Research

I’m interested in how language-model systems can improve themselves without changing their weights, and in what that improvement costs. Concretely, that means two threads:

• Self-improving agent harnesses. The code around a model (how it decomposes a problem, recurses, verifies, and recovers) often matters as much as the model itself. Can a model find the recurring failures in its own behavior and rewrite that code, in a way that still holds up on inputs much longer than it was optimized on?
• Efficient AI. Reasoning gains have an energy and carbon cost that rarely gets reported next to accuracy. I’m working on making that cost measurable and comparable.

Broader interests: recursive self-improvement · agentic systems & harness engineering · recursive language models (RLMs) · long-horizon reasoning · efficient AI systems.

Both projects below are ongoing. I report results only once the experiments are finished.