Joel Julin

I'm a Ph.D. student in Robotics at Carnegie Mellon University, advised by László Jeni. I study how learning systems represent and reason about dynamic visual worlds. My work spans reconstructing, tracking, and controlling 3D scenes, as well as grounding multimodal decisions in evidence across space and time.

I hold an M.S. in Robotics from CMU and a B.S. in Computer Science from the University of Pittsburgh. My research background began while at Pitt, where I worked with Simon Lucey in the Ci2CV Computer Vision Lab at CMU's Robotics Institute on spatial mixing in visual representation learning. I later joined László Jeni's Computational Behavior Lab through the Robotics Institute Summer Scholars Program, supported by an NSF Research Experience Grant.

News

Recent publications, milestones, and research updates.

Oct. 2026 Reconstruct What You Can: Distractor-free 4D Gaussian Splatting accepted at WACV '27!
Sept. 2026 Benchmark paper on Audio-Visually Grounded Chain-of-Thought Reasoning accepted at NeurIPS '26!
Sept. 2026 Co-organized BEAM 2, the 2nd Workshop on Benchmarking Evidence-Aligned Multimodal Reasoning, at ECCV '26.
Aug. 2025 Started Ph.D. in Robotics at CMU.
Jul. 2025 Completed Master's in Robotics at CMU. View the thesis.
Apr. 2025 Paper on Rethinking the Role of Spatial Mixing accepted to eLVM workshop at CVPR '25.
Feb. 2025 Accepted Ph.D. offer from Carnegie Mellon in Robotics, starting Fall 2025.
Jan. 2025 Work on analytical flow supervision for 3DGS (GSLK) accepted to ICLR '25.
Feb. 2024 Work on Controllable Gaussian Splatting (CoGS) accepted to CVPR '24.
Mar. 2023 Accepted M.S. offer from Carnegie Mellon in Robotics.
Feb. 2023 Work on dynamic light field networks (DyLiN) accepted to CVPR '23.
Dec. 2022 Accepted full-time position at CMU to continue with Dr. Jeni's lab in the spring.
Dec. 2022 Graduated from the University of Pittsburgh.
Nov. 2022 Won gold medal at the 8th China International College Students' 'Internet+' Innovation and Entrepreneurship Competition.
Oct. 2022 Attended HackMIT at MIT, building Immersive NeRF.
May 2022 Accepted to CMU's Robotics Institute Summer Scholar program (project).
May 2021 Began working with Dr. Lucey's computer vision lab (see research here).

Research

Learning representations that make dynamic scenes more realistic, controllable, and geometrically grounded.

Reconstruct What You Can: Distractor-free 4D Gaussian Splatting

Joel Julin, Liuyue Xie, Ananya Bal, Mose Sakashita, Koichiro Niinuma, and Laszlo A. Jeni

Reconstructs dynamic scenes while separating transient occluders, reflections, and fog using uncertainty from the learned deformation field.

WACV '27

Paper forthcoming Project

ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark

Joel Julin*, Souraja Kundu*, Liza Dahiya, George Wei, Liuyue Xie, Jingru Yi, Xu Zhang, Hao Yang, László Jeni

Benchmarks whether multimodal models ground each reasoning step in the correct visual and auditory evidence across time.

NeurIPS '26 · E&D

* Joint first authors

Paper forthcoming Project

Rethinking the Role of Spatial Mixing

George Cazenavette, Joel Julin, and Simon Lucey

Shows that learning channel mixing while freezing spatial mixing can preserve classification performance and improve adversarial robustness.

CVPR-W '25

Paper

Gaussian Splatting Lucas-Kanade

Liuyue Xie, Joel Julin, Koichiro Niinuma, and László Jeni

Adapts Lucas-Kanade to dynamic Gaussian splatting with analytical scene-flow constraints for reconstructing highly dynamic scenes from limited camera motion.

ICLR '25

Paper Project