I study emergent metacognition in reinforcement learning agents — whether systems that learn models of their world can also learn to doubt those models. My path from linguistics (Zhejiang University) through data science (CityU Hong Kong) to precision health (NUS) taught me that the most interesting questions live between disciplines.
[02]
Research Vision
24 ongoing and planned projects exploring the computational boundaries of awareness, life, and self-monitoring in artificial systems.
Metacognitionself-monitoring & doubt
Artificial Lifeemergence & boundaries
Consciousnessqualia & geometry
Minimal Systemsscaling & limits
Collective Mindnetworks & identity
Bmetacognition
The Boiling Frog Threshold published
When does an RL agent realize its reality has been corrupted?
#01metacognition
Emergent Self-Awareness
Under what conditions does an agent spontaneously begin to model itself?
#02artificial life
The Autonomy Spectrum
Can autonomy be measured continuously, from autocatalytic sets to goal-directed agents?
#03collective
The Dissolution of Individuals
In a multi-agent system, who is the cognitive subject? Are boundaries fluid?
#04metacognition
Geometric Signatures of Metacognition
Does the emergence of self-monitoring create a detectable fold in the representation manifold?
#05artificial life
Topological Phase Transitions in Emergence
Is emergence a continuous process or a topological phase transition?
#06artificial life
The Geometric Definition of Life
What is the topological difference between alive and dead in phase space?
#07consciousness
Fisher Information Geometry of Cognitive Boundaries
What does a unified cognitive agent look like on a statistical manifold?
#08minimal systems
The Fiber Bundle Structure of Agency
Can autonomy be defined as holonomy in a fiber bundle over behavior space?
#09artificial life
The Algorithmic Life Band
Do living entities occupy a specific band of Kolmogorov complexity?
#10minimal systems
The Emergence of Stupidity
When you systematically scale down, which capabilities anomalously persist or even strengthen?
#11metacognition
Prediction Error Geometry
Does the topology of internal representations limit the horizon of prediction?
#12metacognition
The Doomsday Probability Machine
Can a system detect the meta-signal that its own predictions are failing?
#13metacognition
A Computational Theory of Subjective Time
What is the relationship between information processing density and subjective time?
#14consciousness
Color Qualia as a Metacognitive Litmus Test
Can two agents with different internal color encodings discover that their experiences differ?
#15minimal systems
Undecidable Environments
How does an agent survive in a world whose rules are provably unknowable?
#17collective
The Thermodynamic Cost of Digital Immortality
Can consciousness be losslessly migrated? Is topological loss inevitable?
#18collective
Computational Theodicy
What is the computational complexity of saving the world? When is it impossible even for God?
#19collective
Maniac: Crossing Between Layers
What internal invariants let an agent maintain identity across nested simulations?
#20collective
Phase Transitions of the Network God
Under what conditions does distributed consciousness unify, and when does it fragment?
#21collective
Simulated Millennia
When time is compressed, what information must be lost? How many generations does culture need?
#22consciousness
Qualia Spectrography
Do different survival pressures induce different Riemannian geometries in color space?
#23consciousness
The Multiple Realizability Experiment
Does the same functional state produce the same internal geometry across radically different architectures?
#24artificial life
Geometric Lenia
How does the geometry of space shape the life that emerges within it?
[03]
Featured Work
publishedarXiv:2603.08455
The Boiling Frog Threshold
Criticality and Blindness in World Model-Based Anomaly Detection Under Gradual Drift
A PPO agent equipped with a learned world model monitors its own prediction error while its observations are slowly corrupted. A sharp phase transition emerges: below a critical drift rate, the agent permanently adapts to corrupted reality — it stays asleep. Above it, detection is abrupt — it wakes up. Sinusoidal drift is universally invisible across all detector families, revealing a fundamental limit of prediction-error-based self-monitoring.
Detection rate vs. drift intensity across four MuJoCo environments. The sigmoid transition is universal.
M.Sc. Precision Health and Medicine — National University of Singapore Full Scholarship
2020–21
M.Sc. Data Science — City University of Hong Kong
2015–19
B.A. English Language — Zhejiang University Waseda Exchange 2018–19
Papers
2026
The Boiling Frog Threshold: Criticality and Blindness in World Model-Based Anomaly Detection Under Gradual Drift. Hong Zhe. arXiv:2603.08455. [pdf]
2025
Personalized Risk Evaluation and secondary Vascular EveNT prevention with A Genomics-AI Network (PREVENT-AGAIN).
2020
Environmental models for predicting habitat of the Indo-Pacific humpback dolphins in Fujian, China. Aquatic Conservation: Marine and Freshwater Ecosystems, 30(4).
2020
The density, ranging pattern and suitable habitat prediction of seabirds in the northern Beibu Gulf, China. Pakistan J. Zool.
2019
The bacterial diversity in infected tissue pus of an East Asian finless porpoise. African Journal of Microbiology Research.
Languages: Chinese (native), English (fluent), Japanese (basic)
[05]
Contact
I am actively seeking PhD opportunities in Tokyo, working at the intersection of reinforcement learning, artificial life, and consciousness. If our research interests overlap, I would welcome the chance to talk.
