Qiushi Engine Outlines Data-Efficient Language Modeling Principles

Qiushi Engine conducted an autonomous research program focused on data-efficient language modeling within the BabyLM 2026 Strict-Small benchmark, limited to a 10-million-word corpus and 100 million cumulative word presentations.

The project linked frontier advancement, principle discovery, and model improvement across three stages. Researchers utilized compact restatements, budget reinvestment, and residual incremental learning, observing that exact repetition and aligned restatements yield distinct context-use patterns depending on target relations.

The resulting approach increased the overall benchmark score from 42.02 to 42.25, achieving the top result in the public Strict-Small snapshot as of September 8, 2026. The research artifacts, models, and code have been released on Hugging Face and GitHub.