Qixun Wang 王启迅

Portrait of Qixun Wang

I am a fourth-year Ph.D. student in the School of Intelligence Science and Technology at Peking University.

My current research focuses on advancing the core capabilities of multimodal large language models (MLLMs) and their downstream applications. I am particularly interested in post-training methods that strengthen perception and reasoning, including reasoning in latent space (Monet, SLVR) and reasoning with visual tools (Beacon). Building on these capabilities, I explore MLLM applications in audio-video captioning and understanding (RefCaptioner), enhancing dynamic 3D world generation through programming, assessing the authenticity of AI-generated videos (Artifact-Bench, Verus-Insight), and GUI agents (FocusMem).

Previously, I focused on out-of-distribution (OOD) generalization, conducting theoretical analysis and designing algorithms for visual recognition (MAT & LDAT), graph tasks (CIA-LRA), and in-context learning in large language models (ICL-OOD). This background continues to inform my efforts to develop more robust and generalizable methods for MLLMs.

News

July, 2026 One paper, Beacon, on improving reasoning-mode adaptiveness and achieving genuine tool-induced performance gains in agentic visual reasoning, was released on arXiv.
May, 2026 A paper entitled Artifact-Bench, which evaluates MLLMs for AI-generated video detection, was released on arXiv.
May, 2026 Our paper Semantic-Enriched Latent Visual Reasoning was accepted to ICML 2026.
March, 2026I joined the Kling Team at Kuaishou Technology as a research intern.
February, 2026 Two papers were accepted to CVPR 2026, covering reasoning in latent visual space and a benchmark for unified multimodal models.

Selected papers (see full publication)

  1. arXiv
    Beacon: Knowing When and Why to Perform Agentic Visual Reasoning
    Qixun Wang*, Yang Shi*, Letian Cheng, Zhuoran Zhang, Yan He, Yuqi Tang, Qi Zhang, Xinlei Yu, Ruizhe Chen, Tianrun Xu, Yuanxing Zhang, Pengfei Wan, Haotian Wang, Xianghua Ying
    arXiv preprint, 2026
    • Conduct a comprehensive analysis of reasoning-mode adaptiveness and tool-induced performance changes in existing agentic visual reasoning models.
    • Propose a novel training recipe that achieves state-of-the-art or competitive performance across 13 visual reasoning benchmarks, while improving reasoning-mode adaptiveness and delivering genuine tool-induced performance gains.
    PDF Code
  2. CVPR
    Monet: Reasoning in Latent Visual Space Beyond Images and Language
    Qixun Wang, Yang Shi, Yifei Wang, Yuanxing Zhang, Pengfei Wan, Kun Gai, Xianghua Ying, and Yisen Wang
    CVPR, 2026
    • Propose a new framework for multimodal latent reasoning, including dataset construction, SFT, and RL algorithms, achieving significant improvements on both in-domain and OOD visual reasoning benchmarks
    • 200+ GitHub stars
    PDF Code
  3. ICLR
    Can In-context Learning Really Generalize to Out-of-distribution Tasks?
    Qixun Wang, Yifei Wang, Xianghua Ying, and Yisen Wang
    ICLR, 2025
    • Reveal the capability boundary and algorithm selection mechanism of ICL on OOD tasks through carefully designed experiments and theoretical analysis.
    PDF Code
  4. NeurIPS
    Dissecting the Failure of Invariant Learning on Graphs
    Qixun Wang, Yifei Wang, Yisen Wang, and Xianghua Ying
    NeurIPS, 2024
    • Theoretically and empirically demonstrate the failure modes of classic invariant learning approaches on graph data, and propose a new training objective with significant performance gains and theoretical guarantees.
    PDF Code
  5. NeurIPS Spotlight
    Improving Out-of-distribution Generalization by Adversarial Training with Structured Priors
    Qixun Wang*, Yifei Wang*, Hong Zhu, and Yisen Wang
    NeurIPS, 2022
    • Propose a simple yet effective low-rank adversarial training strategy that improves the OOD generalization of visual recognition models.
    PDF Code
  6. ICML
    Semantic-Enriched Latent Visual Reasoning
    Tianrun Xu, Yue Sun, Qixun Wang, Jingyi Lu, Yuan Wang, Tianren Zhang, Longteng Guo, Fengyun Rao, Jing Lyu, Feng Chen, and Jing Liu
    ICML, 2026
    • Reveal and address the lack of semantic richness in latent embeddings learned by prior latent visual reasoning training paradigms.
    PDF Code
  7. arXiv
    Artifact-Bench: Evaluating MLLMs on Detecting and Assessing the Artifacts of AI-Generated Videos
    Yuqi Tang*, Yang Shi*, Zhuoran Zhang*, Qixun Wang*, and a group of outstanding researchers
    arXiv preprint, 2026
    • Propose a comprehensive benchmark for evaluating MLLMs’ ability to detect and analyze artifacts in AI-generated videos
    PDF Code

Experience

  • Kling Team, Kuaishou Technology — Research Intern http://arxiv.org/abs/2604.03315(March 2026–present)
    Working on multimodal agents.

Education

  • Ph.D. Candidate in Machine Learning and Computer Vision, School of Intelligence Science and Technology, Peking University (2023–present)
  • B.S. in Intelligence Science and Technology, EECS, Peking University (2019–2023)

Awards

  • The Third-Class Scholarship of Peking University (2025)
  • Merit Student at Peking University (2025)
  • Outstanding Graduate of Peking University (2023)
  • Yanchuang Capital Scholarship, Top 6% (2022)
  • Merit Student at Peking University, Top 6% (2022)
  • Academic Innovation Award at Peking University, Top 1% (2022)
  • Award for Academic Excellence (2021)