Tianjun Zhong

I am Tianjun Zhong, a Master’s student in Computer Science at Columbia University. My research focuses on understanding and enhancing reasoning in large language models (LLMs), with broader interests in interpretability and LLM-based agents.

My work spans several directions: investigating the mechanisms that underlie reasoning in LLMs, developing methods to improve interpretability and layer-wise understanding, building collaborative agentic systems for complex tasks, and advancing approaches to post-training that preserve reliability at scale. These efforts aim to deepen our scientific understanding of model internals while creating practical systems that are efficient and trustworthy.

I have had the opportunity to collaborate with Professor Nima Mesgarani (Columbia) on bridging reasoning in LLMs with human neural activity, Professor Zhiting Hu (UC San Diego) on large-scale post-training and reinforcement learning for reasoning, and Professor Baishakhi Ray (Columbia) on designing multi-agent systems for software engineering. Together, these experiences shape my broader interest in aligning LLM development with both interpretability and real-world deployment.

News

Sep 18, 2025	🎉 Our paper on reasoning and interpretability in large language models has been accepted to NeurIPS 2025!

Selected Publications

Preprint

SWE-Spot: Building Small Repo-Experts with Repository-Centric Learning

Jinjun Peng^*, Magnus Saebo^*, Tianjun Zhong, and 5 more authors

2026

arXiv Bib

@misc{peng2026swespotbuildingsmallrepoexperts,
  title = {SWE-Spot: Building Small Repo-Experts with Repository-Centric Learning},
  author = {Peng, Jinjun and Saebo, Magnus and Zhong, Tianjun and Cheng, Yi-Jie and Yang, Junfeng and Ray, Baishakhi and Chen, Simin and Ding, Yangruibo},
  year = {2026},
  url = {https://arxiv.org/abs/2601.21649},
}

Preprint

From Chains to DAGs: Probing the Graph Structure of Reasoning in LLMs

Tianjun Zhong, Linyang He, and Nima Mesgarani

2026

arXiv Bib

@misc{zhong2026chainsdagsprobinggraph,
  title = {From Chains to DAGs: Probing the Graph Structure of Reasoning in LLMs},
  author = {Zhong, Tianjun and He, Linyang and Mesgarani, Nima},
  year = {2026},
  url = {https://arxiv.org/abs/2601.17593},
}

NeurIPS 2025

Brain-Predictive Reasoning Embedding through Residual Disentanglement

Linyang He^*, Tianjun Zhong^*, Richard Antonello, and 3 more authors

In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

arXiv OpenReview Code Bib

@inproceedings{he2025brainpredictive,
  title = {Brain-Predictive Reasoning Embedding through Residual Disentanglement},
  author = {He, Linyang and Zhong, Tianjun and Antonello, Richard and Mischler, Gavin and Goldblum, Micah and Mesgarani, Nima},
  booktitle = {The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year = {2025},
  openreview = {https://openreview.net/forum?id=tPqBnGwTwa},
  url = {https://arxiv.org/abs/2510.22860},
}

Preprint

K2-Think: A Parameter-Efficient Reasoning System

Zhoujun Cheng^*, Richard Fan^*, Shibo Hao^*, and 28 more authors

2025

arXiv Hugging Face Code Bib

@misc{cheng2025k2thinkparameterefficientreasoning,
  title = {K2-Think: A Parameter-Efficient Reasoning System},
  author = {Cheng, Zhoujun and Fan, Richard and Hao, Shibo and Killian, Taylor W. and Li, Haonan and Sun, Suqi and Ren, Hector and Moreno, Alexander and Zhang, Daqian and Zhong, Tianjun and Xiong, Yuxin and Hu, Yuanzhe and Xie, Yutao and Han, Xudong and Wang, Yuqi and Pimpalkhute, Varad and Zhuang, Yonghao and Singh, Aaryamonvikram and Liang, Xuezhi and Xie, Anze and She, Jianshu and Fan, Desai and Gao, Chengqian and Ma, Liqun and Yurochkin, Mikhail and Maggs, John and Ma, Xuezhe and He, Guowei and Hu, Zhiting and Liu, Zhengzhong and Xing, Eric P.},
  year = {2025},
  huggingface = {https://huggingface.co/LLM360/K2-Think},
  url = {https://arxiv.org/abs/2509.07604},
}