ScienceBuddy connects harness revision with task-model training. A separate, fixed GPT-6 Astra auxiliary model diagnoses trajectories and edits instructions, skills, or exposed context settings. Accepted harnesses generate further experience; GRPO updates the task model. Neither step establishes that the harness editor itself improves. Version 1, §2.3–2.5.
The coupled experiment carries the model and harness through three cycles, each with ten harness-search steps and twenty RL updates. Separate experiments hold either model weights or the harness fixed. Scientific tasks include literature reading, database judgments, protocol troubleshooting, and gene or variant assessment. These studies support bounded scientific adaptation; they do not isolate a causal benefit from making the improver stronger. Version 1, §4.2–4.4.
The authors explicitly acknowledge that the reflector remains fixed, so improved task performance does not imply a stronger improvement mechanism. The fixed-model adaptation study uses controlled simulated feedback, rather than establishing longitudinal learning from everyday scientist use. We have not independently reproduced the results. Version 1, §5 and Appendix §7.2.
Why include this boundary case?
Returning better task trajectories to a fixed editor is not sufficient evidence that an accepted capability is operating as a later self-improver. Our interpretation is that ScienceBuddy is relevant model–harness co-adaptation research, with the stronger recursive claim unresolved.
For RSI4Science, the next question is whether an accepted scientific capability can help generate, assess, or execute a subsequent self-modification. An experiment should expose that path and compare it with a fixed improver under matched resources. This is a proposed study, not a reported result.
ScienceBuddy is a preprint we discuss, not a claimed partner or submitted task. Its source date and this Brief’s editorial date are separate.
Propose a scientific problem or read what counts as recursive self-improvement.
Clarification, October 3, 2026: the Brief now identifies the fixed modifier and treats the paper as a boundary case, rather than implying that procedural adaptation establishes RSI.
Editorial publisher: RSI4Science / DUAL Lab. AI-assisted research, drafting and editing; primary sources checked by automated agents.
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