IRSI

Institute for Recursive
Self Improvement

Research on recursive self-improvement in AI.

IRSI studies whether AI systems can improve their own capabilities—and their ability to make further improvements.

Revise the improvement process, build a candidate, and test it. Development evidence feeds into the next revision.
Can a system get better at making its next improvement?

First project

IRSI-Bench is our first project: a benchmark for models’ ability to perform recursive self-improvement. We are developing evaluations that compare improvement across generations, account for the resources used, and test whether gains hold on new tasks.

Read about the benchmark

Research areas

  • Recursive improvementHow changes to an agent affect the improvements it makes next.
  • Agent communicationWhen sharing knowledge or internal state improves joint performance.
  • Evaluation methodsHow to measure progress, cost, and reliability across generations.

About

IRSI was founded by Peter Seelman. Our first project is IRSI-Bench.

peter@irsi.ai