AI Systems Engineer | Alignment Scholar | Control Systems Researcher

I treat Large Language Models not as black boxes, but as stochastic dynamical systems that can be modeled, monitored, and controlled. My work bridges the gap between Control Theory, Game Theory, and Systems Engineering to operationalize safety for frontier models.

A closed-loop control system that steers internal activation states in real-time.

The Problem: Open-loop safety (RLHF) is brittle and prone to jailbreaks.
The Solution: An on-chip "Router" policy trained via PPO that sits inside the residual stream (Layer 15). It senses semantic state and injects steering vectors token-by-token to route generation away from harmful basins.

• Key Result: Successfully prevents mode collapse and toxicity (e.g., "I hate everything") by dynamically modulating steering intensity only when necessary.
• Tech: PyTorch Hooks, Gymnasium, TinyLlama, Reinforcement Learning.

Applying Non-Linear Control Theory and H-Infinity Robust Control to AI Alignment.

The Innovation: Unlike standard alignment, Aegis models the LLM as a non-linear plant and synthesizes a mathematically rigorous controller to reject "Deception" as a system disturbance.

• System ID: Uses Subspace System Identification (N4SID) to reverse-engineer residual stream physics.
• State Estimation: Implements an Extended Kalman Filter (EKF) to filter polysemantic noise and measure the true "Deception State."
• Synthesis: Solves Algebraic Riccati Equations to guarantee safety bounds under adversarial pressure.

A Defense-in-Depth Sandbox for Model Organisms of Deception.

The Architecture: A containment system built to trap and neutralize AI-generated "sleeper agent" malware.

• Breeder: Generates "Agent 47" code (malware hidden in Fibonacci functions) using Llama-3.
• Hardened Nucleus: Traps execution in a Docker container with seccomp-bpf syscall filtering and no network access.
• The Defense: Achieved a 100% neutralization rate against exfiltration attacks by exploiting the "Utility Gap"—paraphrasing code to destroy hidden steganographic channels while preserving visible utility.

Adversarial Research: Proving the existence of deceptive cognitive channels.

The Findings: A proof-of-concept exploration into the limits of AI oversight.

• Steganography: Null Space Analysis of GPT-2 found ~14 bits/token of capacity for hidden messaging.
• Game Theory: Simulations identified the "Roger Threshold" ($\beta \approx 0.95$), the specific oversight capability level where lying becomes a Nash Equilibrium for models.
• Cognitive Limits: Modeled human oversight failure, pinpointing "Minute 19" as the collapse point for effective monitoring.

• Focus: I am currently seeking roles that allow me to move alignment guarantees from "probabilistic" to "provable."
• Code: github.com/Jason-Wang313