A new paper on arXiv argues that data strategy is central to training large language models. Rather than treating data as a static input, the authors frame it as an active control variable: how data is selected, weighted, and scheduled shapes the skills a model ultimately learns.
The paper's key claim is that these choices determine what a model can acquire from a fixed data budget. That makes data strategy a direct lever on capability, not just a preprocessing step. The authors propose a unified perspective that spans the whole training lifecycle, from initial curation to later training phases.
Because the source is a single position-style paper, the claims are presented as a framework rather than empirical results. Still, the emphasis on data as a strategic resource aligns with a broader shift in LLM research toward understanding training dynamics beyond architecture and scale.