Train Long, Think Short: Curriculum Learning for Efficient Reasoning

We study curriculum learning strategies that train large reasoning models with long chains of thought but encourage short, efficient inference. We characterize when reducing test-time compute preserves accuracy, and propose a training schedule that closes the gap between long-train / short-test and long-train / long-test regimes across mathematical reasoning, multi-hop QA, and code generation benchmarks.

Under review (arXiv:2508.08940)