About

I am Junjie Wang, a Ph.D. student in Computer Science at Harbin Institute of Technology, Shenzhen. I am co-advised by Prof. Zhuotao Tian and Prof. Bin Chen. My research focuses on multimodal understanding and generation, multimodal visual perception, and efficient reasoning for large models. I have worked as a research intern at ByteDance (TikTok) and Kling AI at Kuaishou, where I studied image and video generation agents, multimodal LLM distillation, and vision-encoder post-training.

I expect to graduate in 2027 and am actively seeking industrial research positions. If you are interested, please feel free to reach out via email or WeChat (xiaomoguhzz).

Recent News

Experience

Kuaishou, Kling AI 2026.04 – Present

Research InternSupervised by Yuanzheng Ci and Xin Tao

• MLLM Vision-Encoder Post-training: built a three-stage post-training pipeline on MS-Swift with a multi-teacher distillation scheme, aiming to obtain a vision encoder that surpasses the Qwen3.5 ViT in both spatial structure and temporal dynamics.
• Video-Generation Agent: post-trained a Qwen3.5-based MLLM agent that drives video-generation models through planning, reflection, and feedback rectification, and designed the corresponding benchmark and reinforcement learning training method.

ByteDance, TikTok 2025.09 – 2026.04

Research InternSupervised by Xiangtai Li and Guangcan Mai

• Image-Generation Agent: formalized T2I generation as a Reason-Reflect-Rectify loop and built the R³-Bench benchmark together with the R³-Refiner reinforcement learning method, which has been accepted to ICML 2026.
• TinyMLLM Distillation: built a two-stage distillation pipeline on Qwen3-VL and Qwen3.5, covering recipe design, supervision-data construction, teacher training, and 2B-model distillation. The resulting model has been deployed in live-streaming services to replace Gemini and reduce cost.

Publications

ICML 2026

Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation

Wang J*, Lou X*, Tian Z, et al.

Paper Code

ECCV 2026

Scaling Multi-Reference Image Generation with Dynamic Reward Optimization

Huang W, Wang J, Tian Z, et al.

Paper Code

IEEE TPAMI

Generalized Decoupled Learning for Enhancing Open-Vocabulary Dense Perception

Wang J, Chen K, Li Y, Chen B, Zhao H, Qi X, Tian Z.

Accepted Paper Code

CVPR 2025

DeCLIP: Decoupled Learning for Open-Vocabulary Dense Perception

Wang J, Chen B, Tian Z, et al.

Paper Code

AAAI 2025

OV-DQUO: Open-Vocabulary DETR with Denoising Text Query Training and Open-World Unknown Objects Supervision

Wang J, Chen B, Kang B, et al.

Paper Code

ACM MM 2025

CalibCLIP: Contextual Calibration of Dominant Semantics for Text-Driven Image Retrieval

Kang B, Chen B, Wang J, Tian Z, et al.

Oral Paper

NeurIPS 2025

Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior

Li Y, Wang J, Tian Z, et al.

Paper Code

ICLR 2026

Efficient Reasoning with Balanced Thinking

Li Y, Wang J, Tian Z, et al.

Paper Code

Honors

• National Scholarship for Graduate Students
• Provincial Outstanding Master's Thesis
• Outstanding Graduate