Research portfolio

Research & publications

Ongoing investigations and published papers, with my role in each.

Normal and attacked pipelines: input image, imagined video, and downstream IDM action
Conceptual problem overview
Ongoing· Project lead · World-model safety

Targeted attacks on world models

I lead research on world-model hijacking: targeted adversarial attacks that aim to steer model behavior toward an attacker-chosen objective.

Methods and results will accompany a public release.

Standard AttnResidual compared with OASIS
OASIS, Figure 1 · Enlarge ↗
NeurIPS 2026· COLM 2026 Efficient Reasoning Workshop · Spotlight

Attention Sinks and Outliers in Attention Residuals

Haozheng Luo*, Haoran Dai*, Jingyuan Huang*, Shaoyang Zhang, Xi Chen, Eric Hanchen Jiang, Yijiang Li, Chenghao Qiu, Chenwei Xu, Zhenyu Pan, Haotian Zhang, Binghui Wang, and Yan Chen

Understanding how attention sinks and activation outliers interact with cross-layer routing, and using null-aware attention to improve model robustness.

My contributionI contributed controlled interventions and layer-wise representation analysis to study the relationship between attention sinks and activation outliers.
View citation
@misc{luo2026attentionsinksoutliersattention,
  title={Attention Sinks and Outliers in Attention Residuals},
  author={Luo, Haozheng and Dai, Haoran and Huang, Jingyuan and Zhang, Shaoyang and Chen, Xi and Jiang, Eric Hanchen and Li, Yijiang and Qiu, Chenghao and Xu, Chenwei and Pan, Zhenyu and Zhang, Haotian and Wang, Binghui and Chen, Yan},
  year={2026},
  eprint={2605.17887},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2605.17887}
}
Knowledge-state evolution for users with different prior knowledge
KnowSim, Figure 1 · Enlarge ↗
EMNLP 2026· Main Conference

KnowSim: Evaluating Information Calibration in LLM Assistants with User Simulators that Learn

Yoonjoo Lee, Hyoungwook Jin, Tae Soo Kim, Shaoyang Zhang, Philippe Laban, and Q. Vera Liao

Evaluating whether assistants adapt to what a user knows, with user simulators whose explicit knowledge states evolve throughout a conversation.

My contributionI built experimental infrastructure for knowledge-state-aware simulation, conversation analysis, and the initial end-to-end human validation study, then iterated on these systems through pilot feedback and failure analysis.
View citation
@inproceedings{lee-etal-2026-knowsim,
  title={KnowSim: Evaluating Information Calibration in LLM Assistants with User Simulators that Learn},
  author={Lee, Yoonjoo and Jin, Hyoungwook and Kim, Tae Soo and Zhang, Shaoyang and Laban, Philippe and Liao, Q. Vera},
  booktitle={Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year={2026},
  pages={29750--29780},
  publisher={Association for Computational Linguistics},
  url={https://aclanthology.org/2026.emnlp-main.1307/}
}
Applied machine learning conceptual illustration
PSGEC 2024· Power System and Green Energy Conference

Knowledge Graph Enhanced Interactive Fault Diagnosis O&M Approach Based on Large Models of Electrical Power

Xiaoyun Tian, Bin Liu, Jianzhong Yang, Bo Ai, Huailiang Zhao, and Shaoyang Zhang

Combining knowledge graphs and large models for interactive fault diagnosis in electrical power systems.

My contributionCo-author on applied research at the intersection of domain knowledge and machine learning for power-system operations.
View citation
@inproceedings{tian2024knowledge,
  title={Knowledge Graph Enhanced Interactive Fault Diagnosis O&M Approach Based on Large Models of Electrical Power},
  author={Tian, Xiaoyun and Liu, Bin and Yang, Jianzhong and Ai, Bo and Zhao, Huailiang and Zhang, Shaoyang},
  booktitle={2024 4th Power System and Green Energy Conference (PSGEC)},
  year={2024},
  url={https://ieeexplore.ieee.org/document/10721152}
}

* Equal contribution. Publication links point to public versions.