Mingyang Song

Mingyang Song

School of Computer Science,
Fudan University

About me

I am a third-year Ph.D. candidate at the School of Computer Science, Fudan University, advised by Prof. Yu Cheng. I received my B.Eng. in Computer Science and Technology from Harbin Institute of Technology in 2023. From January to May 2026 I was a visiting student at the University of Illinois Urbana-Champaign (UIUC), and I have collaborated extensively with the Shanghai Artificial Intelligence Laboratory and The Chinese University of Hong Kong throughout my Ph.D.

My research aims to build reliable large vision–language models that can perceive, plan, and reason in real-world settings. The work spans four interlocking threads: (i) visual agents and multimodal tool use, (ii) multi-step and step-level reasoning for LLMs and VLMs, (iii) hallucination mitigation, and (iv) long-tail data calibration. My papers have appeared at top-tier venues including ICLR, CVPR, and ACL — representative projects include AdaReasoner (ICLR 2026), PRMBench (ACL 2025), and From Head to Tail (CVPR 2025).

I am currently a research intern at the Multimodal Model Department, Tencent Hunyuan, jointly selected for Tencent's “Project Up” Talent Program (青云计划) and the Rhino Bird Elite Talent Program (犀牛鸟精英人才计划), where I work on multimodal foundation models and agentic reasoning. I also serve as a reviewer for NeurIPS, CVPR, ACL, and TPAMI.

I welcome research discussions and collaborations from both academia and industry. For papers and code, please see the Publications section; for collaboration, recruiting, or academic-service inquiries, feel free to reach me via email.

News

  • [2026] Joined the Multimodal Model Department, Tencent Hunyuan as a research intern under the “Project Up” (青云计划) and Rhino Bird Elite Talent (犀牛鸟精英人才计划) programs.
  • [Jan. – May 2026] Visiting Student at the University of Illinois Urbana-Champaign (UIUC).
  • [2026] Our paper AdaReasoner has been accepted at ICLR 2026.
  • [Sep. 2025] Joined The Chinese University of Hong Kong as a Research Assistant, working on visual agents.
  • [2025] Received the BD Scholarship.
  • [2025] Our paper on long-tail representation in LVLMs has been accepted at CVPR 2025.
  • [2025] PRMBench accepted at ACL 2025 Main.

Experiences

  • [2026.5 – Present] Research Intern, Multimodal Model Department, Tencent Hunyuan (Project Up 青云计划 & Rhino Bird Elite Talent 犀牛鸟精英人才计划)
  • [2026.1 – 2026.5] Visiting Student, University of Illinois Urbana-Champaign (UIUC)
  • [2025.9 – 2026.1] Research Assistant, The Chinese University of Hong Kong
  • [Feb. 2023 – Aug. 2025] Research Intern, Shanghai Artificial Intelligence Laboratory

Education

  • [2023 – Present] Ph.D. Candidate in Artificial Intelligence, Fudan University
  • [2019 – 2023] B.Eng. in Computer Science and Technology, Harbin Institute of Technology (96.2/100, ranked 3/92)

Research Interests

  • Visual agents and multimodal tool use
  • Multi-step reasoning for LLMs and VLMs
  • Reliability of large vision-language models
  • Hallucination mitigation and evaluation

Selected Publications

AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning

ICLR 2026 · Mingyang Song*, Haoyu Sun*, Jiawei Gu*, Linjie Li*, Ranjay Krishna, Yu Cheng

A unified framework for data curation, reinforcement learning, and evaluation of tool-planning models. The model exhibits self-adaptive tool use and achieves substantial gains on visual spatial planning tasks.

Paper / Project / Code

PRMBench: A Fine-grained and Challenging Benchmark for Process-Level Reward Models

ACL 2025 Main · Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, Yu Cheng

A comprehensive benchmark for process-level reward models with 6,216 curated samples and 83,456 step-level labels, revealing key limitations in current PRMs.

Paper / Project / Code

From Head to Tail: Towards Balanced Representation in Large Vision-Language Models through Adaptive Data Calibration

CVPR 2025 · Mingyang Song, Xiaoye Qu, Jiawei Zhou, Yu Cheng

An in-depth analysis of the long-tail problem in current LVLM training data, with a fine-grained adaptive calibration framework for more balanced representation learning.

Paper / Project / Code

Mitigating Multilingual Hallucination in Large Vision-Language Models

TOMM · Xiaoye Qu*, Mingyang Song*, Wei Wei, Jianfeng Dong, Yu Cheng

An early approach for mitigating multilingual hallucinations in LVLMs, with an analysis of the main factors that drive non-English hallucination behavior.

Paper / Code

GenReasoner: Generative Visual Agent Planning

Nature Portfolio Submission

Studies whether tool-planning capabilities learned under specific task-tool combinations can generalize to unseen tasks, with a focus on transferable visual agent planning.

Awards

  • [2025] BD Scholarship
  • [2022] Finalist, The Interdisciplinary Contest in Modeling (COMAP)
  • [2021] First-Class People's Scholarship
  • [2020, 2021, 2022] National Encouragement Scholarship

Service

  • Reviewer: NeurIPS 2024, 2025, 2026; CVPR 2025; ACL 2025, 2026; TPAMI

Languages

  • Chinese: native
  • English: proficient
  • Japanese: working proficiency