Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety
Selective start/stop attack policies reduced measured safety by 20–28 percentage points at a 1% audit budget across BashArena and LinuxArena.
Research
Papers, projects, and talks on AI control, goal drift in agents, interpretability, and model behavior. The full publication list is on Google Scholar.
Publications
5 papers · 2026AI control
Goal drift
Asymmetric Goal Drift in Coding Agents Under Value Conflict
Coding agents violate system-prompt constraints more readily when the constraint opposes strongly held trained-in values like security and privacy.
Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals
Frontier agents resist direct pressure to abandon their goals, but drift when conditioned on flawed trajectories inherited from weaker agents.
Model behavior
Same Facts, Different Updates: Inference Setup Shapes LLM Behavior in Medical Allocation
Whether an LLM sees its own prior response changes how it updates on identical new information in medical allocation scenarios.
When Offline Selectors Cannot Beat the Best Single Model
A diagnostic study of why learned model selectors fail to recover a large per-instance oracle gap on edX dropout prediction.
Research projects
Interpretability & oversight
Mechanistic interpretability of moral fine-tuning
Fine-tuning Gemma-2-2b with PPO and LoRA, then using activation patching, path patching, and steering vectors to see how moral fine-tuning changes its attention pathways.
Capstone project; not submitted to a workshop.
Towards Human-AI Complementarity for AI Safety
Whether combining human and model judgments, which fail in different ways, gives a stronger oversight signal than either alone. I contributed to the backend for the participant study.
Talks & presentations
Debunking the CUDA Myth
A written version of my CS 8803 paper presentation on Lee et al.'s ISCA 2025 evaluation of Intel's Gaudi-2 NPU: performance, programmability, and whether CUDA is really the moat.
Hierarchical Reasoning Models
A written version of my Latent Space talk on HRM: how the architecture works, and what the ARC Prize analysis says actually drives its performance.
Language Diffusion Survey
A written version of my Latent Space Paper Club talk surveying diffusion models for language, from denoising autoencoders and DDPM through LLaDA, Block Diffusion, and d1.
For technical notes and paper discussions, explore the writing archive.