenReview GitHub

Research Interests

MLLM Evaluation Building benchmarks that reveal fine-grained visual perception and reasoning failures in multimodal models.

Visual Reasoning Studying visual discrepancy sensitivity, number sense, spatial reasoning, and perception-heavy problem solving.

MLLM Training Improving multimodal reasoning with verifiable supervision, reinforcement learning, and instruction-guided generation.

Selected Publications

CVPR 2026 Highlight First Author

OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models

Tengjin Weng, Wenhao Jiang, Jingyi Wang, Ming Li, Lin Ma, Zhong Ming

A controllable benchmark for testing fine-grained visual discrepancy sensitivity in multimodal large language models, together with OddGrid-GRPO for improving visual discrimination.

arXiv Project

ICCV 2025 First Author

VisNumBench: Evaluating Number Sense of Multimodal Large Language Models

Tengjin Weng, Jingyi Wang, Wenhao Jiang, Zhong Ming

A benchmark for evaluating visual number sense in multimodal large language models.

Paper Project

AIIM 2026 Co-first Author

Multi-annotation agreement and prediction consistency networks

Shuai Wang*, Tengjin Weng*, Jingyi Wang, Kai Zhao, Yang Shen, et al.

Semi-supervised medical image segmentation with ambiguous boundaries and multi-annotated data.

DOI OpenReview

BSPC 2024 First Author

Enhancing point annotations with superpixel and confident learning guided for improving semi-supervised OCT fluid segmentation

Tengjin Weng, Yang Shen, Kai Jin, Yaqi Wang, Zhiming Cheng, et al.

A semi-supervised OCT fluid segmentation method using point annotations, superpixels, and confident learning.

DOI arXiv

ECCV 2026

LlamaSeg: Image Segmentation via Autoregressive Mask Generation

Jiru Deng, Tengjin Weng, Tianyu Yang, Wenhan Luo, Zhiheng Li, Wenhao Jiang

A visual autoregressive framework that unifies image segmentation tasks through natural language instructions.

OpenReview

Education

PhD Student, Shenzhen University, 2024 - Present
MS Student, Zhejiang Sci-Tech University, 2021 - 2024
Undergraduate Student, Wenzhou University, 2017 - 2021