A research lab for self-learning systems
Recursive improvementfor machinesthat act.
When an AI agent acts in the world, it shouldn't just finish a job—it should get better at it. We study how autonomous systems learn from real-world practice to improve themselves with every run.
An agent can succeed and still learn nothing
Completion is not learning.
A self-improving system must become different because it acted—more capable where experience holds, less certain where it does not, and ready to revise both.
act→observe→test→revise→transfer
Question in progress
What does recursive self-improvement require?
01Learning from interaction
02Revising beliefs with evidence
03Transferring across encounters
Self-learningMemory & transfer
Memory is not yet learning
Remembering preserves experience. Self-learning must decide what the experience means—and whether it should change the next attempt.
Research questionsRecursive self-improvement
What changes after the first attempt?
A research agenda for agents that form hypotheses, test them through action, and revise what they carry forward.