International Conference on Learning Representations (ICLR) 2025
About
I am a Senior Research Fellow and the Director of the AI Safety Research Center at Beijing-AISI. My work integrates brain-inspired intelligence with AI safety & alignment, focusing on embedding human moral and ethical values into large language models through both internal alignment and external oversight. In the short term, I develop secure, hardware–software co-designed platforms for safe and efficient deployment. Long term, my goal is to build real intelligence: brain-inspired systems with intrinsic safety and genuine cognitive abilities.
Contact
Email: dongcheng.zhao@beijing-aisi.ac.cn
Prospective students: include a short bio and links to code/publications. Typical response time: 1–2 weeks.
Education
Ph.D., CAS Institute of Automation (supervised by Prof. Yi Zeng)
B.Sc., School of Math & Stats, Xidian University
News
07/2025
One paper accepted by ICCAD.
06/2025
One paper accepted by ICCV.
05/2025
One paper accepted by iScience.
02/2025
One paper accepted by IJCAI 2025.
02/2025
One paper accepted by ICLR 2025.
02/2025
Two papers accepted by AAAI 2025: StressPrompt and EventZoom.
Awards & Honors
- 2023 Cell Press China Paper of the Year Award (Interdisciplinary Science)
- 2022 Cell Press China Paper of the Year Award (Interdisciplinary Science)
Selected Publications
Pattern Recognition 2024, 116: 108–123
Proceedings of the National Academy of Sciences (PNAS) 2023 120(18)
Grants & Projects
2025.01–2027.12
国家自然科学基金委员会 青年科学基金 — 融合神经元精细建模的多神经环路以及学习机制协同的脉冲神经网络研究 (主持)
2025.07–2026.07
企业横向 — 大模型对齐机制研究 (主持)
2024.01–2027.12
国家自然科学基金委员会 面上项目 — 基于大脑组织结构和神经回放多样性的连续学习方法研究 (参与)
2024.09–2026.09
北京市科学技术委员会 中央引导地方专项 — 基于外部监督对齐与内部机制对齐的超级对齐关键技术研究及示范应用 (参与)
Research Projects
PANDAGUARD
PandaGuard structures LLM safety evaluation as an interactive system with Attackers, Defenders, and Judges. It supports plug-and-play algorithms, multiple inference backends, and flexible human-in-the-loop interfaces.
Spiking Transformer Evaluation Platform (STEP)
STEP reproduces state-of-the-art Spiking Transformer models and offers a unified pipeline for classification, segmentation, and detection. It enables fair, reproducible comparisons with energy modeling and easy extensibility.
Brain-inspired Cognitive Intelligence Engine (BrainCog)
BrainCog is an open-source SNN-based cognitive intelligence engine for brain-inspired AI, embodied AI, and brain simulation, supporting large-scale spiking computation and neuro-symbolic components.
Open Positions
I'm looking for highly motivated students and collaborators in:
- LLM Safety & Alignment: red-teaming, defense, safety–utility tradeoffs
- Brain-inspired & Neuromorphic AI: spiking transformers, energy-efficient learning
- Multimodal Learning: representation learning, alignment, dataset/benchmark design
How to apply send an email with a short bio (1–2 paragraphs), and links to representative code/papers or GitHub/Google Scholar. If available, include a transcript or project portfolio. Please include “[Application] YourName – Topic” in the subject.
Professional Service
- Program Committee / Reviewer: IJCAI, AAAI, ICML, NeurIPS, ICLR
- Journal Reviewer: IEEE TNNLS, IEEE TIP, IEEE TETCI
- Open-source: Core contributor to PandaGuard, STEP, BrainCog
Contact
Email
dongcheng.zhao@beijing-aisi.ac.cn
Affiliation
Beijing-AISI — AI Safety Research Center
Links