Dengming Zhang

Hello there, I'm Dengming Zhang from Zhejiang University, and I recently joined Tencent as an Large Language Models (LLMs) Post-Training Engineer.

My primary research focuses on Multimodal Large Models under low-data and low-compute constraints. On the low-data side, I study how to merge multiple domain-specialized, fine-tuned expert LLMs into one generalist model using only 1–5 samples, while retaining SOTA-level performance (ICLR 2026). On the low-compute side, I explore how to equip vision foundation models with audio using a single RTX 4090, and improve audio-visual affective understanding to a SOTA-level. I am also interested in Generative AI (Image/Music), Affective Computing, Meta-learning, and HCI.

By the way, I am good at combining scientific research with engineering implementation, and I have rich experience in front-end development, back-end development, and cluster devops. Some of the open source projects that I lead/participate in can be found on my GitHub.

Connect

News

2026/08

Our work on Generalizable Search Agents at Tencent is now available on arXiv.

2026/04

Joined Tencent as an Large Language Models (LLMs) Post-Training Engineer.

2026/03

Received my Master's degree from Zhejiang University.

2026/02

A paper on Efficient VLM Token Pruning got accepted to CVPR 2026.

2026/01

A paper on Model Merging got accepted to ICLR 2026.

2025/11

A preprint on Emotion Multimodal LLM is now on arXiv.

2025/09

A preprint on Model Merging is now on arXiv.

2025/07

Two papers on Controllable Music Generation got accepted to ACM MM 2025 and ACM MM 2025.

2025/06

A paper on Controlled Text Rendering Image Generation got accepted to ICCV 2025.

2024/12

A paper on Personalized Emotion Recognition got accepted to AAAI 2025.

2024/08

A paper on Controlled Style Image Generation got accepted to UIST 2024.

Research Highlights

Research Area

LLM & Agentic AI

1 paper

#01Technical Report2026

Cross-Domain Hybrid OPD for Generalizable Search Agents

Hongzhan Chen, Xiaoyu Liu, Dengming Zhang, Minzhou Huang, Dongliang Xu, Jingcheng Xie, Dongxiang Fang, Bowen Qin, Minsheng Hao, Yaozong Shen, Xiaojun Quan, Mona Zhou, Haosheng Zou, Jeff Chen

Motivation: Search-oriented RL improves retrieval but can erode the general reasoning and instruction-following skills expected from universal assistants.
Method: Hybrid-OPD couples agentic search RL with cross-domain expert distillation, preserving search specialization while recovering broad capabilities.

Research Area

Multimodal Large Models

3 papers

Expert Merging: Model Merging with Unsupervised Expert Alignment and Importance-Guided Layer Chunking thumbnail

ICLR 26

#02International Conference On Learning Representations (ICLR)2026

Expert Merging: Model Merging with Unsupervised Expert Alignment and Importance-Guided Layer Chunking

Dengming Zhang, Xiaowen Ma, Zhenliang Ni, Zhenkai Wu, Han Shu, Xin Jiang, Xinghao Chen

Motivation: Existing model-merging methods rely on hand-tuned weights and treat every layer alike despite large differences in layer importance.
Method: Expert Merging learns layer-wise coefficients from unlabeled calibration data and assigns finer merging granularity to high-impact layers.

#03Preprint2025

Learning to Hear by Seeing: It's Time for Vision Language Models to Understand Artistic Emotion from Sight and Sound

Dengming Zhang, Weitao You, Jingxiong Li, Weishen Lin, Wenda Shi, Xue Zhao, Heda Zuo, Junxian Wu, Lingyun Sun

Motivation: Artistic emotion emerges from sight and sound together, but most audio-visual models depend on expensive, large-scale audio pretraining.
Method: VAEmotionLLM aligns a new audio pathway to visual signals, then uses a lightweight adapter to learn cross-modal emotion.

VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm thumbnail

CVPR 26

#04IEEE/CVF Conference On Computer Vision And Pattern Recognition (CVPR)2026

VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm

Zhenkai Wu, Xiaowen Ma, Zhenliang Ni, Dengming Zhang, Han Shu, Xin Jiang, Xinghao Chen

Motivation: VLMs process many redundant visual tokens, while importance-only pruning can retain duplicates and miss critical image regions.
Method: VLM-Pruner balances redundancy and spatial coverage through centrifugal selection, spatial-sparsity buffering, and selective token fusion.

Research Area

Affective Computing & Music Emotion

1 paper

Personalized Dynamic Music Emotion Recognition with Dual-Scale Attention-Based Meta-Learning thumbnail

AAAI 25

#05Proceedings Of The AAAI Conference On Artificial Intelligence (AAAI)2025

Personalized Dynamic Music Emotion Recognition with Dual-Scale Attention-Based Meta-Learning

Dengming Zhang, Weitao You, Ziheng Liu, Lingyun Sun, Pei Chen

Motivation: Dynamic music emotion changes over time and across listeners, yet existing models capture neither long-range context nor personal perception well.
Method: DSAML combines dual-scale attention with annotator-specific meta-learning to personalize emotion prediction from a single user annotation.

Research Area

Meta-learning & Diagnosis

1 paper

Few-shot bearing fault diagnosis based on meta-learning with discriminant space optimization thumbnail

MST 22

#06Measurement Science And Technology2022

Few-shot bearing fault diagnosis based on meta-learning with discriminant space optimization

Dengming Zhang, Kai Zheng, Yin Bai, Dengke Yao, Dewei Yang, Shaowang Wang

Motivation: Real-world bearing faults provide few labeled examples and appear under heavy noise and shifting operating conditions.
Method: MLDSO optimizes a discriminative, fault-tolerant feature space for meta-learning, enabling robust diagnosis from only a few samples.

Research Area

Controllable Generation & Creative AI

4 papers

StyleFactory: Towards Better Style Alignment in Image Creation through Style-Strength-Based Control and Evaluation thumbnail

UIST 24

#07Proceedings Of The 37th Annual ACM Symposium On User Interface Software And Technology (UIST)2024

StyleFactory: Towards Better Style Alignment in Image Creation through Style-Strength-Based Control and Evaluation

Mingxu Zhou, Dengming Zhang, Weitao You, Ziqi Yu, Yifei Wu, Chenghao Pan, Huiting Liu, Tianyu Lao, Pei Chen

Motivation: People and image generators often disagree on what a style means and how strongly it should appear.
Method: StyleFactory turns style strength into an interactive ranking and visualization workflow for iteratively aligning personal style.

FonTS: Text Rendering with Typography and Style Controls thumbnail

ICCV 25

#08International Conference On Computer Vision (ICCV)2025

FonTS: Text Rendering with Typography and Style Controls

Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, Xingxing Zou

Motivation: Text-to-image models struggle to keep fonts and styles consistent or apply typography to specific words.
Method: FonTS pairs parameter-efficient typography tuning with control tokens and a text-agnostic adapter for precise type and style control.

Controllable Video-to-Music Generation with Multiple Time-Varying Conditions thumbnail

ACM MM 25(Oral)

#0933rd ACM International Conference On Multimedia (ACM MM)2025

Controllable Video-to-Music Generation with Multiple Time-Varying Conditions

Junxian Wu, Weitao You, Heda Zuo, Dengming Zhang, Pei Chen, Lingyun Sun

Motivation: Video-to-music systems often generate in a black box, giving users little control over musical attributes that evolve over time.
Method: The framework combines temporal alignment, dynamic condition fusion, and guided decoding to follow multiple time-varying controls.

Spatial-Temporal Decomposition and Alignment in Controllable Video-to-Music Generation thumbnail

ACM MM 25

#1033rd ACM International Conference On Multimedia (ACM MM)2025

Spatial-Temporal Decomposition and Alignment in Controllable Video-to-Music Generation

Weitao You, Heda Zuo, Junxian Wu, Dengming Zhang, Zhibin Zhou, Lingyun Sun

Motivation: Most video-to-music methods prioritize output quality but make the generated music difficult to adapt to user intent.
Method: The model decomposes and aligns spatial-temporal video-music features through flow matching, enabling flexible control without explicit labels.

Experience

Tencent

LLMs Post-Training Engineer2026.04 - Present

Shanghai, China

Work on Post-training for Large Language Models (LLMs).

Huawei Noah's Ark Lab

Research Intern2025.06 - 2025.12

Shanghai, China

Research internship on Model Merging (Expert Merging)^[1].

Tencent

Algorithm Engineer Intern2025.04 - 2025.06

Shenzhen, China

Work on Game Character Material Generation with animation.

Tencent

2026.04 - Present

LLMs Post-Training EngineerShanghai, China

Work on Post-training for Large Language Models (LLMs).

Huawei Noah's Ark Lab

2025.06 - 2025.12

Research InternShanghai, China

Research internship on Model Merging (Expert Merging)^[1].

Tencent

2025.04 - 2025.06

Algorithm Engineer InternShenzhen, China

Work on Game Character Material Generation with animation.

Awards

First Zhaoyuan Chengen Technology Innovation Scholarship (Top 1)
2021.12, University-wide Unique Award
First-Class Academic Scholarship (Top 5%)
2022.12, 2021-2022 Academic Year
First-Class Academic Scholarship (Top 5%)
2021.12, 2020-2021 Academic Year
First-Class Academic Scholarship (Top 5%)
2020.12, 2019-2020 Academic Year
First Prize in Chongqing, National Electronic Design Contest
2022.01, Chongqing Municipal Education Commission
First Prize, TI Cup Electronic Design Contest
2020.11, Chongqing Municipal Education Commission
Chongqing Excellent Undergraduate Graduation Thesis
2023.06, Chongqing Municipal Education Commission

First Zhaoyuan Chengen Technology Innovation Scholarship (Top 1)
2021.12, University-wide Unique Award
First-Class Academic Scholarship (Top 5%)
2022.12, 2021-2022 Academic Year
First-Class Academic Scholarship (Top 5%)
2021.12, 2020-2021 Academic Year
First-Class Academic Scholarship (Top 5%)
2020.12, 2019-2020 Academic Year
First Prize in Chongqing, National Electronic Design Contest
2022.01, Chongqing Municipal Education Commission
First Prize, TI Cup Electronic Design Contest
2020.11, Chongqing Municipal Education Commission
Chongqing Excellent Undergraduate Graduation Thesis
2023.06, Chongqing Municipal Education Commission

Dengming Zhang

News

Research Highlights

LLM & Agentic AI

Multimodal Large Models

Affective Computing & Music Emotion

Meta-learning & Diagnosis

Controllable Generation & Creative AI

Experience

Tencent

Huawei Noah's Ark Lab

Tencent

Tencent

Huawei Noah's Ark Lab

Tencent

Awards

First Zhaoyuan Chengen Technology Innovation Scholarship (Top 1)

First-Class Academic Scholarship (Top 5%)

First-Class Academic Scholarship (Top 5%)

First-Class Academic Scholarship (Top 5%)

First Prize in Chongqing, National Electronic Design Contest

First Prize, TI Cup Electronic Design Contest

Chongqing Excellent Undergraduate Graduation Thesis

First Zhaoyuan Chengen Technology Innovation Scholarship (Top 1)

First-Class Academic Scholarship (Top 5%)

First-Class Academic Scholarship (Top 5%)

First-Class Academic Scholarship (Top 5%)

First Prize in Chongqing, National Electronic Design Contest

First Prize, TI Cup Electronic Design Contest

Chongqing Excellent Undergraduate Graduation Thesis