The "data mix ratio" in large-model training is almost a black art — slight changes in the web/code/scientific-literature ratio cause downstream capabilities to vary dramatically. Mainstream method RegMix depends on an implicit assumption: the data pool is static. Once a data source is refreshed or expanded, this assumption breaks directly — all the previous hundreds of small-model proxy experiments become void and can only be re-run from scratch. The Tsinghua team, in CausalMix posted to arXiv (2607.01104) on July 1, takes a different approach: treating the statistical features of the data pool as "covariates", the domain mix weights as "treatment", the whole problem is a standard causal inference task. They use 512 Qwen2.5-0.5B training runs to fit the Conditional Average Treatment Effect (CATE), then extrapolate the optimal mix to a new data pool of 800,000 documents and directly train a 7B model — no more rerunning proxy sweeps throughout. CausalMix's really interesting point is that it isolates the "confounding bias" from data pool changes — what it fits isn't "which ratio won" but "how does a change in ratio causally drive performance". The same framework directly extrapolates to Qwen3-4B-Base's long-chain-of-thought training without redesign; it's stably better than RegMix and other baselines on multiple downstream tasks. For those working on training infrastructure, this is the first paper worth reading in July. The most direct selling point is cost: running hundreds of small-model experiments for every new data pool isn't a small number in ten-million-dollar pretraining. CausalMix makes "tuning ratios" potentially decouple from data pool updates for the first time. Code is currently under review and not yet public.