ScienceRSI: Small expert models lift a 27B target by 6.7 points across seven benchmarks
jiqizhixin · x · 2026-11-04
- A preprint from Scinetics and university collaborators studies whether compact scientific specialist models can supply useful training signals to another model, organized around specialization, verification, transfer and reuse.
- Instead of copying expert answers, contributions are checked against shared scientific requirements; a specialist's locally better calculation, mechanistic explanation or evidence judgment guides the target only where it is weaker, combined with task rewards.
- Experiments cover equal-size consolidation (4B→4B) and cross-scale transfer (4B→8B, 8B→27B with a Qwen3.5 target).
- In a separate 8B→27B transfer, the unweighted mean across seven benchmarks rises from 87.3% to 94.0%, a 6.7-point gain; the 27B extension changes model generation, so it is an independent setting.
More from Research
- New Thread Tackles Long Video Generation Drift: Let Every Frame See the Whole Clip — mangahomanga · 2026-11-04
- Researchers seek small-scale Chinchilla scaling law replications in PyTorch or JAX — turing-math-labs · 2026-11-04
- TLAPS-Bench: 265 real-world TLA+ proof targets; top model GPT-6.1 Sol solves only 44.3% — tianyin_xu · 2026-11-04
- StreamDuo pairs a commentator with a wiki research agent to commentate unseen games — Zixian Gao · 2026-11-04
- Redwood researcher: AI research automation and runaway recursive self-improvement can come apart — deanwball · 2026-11-04
- UMAP 0.6 adds Adam and momentum layout optimizers, becoming defaults in 0.7 — leland_mcinnes · 2026-11-04