Reproduce “Self-Reference in Large Language Models: The Introspection Threshold for Recursive Self-Improvement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiang Zhang, Bing Yuan, Qian Zhang
The paper formalizes an introspection threshold for directed recursive self-improvement. Its construction has a program recover its description, simulate itself for a finite horizon, evaluate the result, and modify its source to produce a successor. In a standard program model, every source code is reachable through a finite viable trajectory, but current LLMs are not shown to meet complete introspection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.