RSI means recursive self-improvement. A system improves a capability and uses the accepted capability in a later improver role to help improve itself again. The feedback into subsequent self-improvement is the central connection.
A scientific agent may retain useful notes, transfer a procedure, or find a better answer through repeated trials. These are valuable behaviors, but they do not by themselves establish that connection. Nor is it enough for a fixed external editor to keep modifying a target: the accepted capability must actually help perform later improvement work.
The capability might reside in model weights, code, tools, or procedures. RSI does not require weight updates or rewriting updater code every round. For example, an accepted diagnosis procedure could help the system identify and repair a later weakness in its own research process.
A loop is a mechanism; its benefit needs an experiment
First identify what changes, who or what proposes and validates the change, and where the accepted capability enters the next improver role. Version records and execution traces should show this path across cycles. This establishes the proposed recursive architecture, not that recursion improves results.
Then test the causal claim. Compare subsequent improvement with a fixed improver and with a condition that withholds the accepted capability from the later improver role. Match tasks and resources, include all improvement costs, and report rejected changes and regressions. A rising final score alone cannot isolate the benefit of recursion.
Scientific transfer is a further claim. It needs assessment on new scientific problems, with appropriate reference methods and checks of scientific validity. An agent benchmark outside science cannot establish that transfer.
The distinctions matter in current research. MetaSkill-Evolve v1, §3.4–3.5 reuses updated meta-skills in its improvement pipeline. AIDE² v1, §3.6 tests an evolved agent as a subsequent improver, but does not establish superiority over its reference improver. ScienceBuddy v1, §5 explicitly notes that its fixed reflector prevents stronger task performance from establishing a stronger improvement mechanism. These are distinct evidence boundaries, not interchangeable demonstrations of RSI for science.
Start with a scientific question
A contribution can begin with a real research question, an available source of feedback and a plausible way to check an answer. A finished agent implementation is not required. Contributors can leave the recursive mechanism open while developing the scientific environment.
The molecular and oscillator teaching examples illustrate scientific inputs and checks; they are not RSI benchmarks. The materials experiment-planning direction is a proposal with source selection still pending. A campaign must add the changing system, the feedback into its improver role, repeated-cycle records, controls, and independent scientific assessment before it can test an RSI claim.
Our current stage
We are developing the protocol and starter examples. There is no independently reviewed RSI4Science benchmark release, leaderboard or demonstrated RSI4Science result yet. The collection links to the original MolP2SBench benchmark and provides a teaching adaptation. The teaching resources and original-source links are available now. You can prepare a dataset or source proposal and contact the team while the collection and review process are developed.
Browse datasets, read the draft protocol, or propose a scientific problem.
Clarification, October 3, 2026: this note now defines the recursive connection explicitly and separates it from experience retention, task improvement, and evidence of a causal benefit.
Correction, October 12, 2026: the materials direction remains a proposal awaiting source selection, not an available starter resource. This does not change the evidence required for an RSI claim.
Editorial publisher: RSI4Science / DUAL Lab. AI-assisted research, drafting and editing; primary sources checked by automated agents. Updated .
Corrections are welcome. Prepare a source-backed correction. Publication does not imply independent experimental reproduction or endorsement of the work discussed.
Turn a research question into a task