A small expert can teach a bigger model: ScienceRSI lifts a 27B model to 94.0% across seven benchmarks
Synopsis
ScienceRSI splits scientific capability into specialization, verification, transfer and reuse, so an expert model contributes feedback only where it is genuinely better on the same scientific requirement. In the 4B-to-4B setting, scientific-reasoning average reached 55.0%, 2.0 points above GRPO; transferring from Qwen3-8B experts to a Qwen3.5-27B target raised the equal-weight average over seven benchmarks from 87.3% to 94.0%. Removing expert-contribution verification dropped the same 4B student's scientific-reasoning average from 55.0% to 50.6%.
Interpretation
Differences in scientific-model capability are handled at the level of specific steps rather than by comparing two models as a whole. Training one model good at computation and another good at mechanistic analysis does not automatically yield a more well-rounded scientific assistant; their strengths may complement each other, or their errors may enter the new model together. Across the 4B and 8B experiment groups, ScienceRSI ranked first among the compared methods on five scientific benchmarks; in the 4B-to-4B setting its scientific-reasoning average was 55.0%, 2.0 points above GRPO.
Expert feedback is verified before transfer, and a contribution is adopted only when the expert beats the current answer on the same scientific requirement. Handing all expert outputs to the student model does not guarantee improvement: on the same 4B student, GRPO+MOPD scored 50.1% on scientific reasoning, below GRPO's 53.0%. The ablation used the same expert pool; removing expert-contribution verification cut the scientific-reasoning average from 55.0% to 50.6%, and keeping the same supervision coverage while randomly relocating feedback lowered scientific-reasoning and domain-science averages by 6.3 and 12.1 points respectively.
Perspective
The demonstrated scope is 4B-to-4B same-scale integration, 4B-to-8B cross-scale transfer, and an extension transferring from Qwen3-8B experts to a Qwen3.5-27B target, evaluated on seven benchmarks spanning scientific reasoning, domain science and general reasoning. The 8B-to-27B group changes both target scale and model generation, so it is presented separately.
The report verifies capability transfer at different scales but does not show that repeated recursive rounds would keep accumulating gains; as target models strengthen, specialization training, task difficulty and verification capability must rise together. The best and runner-up marks in Table 2 are point estimates within the same method group and do not indicate statistical significance.