I am Livan Zhou, a First Class Honours Computer Science graduate from Coventry University. My work focuses on developmental agency—especially how agents decide what capability to develop next—governed memory, multi-agent coordination, and uncertainty-aware AI systems.
Each project is framed as a falsifiable research object with public evidence, explicit limitations, and a concrete path toward stronger evaluation.
Uncertainty-Gated Evidence Escalation for Open-World AI-Generated Text Detection
A two-stage detection architecture that escalates ambiguous cases from lightweight distributional evidence to auditable cross-family comparison, rather than forcing every sample through a single opaque classifier.
Evidence-weighted reopening for long-horizon agent memory
A governed-memory mechanism for agents that must preserve useful conclusions without turning earlier reasoning into permanent dogma. Memories remain revisable when later evidence crosses an explicit reopening threshold.
Sparse intervention-calibrated developmental leverage
A compute-aware developmental-agency project that estimates how training one currently learnable skill may transfer to capabilities the agent still lacks. It combines a low-cost, agent-specific transfer proxy with sparse real interventions to decide what the agent should develop next under a fixed budget.
The long-term objective is not another workflow wrapper around an LLM, but an agent architecture that persists across time and can reorganise its own capabilities from consequences.
How can an agent estimate which currently learnable capability will most expand the competence it still lacks, then revise that developmental priority as its own skills and experience change?
How can memory preserve identity and accumulated competence while remaining auditable, revisable, and resistant to contamination or premature lock-in?
How should agents communicate memory, divide emerging competence gaps, and coordinate under decentralised information without collapsing into one shared prompt?
I treat repositories, experiment logs, manuscripts, demonstrations, and negative results as parts of one evidence system. The aim is to make the path from claim to artefact visible.
Public repositories before broad claims
Controlled experiments before narrative conclusions
Explicit uncertainty instead of forced certainty
Reproducible artefacts that a reviewer can inspect
I graduated with First Class Honours in Computer Science from Coventry University. My recent work moves from robust AI-generated text detection toward the deeper problem of agents that maintain cognitive continuity, estimate their own developmental leverage, and coordinate as teams.
I am particularly interested in research environments where conceptual ambition is matched by controlled simulation, reproducible engineering, and clear empirical baselines.