Mechanistic interpretability · Latent reasoning · Learning agents

Jiayu Yang杨家宇

Ph.D. student building a causal understanding of how language models reason, remember, and learn.

Guangzhou, China

I am a Ph.D. student in the Information Hub at HKUST (Guangzhou), working with Chengwei Qin, Zhijiang Guo, and Yutao Yue. I am also a research intern at Tencent Hunyuan.

I study the internal mechanisms of large language models and turn that understanding into better post-training. I care about explanations that are causal and testable, especially for long-horizon reasoning and agentic reinforcement learning.

Mechanistic interpretability

Tracing factual recall and knowledge editing through concept-level and neuron-level mechanisms.

Latent reasoning

Making hidden-state recurrence trainable with on-policy RL and open to causal intervention.

Learning agents

Understanding credit assignment and policy behavior in long-horizon software tasks.

Updates

Latest news

View all

Aug 24, 2026	Three recent milestones: SWITCH was accepted to EMNLP 2026, Dynamic-V2C was accepted to ECCV 2026, and CAT was published in TMLR.
Feb 01, 2026	I joined Tencent Hunyuan as a research intern, working on LLM reasoning and world models.
Jan 27, 2026	Our paper ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall was accepted to ICLR 2026.
Mar 24, 2025	I joined LARK Lab and began working with Zhijiang Guo on LLM reasoning and mechanistic interpretability.
Sep 01, 2024	I started my graduate studies in Artificial Intelligence at HKUST (Guangzhou).

Research

Selected publications

View all

ICLR

ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall

Jiayu Yang, Yuxuan Fan, Songning Lai, Shengen Wu, Jiaqi Tang, and 3 more authors

In International Conference on Learning Representations, 2026

arXiv Bib

@inproceedings{yang2026ace,
  title = {ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall},
  author = {Yang, Jiayu and Fan, Yuxuan and Lai, Songning and Wu, Shengen and Tang, Jiaqi and Kang, Chun and Guo, Zhijiang and Yue, Yutao},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
}

TMLR

CAT: Concept-Level Backdoor Attacks for Concept Bottleneck Models

Songning Lai, Jiayu Yang, Yu Huang, Lijie Hu, Tianlang Xue, and 4 more authors

Transactions on Machine Learning Research, 2026

arXiv Bib

@article{lai2024cat,
  title = {CAT: Concept-Level Backdoor Attacks for Concept Bottleneck Models},
  author = {Lai, Songning and Yang, Jiayu and Huang, Yu and Hu, Lijie and Xue, Tianlang and Hu, Zhangyi and Li, Jiaxu and Liao, Haicheng and Yue, Yutao},
  journal = {Transactions on Machine Learning Research},
  year = {2026},
}

EMNLP

Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning

Jiayu Yang, Chao Chen, Shengen Wu, Yinhong Liu, Yuxuan Fan, and 4 more authors

In Conference on Empirical Methods in Natural Language Processing, 2026

arXiv Bib

@inproceedings{yang2026switch,
  title = {Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning},
  author = {Yang, Jiayu and Chen, Chao and Wu, Shengen and Liu, Yinhong and Fan, Yuxuan and Li, Lujundong and Lai, Songning and Qin, Chengwei and Guo, Zhijiang},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2026},
}

ECCV

Dynamic-V2C: Editable and Continual Vision-to-Concept Bottleneck Models via Influence Functions

Songning Lai, Shaofeng Liang, Jiayu Yang, Ninghui Feng, Yuxuan Fan, and 1 more author

In European Conference on Computer Vision, 2026

Bib

@inproceedings{lai2026dynamic,
  title = {Dynamic-V2C: Editable and Continual Vision-to-Concept Bottleneck Models via Influence Functions},
  author = {Lai, Songning and Liang, Shaofeng and Yang, Jiayu and Feng, Ninghui and Fan, Yuxuan and Chen, Wenshuo},
  booktitle = {European Conference on Computer Vision},
  year = {2026},
}

ACM MM

Learning New Concepts, Remembering the Old: Continual Learning for Multimodal Concept Bottleneck Models

Songning Lai, Mingqian Liao, Zhangyi Hu, Jiayu Yang, Wenshuo Chen, and 4 more authors

In Proceedings of the 33rd ACM International Conference on Multimedia, 2025

Oral Bib

Selected as an oral presentation at ACM Multimedia 2025.

@inproceedings{lai2025learning,
  title = {Learning New Concepts, Remembering the Old: Continual Learning for Multimodal Concept Bottleneck Models},
  author = {Lai, Songning and Liao, Mingqian and Hu, Zhangyi and Yang, Jiayu and Chen, Wenshuo and Xiao, Hongru and Tang, Jianheng and Liao, Haicheng and Yue, Yutao},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  pages = {12314--12322},
  year = {2025},
}