AI systems can increasingly self-evaluate and modify parts of their own prompts, tools, memory, code, data, and workflows based on feedback. The unresolved question is how to make sure these feedback loops produce genuine, repeatable improvement rather than benchmark overfitting or unstable changes. This track is for researchers and senior engineers experimenting with agents, automated research systems, or iterative model-improvement pipelines.

Talks cover definitions and levels of recursive self-improvement; generate–evaluate–select loops for system modification; modification of prompts, tools, memory, code, data, and models; measurement, credit assignment, regressions, and benchmark overfitting; and sandboxing, oversight, stopping conditions, and control boundaries. Expect discussion of the tradeoffs involved, experiments happening at the frontier, and a sober view of current limitations. Attendees will understand how to distinguish RSI from ordinary iterative optimization, and how to design self-improvement loops with hard-to-hack metrics and stopping conditions.
Notes: Nothing to fact-check. Suhas’s text used as written; the hook is his first two sentences, lightly shortened.