A new training method called Transfer-Aware Curriculum cuts the guesswork out of how AI models divide their study time across math, code, and science.
Reinforcement learning with verifiable rewards - RLVR - lets models train on problems with checkable answers. The wrinkle: when you train across multiple subjects at once, the order and frequency of those subjects matters a lot, and most teams either fix that schedule in advance or tune it by hand. Researchers behind TAC instead treat subject selection as a live optimization problem. The system monitors not just where a model is improving, but whether progress on one subject spills over into others - using gradient geometry signals already produced during training, adding less than 1% overhead. Tested on Qwen3-1.7B and Llama3.2-3B across six reasoning domains, TAC beat proportional random sampling, a hand-crafted schedule, and a learnability-only approach by up to 2.8 percentage points, a 10% relative gain.
The result matters because multi-domain reasoning is where the real capability gaps show up. A model that aces math but stumbles on code - or vice versa - is brittle in practice. TAC offers a principled, low-cost way to close those gaps without a human scheduling consultant. The finding that removing the transferability signal sharply degrades performance is the more interesting result: it suggests current curricula are leaving meaningful gains on the table.
The technique works on small models, which makes it accessible, though whether it holds at frontier scale remains untested.