Research

The core challenge of intelligence is determining which past experiences are relevant to the current moment. If an agent assumes too much similarity, it over-generalizes. If it assumes too little, it fails to leverage its prior learning.

My work investigates how latent cause inference acts as a gatekeeper for this process. I model how the mind and artificial agents use context, curricula, and task similarity to decide whether to update an existing representation or create a new one. Understanding this decision boundary is crucial for understanding how and when agents fail on meaningful tasks.

One hypothesis guiding my work is that some hallucinations in LLMs may arise from lossy compression and context-dependent reconstruction—mechanisms that also contribute to human false memories. I test when compressed representations overgeneralize across distinct details and whether contextual inference can keep them separate.

AI systems can fail not only by complying with harmful requests, but also by selecting the wrong behavioral policy for the context. My work studies prompt-conditioned task selection: how system prompts, task framing, and clarity route a model among compliance, refusal, and other safety-relevant behaviors.

The goal is to distinguish robust safety from prompt-sensitive behavior and to build evaluations that reveal how context changes an agent's error profile.

There is a cost in using cognitive control. Allocating it effectively is key to both human skill acquisition and efficient machine learning.

I investigate how agents compute the Expected Value of Control (EVC) to decide when to exact effort versus rely on habit. In AI systems, this translates to dynamic compute allocation—knowing when a "System 2" deep-dive is necessary versus a "System 1" rapid response. This balance is critical for scaling performance without exponential cost.

Scientific discovery often involves navigating high-dimensional search spaces. We are building metascience tools to map these design spaces systematically.

By formalizing experimental design as a search problem, we can use automated methods to identify "blank spots" in the literature—regions of the parameter space that have been theoretically neglected but are physically plausible. This approach accelerates discovery by guiding researchers toward high-value, unobserved phenomena.