Zijian Feng

About Me

I am a Staff Applied Scientist at TikTok Business Integrity. We build LLMs and MLLMs through both pre-training and post-training, that power the moderation of TikTok’s monetization ecosystem, keeping the platform safe for users and the business sustainable.

Before this role, I was a Staff Research Engineer at Gaussian Robotics, where I built the machine learning stack for low-compute robots: multi-task learning on edge devices, INT8 quantization and hardware-in-the-loop testing. Earlier I was a Senior Research Scientist at TikTok AI Lab, working on novel view synthesis and visual search, and before that I worked on large-scale visual search at ViSenze and led pedestrian recognition and tracking at Tuputech.

I received a bachelor’s degree in Computing Science from the University of Glasgow in 2016.

These days I care most about how large models are trained: making pre-training and post-training fast, correct and explainable at scale. In my spare time I am building mew (by hands), an LLM training stack written from scratch, and blogging what I learn along the way.

Featured Projects

mew

An LLM training stack written from scratch: BPE tokenizer, Transformer, Triton FlashAttention and data-parallel training today, growing toward 4D parallelism (DP × TP × PP × CP), MoE with expert parallelism, SFT and GRPO. Every component ships with a cost model, a parity test and a benchmark against TorchTitan.

code · roadmap · blog series

news

Oct 08, 2026	Launched the project page for mew, an LLM training stack I’m building from scratch. A blog series on each phase is on the way.
Sep 23, 2026	Released the Pistis Technical Report: 27B and 9B multimodal LLMs post-trained with interleaved distillation and RL.
May 20, 2026	Released TextSculptor, a dataset and benchmark for training and evaluating scene text editing.
Jul 01, 2022	A Deep-Learning-based System for Indoor Active Cleaning was accepted to IROS 2022.
Jul 23, 2021	MINE: Towards Continuous Depth MPI with NeRF for Novel View Synthesis was accepted to ICCV 2021.

selected publications

Pistis Technical Report

Heyun Chen, Xiaohan Lan, Jiaxi Li, Zhilin Lu, Qi She, Weiwen Xu, Fei Yu, Yujie Zhong, Jinghuan Chen, Zijian Feng, Siyu Jiao, Yiheng Lin, Xinhao Wang, Sihan Yang, Jieyu You, Changbin Zhang, Hengyu Zhang, Xudong Zhang, Yunqing Zhao, and Shuai Zheng

arXiv preprint arXiv:2609.28554, Sep 2026

Abs arXiv Bib

We introduce the Pistis model family, comprising 27B- and 9B-parameter multimodal large language models built on Qwen3.6 and Qwen3.5, respectively, and developed through a general and scalable post-training framework. The framework first establishes a strong foundation through large-scale multimodal supervised fine-tuning (SFT). Building on this SFT foundation, we propose Interleaved Distillation and Reinforcement Learning (IDRL), a novel post-training paradigm that tightly integrates on-policy distillation and reinforcement learning within a single training loop.
@article{chen2026pistis, title = {Pistis Technical Report}, author = {Chen, Heyun and Lan, Xiaohan and Li, Jiaxi and Lu, Zhilin and She, Qi and Xu, Weiwen and Yu, Fei and Zhong, Yujie and Chen, Jinghuan and Feng, Zijian and Jiao, Siyu and Lin, Yiheng and Wang, Xinhao and Yang, Sihan and You, Jieyu and Zhang, Changbin and Zhang, Hengyu and Zhang, Xudong and Zhao, Yunqing and Zheng, Shuai}, journal = {arXiv preprint arXiv:2609.28554}, year = {2026}, month = sep, }
TextSculptor: Training and Benchmarking Scene Text Editing

Yiheng Lin, Siyu Jiao, Xiaohan Lan, Wei Zhou, Qi She, Fei Yu, Heyun Chen, Zhengwei Wang, Jinghuan Chen, Moran Li, Yingchen Yu, Zijian Feng, Yao Zhao, Yunchao Wei, and Yujie Zhong

arXiv preprint arXiv:2605.21090, May 2026

Abs arXiv Bib Code

Recent advances in Multimodal Large Language Models (MLLMs) and diffusion-based generative models have substantially improved prompt-driven image editing. However, scene text editing remains challenging, as it requires models to precisely modify textual content while preserving visual realism and non-target regions. We present TextSculptor, a comprehensive framework for data construction and evaluation of scene text editing: an automated data construction pipeline, TextSculpt-Data (3.2M training samples), and TextSculpt-Bench, a benchmark covering text addition, replacement, removal and hybrid editing.
@article{lin2026textsculptor, title = {TextSculptor: Training and Benchmarking Scene Text Editing}, author = {Lin, Yiheng and Jiao, Siyu and Lan, Xiaohan and Zhou, Wei and She, Qi and Yu, Fei and Chen, Heyun and Wang, Zhengwei and Chen, Jinghuan and Li, Moran and Yu, Yingchen and Feng, Zijian and Zhao, Yao and Wei, Yunchao and Zhong, Yujie}, journal = {arXiv preprint arXiv:2605.21090}, year = {2026}, month = may, }

A Deep-Learning-based System for Indoor Active Cleaning

Yike Yun^*, Linjie Hou^*, Zijian Feng^*, Wei Jin, Yang Liu, Heng Wang, Ruonan He, Weitao Guo, Bo Han, Baoxing Qin, and Jiaxin Li

In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Oct 2022

Bib PDF Code Website

@inproceedings{yun2022active,
  title = {A Deep-Learning-based System for Indoor Active Cleaning},
  author = {Yun, Yike and Hou, Linjie and Feng, Zijian and Jin, Wei and Liu, Yang and Wang, Heng and He, Ruonan and Guo, Weitao and Han, Bo and Qin, Baoxing and Li, Jiaxin},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2022},
  month = oct,
}

MINE: Towards Continuous Depth MPI with NeRF for Novel View Synthesis

Jiaxin Li^*, Zijian Feng^*, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee

In IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2021

arXiv Bib Code Website

@inproceedings{li2021mine,
  title = {MINE: Towards Continuous Depth MPI with NeRF for Novel View Synthesis},
  author = {Li, Jiaxin and Feng, Zijian and She, Qi and Ding, Henghui and Wang, Changhu and Lee, Gim Hee},
  booktitle = {IEEE/CVF International Conference on Computer Vision (ICCV)},
  year = {2021},
  month = oct,
}