Li Xiao (李潇)

ByteDance Seed Team

Member of TopSeed Talent Program

xiaoli.cst@gmail.com
Google Scholar
Dinghao Building, Tower B, Beijing, China

NOTE: Our team (Seed LLM) is hiring interns and full-time researchers. If you are interested in LLM pretraining/reasoning/agent, feel free to contact me!

I am a researcher at ByteDance Seed, as a member of TopSeed Program, working with Ke Shen since 2024. I received my Ph.D. in 2025 from the Department of Computer Science and Technology at Tsinghua University (THU). I was a member of TSAIL Group, led by Prof. Bo Zhang and Prof. Jun Zhu, where I worked closely with Prof. Xiaolin Hu and Prof. Bo Zhang. I obtained my Bachelor’s degree from Tsinghua University.

My long-term goal is to build scalable, robust, and generalizable autonomous agents that operate reliably in both digital and real world and fundamentally relieve humans from tedious labor. In the near term, I focus on building world-class foundation models that create real economic value and measurably improve productivity (not models optimized merely for leaderboard performance). I believe accelerating progress toward this goal requires rethinking the current paradigm along these directions:

End-to-end system optimization: Re-examining pretraining, post-training, and evaluation from a unified perspective, jointly optimizing for scalability and generalization rather than treating stages independently.
Predictable scaling ladders: Developing principled, systematic scaling strategies to accelerate model iteration while improving reliability and reducing empirical trial-and-error.
Diving deeper: Exploring better training paradigms (objectives, optimizers, data strategies, etc) that fully leverage pretraining data, RL signals, and human supervision while enabling continuous knowledge and skill acquisition.

Current Work & Progress

Core contributor to the development of flagship foundation models (Seed 1.6, Seed 1.8, and the upcoming Seed 2.0) as well as open-source models (Seed-OSS).
Research on the emergence and enhancement of reasoning patterns in pretrained models and end-to-end strategies to extend their capability ceilings.

Current Interns

Huanran Chen (Tsinghua University, LLM pretraining dynamics)

Previous Experiences

2024.6 - 2024.9, 01.AI, VLM Pretraining Group
2023.5 - 2024.4, Huawei 2012 Lab, Foundation Model Research Group

news

Apr 13, 2026	We introduce Nexus, a novel optimizer that achieves better downstream generalization at the same pretraining loss by aligning gradients and encouraging convergence to shared minima.
Feb 14, 2026	We released Seed 2.0. Building on our efforts to improve generalizable reasoning in Seed 1.8, I was responsible for extending these capabilities across domains and modalities.
Jan 26, 2026	We propose the COD framework that accurately predicts LLM downstream performance before training, achieving a 1.55% average prediction error on a 70B parameter model. This work has been accepted in ICLR 2026. Read more
Dec 18, 2025	We released Seed 1.8. I was responsible for improving the model’s generalizable reasoning capabilities.
Aug 21, 2025	We released Seed-OSS model. I was responsible for enhancing the model’s reasoning density, aenabling it to maintain longer chains of thought and tackle more challenging problems.

selected publications

2026

arXiv

Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima

Huanran Chen, Huaqing Zhang, Xiao Li, Yinpeng Dong, Ke Shen, and Jun Zhu

arXiv preprint arXiv:2604.09258, 2026

Bib

@article{nexus,
  title = {Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima},
  author = {Chen, Huanran and Zhang, Huaqing and Li, Xiao and Dong, Yinpeng and Shen, Ke and Zhu, Jun},
  journal = {arXiv preprint arXiv:2604.09258},
  year = {2026},
}

ICLR

Unveiling downstream performance scaling of llms: A clustering-based perspective

Chengyin Xu, Kaiyuan Chen, Xiao Li, Ke Shen, and Chenggang Li

International Conference on Learning Representations (ICLR), 2026

Bib arXiv

@article{xu2025unveiling,
  title = {Unveiling downstream performance scaling of llms: A clustering-based perspective},
  author = {Xu, Chengyin and Chen, Kaiyuan and Li, Xiao and Shen, Ke and Li, Chenggang},
  journal = {International Conference on Learning Representations (ICLR)},
  year = {2026},
}

2025

Report

Seed1.8 Model Card: Towards Generalized Real-World Agency

Bytedance Seed

Model Card, 2025

Bib

Report

Technical Introduction to the Seed1.6 Model Series

Bytedance Seed

Technical Report, 2025

Bib

@article{seed2025technical,
  title = {Technical Introduction to the Seed1.6 Model Series},
  author = {Seed, Bytedance},
  journal = {Technical Report},
  year = {2025},
  link = {https://seed.bytedance.com/en/seed1_6},
}

ICLR

ADBM: Adversarial diffusion bridge model for reliable adversarial purification

Xiao Li, Wenxuan Sun, Huanran Chen, Qiongxiu Li, Yining Liu, Yingzhe He, Jie Shi, and Xiaolin Hu

International Conference on Learning Representations (ICLR), 2025

Bib arXiv Code

@article{li2024adbm,
  link = {https://openreview.net/forum?id=g0rnZeBguq},
  title = {ADBM: Adversarial diffusion bridge model for reliable adversarial purification},
  author = {Li, Xiao and Sun, Wenxuan and Chen, Huanran and Li, Qiongxiu and Liu, Yining and He, Yingzhe and Shi, Jie and Hu, Xiaolin},
  journal = {International Conference on Learning Representations (ICLR)},
  year = {2025}
}

2024

arXiv

Faster-GCG: Efficient Discrete Optimization Jailbreak Attacks against Aligned Large Language Models

Xiao Li, Zhuhong Li, Qiongxiu Li, Bingze Lee, Jinghao Cui, and Xiaolin Hu

arXiv preprint arXiv:2410.15362, 2024

Bib arXiv

@article{fastergcg,
  title = {Faster-GCG: Efficient Discrete Optimization Jailbreak Attacks against Aligned Large Language Models},
  author = {Li, Xiao and Li, Zhuhong and Li, Qiongxiu and Lee, Bingze and Cui, Jinghao and Hu, Xiaolin},
  journal = {arXiv preprint arXiv:2410.15362},
  year = {2024}
}

2023

TPAMI

Recognizing Object by Components With Human Prior Knowledge Enhances Adversarial Robustness of Deep Neural Networks

Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, and Xiaolin Hu

IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023

Bib arXiv Code

@article{rock,
  author = {Li, Xiao and Wang, Ziqi and Zhang, Bo and Sun, Fuchun and Hu, Xiaolin},
  title = {Recognizing Object by Components With Human Prior Knowledge Enhances
                    Adversarial Robustness of Deep Neural Networks},
  journal = {{IEEE} Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  volume = {45},
  number = {7},
  pages = {8861--8873},
  year = {2023}
}