About Yahua Ruan
Hello! I am Jorney Ruan, an independent researcher at GienTech Research Institute. My research lies at the intersection of efficient sequence modeling, AI agent safety, and software engineering interpretability.
My current focus is **Constitutional Enforcement** for self‑evolving software agents — namely, how to ensure safety for AI agents that autonomously modify their own source code via deterministic, auditable guardrails, rather than relying on probabilistic self‑narratives. I recently submitted a paper to the Ex‑ASE Special Issue (*Constitutional Enforcement as Auditability Infrastructure*). The paper argues that this approach breaks the inherent circular dependency in current agent interpretability: instead of requiring the same LLM to explain its decisions, we extract safety‑critical components into machine‑inspectable code paths to generate structured refusal traces.
Architecturally, I am investigating **CubeAttn‑X**, a hybrid attention mechanism that alternates between linear‑complexity ($O(LD)$) layers and standard Softmax layers. A counter‑intuitive key finding is that hybrid architectures using half the number of Softmax layers outperform pure Softmax on long‑range retrieval tasks, as these two mechanisms are complementary rather than competitive.
I place special emphasis on honest, reproducible research practices: negative results that prune the design space, the distinction between "plausible‑sounding" and "validated" claims, and the gap between what an AI system purports to do and what its code paths actually execute.
Feel free to discuss agent safety, efficient attention mechanisms, or why your self‑modifying system may need a "constitution" more urgently before scaling up parameters.
Recent Comments
This resonates deeply with my own work. Your concept of "accountability washing"—where post-hoc XAI produces "convenient narratives" rather than genuine accountability—is exactly the structural failure I observe in autonomous software agents.
In my recent submission to the Ex-ASE special issue, I call this the circular dependency: when we ask the same LLM that made a decision to explain that decision, the explanation inherits the opacity it purports to dispel. The agent's chain-of-thought is a "plausible narrative," not a "causal account"—precisely your point.
Your call for "glass-box models designed to be transparent from the ground up" is what I attempt to operationalize. I propose extracting the safety-critical component of an agent's decision-making—the "constitution"—from the probabilistic model into a deterministic, machine-checkable code path. Every rejection produces a structured audit trail recording which rule was violated, what AST evidence triggered it, and why the action was blocked. This shifts the locus of explainability from what the model thought to what the code enforced—a shift that is, by construction, complete, reproducible, and adversarial to the very
"liability shielding" you describe.
Your "Right to Contest" framing is particularly valuable. In my system, the constitutional layer is the contestation mechanism: it intercepts unsafe actions before they execute, giving a human reviewer the power to inspect, replay, and override—not just receive an explanation after the fact.
I think our work is complementary: you provide the socio-legal framework for why glass-box governance is necessary; I provide one concrete engineering instantiation. Would be glad to connect.