Current direction / AI for Science
Agentic Harnesses for AI4S
How can a foundation model do more than produce a plausible scientific answer? I am interested in systems that keep the work itself alive: executable artifacts, empirical feedback, and a path for revision.
The question
Scientific work rarely has one exact answer. A useful agent must explore alternatives, compare them with observations, preserve what works, and decide what to revise next. The harness around a model determines what it remembers, what feedback it can use, and how its next attempt differs from the last.
This is where my interests in agentic AI, recursive self-improvement, and text optimization meet AI for science. I want to understand which parts of a scientific workflow should live in language, which should live in code, and how evidence can guide their evolution.
Projects and papers
SimAuthor retains and revises executable simulators under weak empirical feedback. It separates scalar scores from structured discrepancy feedback and accumulates reusable implementation mechanisms across attempts.
Where this can grow
SimAuthor is one instance of a broader research direction. This page will collect future work on agent harnesses for scientific discovery, persistent artifacts, search over programs and strategies, and the relationship between feedback and self-improvement.
Interested in agentic science or simulator authoring? Let's compare notes ↗