Reproduce “The RSI Tax: When Recursive Self-Improvement Is Not Worth the Cost” 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
Yitong Zhang, Ang Li, Yisen Wang
The authors measure an agent’s total token cost as task execution plus the RSI Tax for self-improvement, including rejected update attempts. Simple Reflection with Notes (SRN) revises one bounded note from task failures and supplies it later. Across four model–environment settings, SRN uses 68.2% fewer total tokens than three other update methods, with higher or tied observed accuracy; results cover only the studied streams.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.