ScienceRSI: Recursive Scientific Self-Improvement

Abstract

Scientific AI can improve cumulatively when capabilities developed in one model become useful learning signals for the next. We introduce ScienceRSI, a framework for recursive scientific self-improvement organized around specialize, verify, transfer, and reuse. Compact experts develop complementary scientific capabilities; shared evaluation identifies their useful contributions; a target model integrates those contributions and can serve as the starting point for further specialization. This organization connects reasoning, model scaling, and iteration through reusable scientific capabilities. We evaluate its transfer step in 4B→4B, 4B→8B, and 8B→27B configurations. ScienceRSI ranks first among the compared methods on all five science benchmarks for the 4B and 8B students. Relative to rubric-based reinforcement learning, scientific-reasoning averages improve by 2.0 and 1.9 percentage points, and domain-science averages by 3.3 and 2.1 points, respectively. In a separate cross-generation 27B extension, reported scores improve over Vanilla on all seven benchmarks, with the unweighted mean rising from 87.3% to 94.0%. These results support verified capability transfer across model scales as a practical foundation for ScienceRSI. Accumulating these gains over repeated rounds is the next step toward sustained recursive scientific improvement.