Schölkopf Team's TAC Curriculum Lifts Multi-Domain RLVR ~10%

Curriculum design in RLVR has always been the part of the recipe that never quite feels principled. You mix your training domains, you tweak the ratios by hand, you hope the mixture that worked last month still works this month. A new arXiv preprint from Yongjin Yang, Bernhard Schölkopf and collaborators tries to automate that decision for multi-domain RLVR, and the interesting part is less the size of the reported win than how cheap the authors claim it is.

The method is called Transfer-Aware Curriculum, or TAC. Rather than asking which domain the model is currently learning fastest, TAC asks which domain, if sampled now, would also help the rest of the suite. It estimates that using per-domain advantages and projected gradients, framed as a bandit-style online curriculum, and the paper reports the extra machinery introduces under 1% additional compute on top of standard RLVR. On a six-domain reasoning benchmark trained with Qwen3-1.7B and Llama3.2-3B, the authors see up to a 2.8 point improvement over learnability-only baselines, which they frame as roughly a 10% relative gain, and they report it is robust on imbalanced training datasets.

Why this is interesting if you are training reasoning models rather than reading papers: hand-tuned RLVR mixtures are one of the least reproducible parts of the current post-training loop, and they scale badly as you add domains. A near-free signal for which domain to sample next changes the ergonomics of that loop, and it changes it in the direction of "you can add more domains without doing more manual work."

The honest caveats are the usual ones for a small-model paper. The largest model tested is 3B, and gradient-geometry methods have a habit of behaving differently at scale. What the paper does not give you is a picture of how TAC behaves when domains actively conflict, or how the domain-selection trajectory evolves late in training when the easy transfer has been eaten.

If the result holds on larger models, it is the sort of thing that quietly becomes an opt-in default in open-source RLVR pipelines, because the overhead is small enough that there is no real reason not to try it.