In the BabyLM 2026 Strict‑Small track, Qiushi Engine carried out a long‑horizon, end‑to‑end autonomous research program under a budget of ten million corpus words and one hundred million cumulative word presentations. The workflow is divided into three stages.
Stage I built a frontier model by using compact restatements, reinvesting the budget, and applying residual incremental learning. Stage II revealed that exact repetition and aligned restatement lead to different patterns of context usage depending on the target relation and prediction window. Controlled tasks showed that recovering performance on familiar inputs does not guarantee that the model can still exploit the learned computations on unseen inputs. These observations support a testable data‑efficient learning principle: organize experience around the contextual dependencies required for prediction; design visible information, supervision, and preservation separately; and evaluate learning, generalization, and retention independently.
Stage III retained the source text, masked more local clues, supervised selected targets, and preserved predictions on ordinarily masked inputs. Two continuation seeds derived from the same parent outperformed ordinary continuation on the nine‑metric aggregate, raising the overall score from 42.02 to 42.25. The second generation achieved the highest Overall in the public Strict‑Small snapshot released on 8 September 2026.
Further studies addressed compression, relational anchors, shared representations, and evaluation metrics. The models are released on Hugging Face, with code and research records on GitHub. Together, these stages illustrate Research RSI (recursive self‑improvement of the research process): scientific understanding and methodological innovations reshape subsequent questions and designs, while new experiments test and refine them.
Review