danielxu0208/rsi_eval is a lightweight Python SDK and CLI for evaluating the recursive self-improvement (RSI) technique — that is, approaches that improve by repeating a refinement procedure — in agent benchmarks. The user picks the improvement loop and the tasks to evaluate, and evaluates candidates within a single interface. There is no formal release yet; the repository was created on October 12, 2026.
Discussions about RSI often stop at qualitative claims. This tool attempts to run several benchmarks under one measurement framework so that they can serve as a common evaluation standard. The division of responsibility is clear: the improvement procedure itself is left to the user, while the tool only runs the evaluation and collects the results. It records evaluation results per candidate along with candidate snapshots, makes the task feedback it obtains usable for the next candidate, and allows interrupted tasks to be explicitly resumed. It comes in both SDK and CLI form, is written in Python, and is licensed under Apache-2.0.
That said, the public information contains no actual benchmark results and no comparison figures between candidates. Because the improvement procedure is the user's responsibility, it is hard to reproduce the self-improvement process as-is with this SDK alone. The repository has 137 stars, and the inertia score tops out at 4.7. The qualitative evaluation came out at novelty 7.0, technical 7.0, and quality 6.0. The attempt to unify the measurement frameworks of several studies is visible, but that standard will need more accumulated results before it takes hold.