Jianxiong Shen

Jianxiong Shen

Computer Vision · Multi-Modal Generative AI · Reinforcement Learning

Jianxiong Shen

I am currently a Research Scientist at Tencent in Shenzhen. My work spans Game Agents, Multi-Modal Generative AI and Reinforcement Learning.

I received my Ph.D. from the Polytechnic University of Catalonia (UPC) in 2024, advised by Francesc Moreno-Noguer and Adria Ruiz. Before that, I received my B.Eng. and M.Eng. degrees from Harbin Institute of Technology.

Selected work Google Scholar GitHub
Projects

Recent research

Experiment-driven studies of RL post-training across LLMs, VLMs and Diffusion Models.

Multimodal post-training experiment curves
Recent ResearchVision-Language Models

On-Policy Distillation + GRPO for Geometric Reasoning

Project paper and reproducible training artifacts

Compared sparse sequence rewards with dense teacher feedback on Qwen2.5-VL-7B, then combined both stages to improve Geometry3K accuracy from 37.8% to 54.2%.

Paper Project
Diffusion RL OCR reward examples
Recent ResearchDiffusion Models

Reward Hacking or Forgetting?

Short note on reward-conditioned scene collapse

Studied scene collapse under verifiable OCR reward fine-tuning, separating the observed failure from simple catastrophic-forgetting explanations.

Paper Project
Text reasoning RL training curves
Recent ResearchLanguage Models

R1-Zero Style Reasoning and Transfer

Project report on emergent reasoning and transfer

Reproduced emergent GRPO reasoning at 3B scale and measured where the learned search behavior transfers, including both positive and negative results.

Project
Publications

Selected first-author work

Four projects tracing a path from reliable 3D reconstruction to efficient neural rendering.

LOD-GS levels-of-detail teaser
CVPR 20253D Gaussian Splatting

LOD-GS: Achieving Levels of Detail using Scalable Gaussian Soup

Jianxiong Shen, Yue Qian, Xiaohang Zhan

Structures Gaussians with scalable triangle primitives to maintain high rendering quality across progressively smaller memory budgets.

Paper DOI
Estimating 3D Uncertainty Field teaser
ICRA 2024Uncertainty

Estimating 3D Uncertainty Field: Quantifying Uncertainty for Neural Radiance Fields

Jianxiong Shen, Ruijie Ren, Adria Ruiz, Francesc Moreno-Noguer

Models spatial uncertainty beyond individual rendered views, producing a queryable 3D uncertainty field for neural scenes.

Paper Code
Conditional-Flow NeRF teaser
ECCV 2022NeRF

Conditional-Flow NeRF: Accurate 3D Modelling with Reliable Uncertainty Estimation

Jianxiong Shen, Antonio Agudo, Francesc Moreno-Noguer, Adria Ruiz

Uses conditional normalizing flows to improve both reconstruction accuracy and calibrated uncertainty estimation in NeRF.

Project Paper Code Slides
Stochastic Neural Radiance Fields teaser
3DV 2021Neural Rendering

Stochastic Neural Radiance Fields: Quantifying Uncertainty in Implicit 3D Representations

Jianxiong Shen, Adria Ruiz, Antonio Agudo, Francesc Moreno-Noguer

Introduces stochastic radiance fields for estimating predictive uncertainty in implicit 3D scene representations.

Paper Poster
Background

Experience & education

2024 — Present

Research Scientist · Tencent

Game reinforcement learning, multimodal agents, and post-training research.

2019 — 2024

Ph.D. · Polytechnic University of Catalonia

Computer vision and 3D scene modelling at the Institut de Robòtica i Informàtica Industrial. Thesis awarded Excellent Cum Laude.

2013 — 2019

B.Eng. & M.Eng. · Harbin Institute of Technology

Engineering education and early research in computer vision.

Updates

Recent news

Released a short empirical note on scene collapse under OCR-reward diffusion RL.

LOD-GS was published at CVPR 2025.

Joined Tencent as a Research Scientist.

Completed my Ph.D. with Excellent Cum Laude.

Presented our work on 3D uncertainty fields at ICRA 2024.