Effects of data proportion at scales near model capability limits

Investigate how the proportion of mathematics versus computer‑use data affects performance when training multimodal reasoning models at scales near their capability limits, and ascertain whether uniform performance persists or trade‑offs between these reasoning tasks emerge.

Background

Experiments conducted at a relatively small data scale show that increasing total data improves overall performance and that adding more data from one domain can also benefit another (e.g., math data improving some computer‑use benchmarks).

The authors emphasize that their present scale may not challenge model capacity; thus, it remains uncertain whether observed uniform gains will hold at larger scales or whether inter‑task trade‑offs will appear as models approach saturation.