Large reasoning models (LRMs) achieve strong performance by generating explicit reasoning traces before producing an answer. However, this comes at a cost: they tend to overthink, applying uniformly heavy computation to queries regardless of difficulty. The new paper identifies this as a key inefficiency in current reasoning systems.

The authors propose Root-token Policy Optimization, a method designed to make reasoning adaptive. By optimizing decisions at the level of root tokens, the model learns when to engage in explicit reasoning and when to skip it, potentially reducing unnecessary computation while preserving performance.

The approach directly targets the uniform allocation of computation that characterizes overthinking. While the abstract does not include experimental details, the proposed method signals a shift toward more efficient, query-dependent reasoning in large models.