Training reasoning models efficiently with difficulty-matched questions

Training LLMs to improve reasoning via RL is hindered by scarcity of appropriately difficult samples and the problem that samples become too easy as models improve. D²Evo addresses this by jointly training a Questioner (generates training problems) and Solver (answers them), with the Questioner calibrated to the Solver's current capability. The method mines medium-difficulty anchors from the Solver's mistakes and generates diverse questions at matching difficulty levels. On mathematical reasoning benchmarks, D²Evo outperforms existing generation-based approaches using fewer than 2K real samples and generalizes to broader reasoning tasks. Code availability not mentioned.