How do token limits affect reasoning traces?

Token limits significantly influence reasoning traces in Large Reasoning Models (LRMs). As problem complexity increases, there is an observable pattern where LRMs initially use more tokens for reasoning but then exhibit a counterintuitive reduction in reasoning effort despite remaining below their generation limits. This trend indicates a fundamental scaling limitation in their reasoning capabilities relative to problem complexity, leading to performance collapse at higher complexities^[1].

Moreover, the inefficiencies in reasoning processes become evident as models often explore incorrect solutions, wasting token budgets, and their ability to self-correct diminishes^[1]. Ultimately, these token dynamics affect the overall effectiveness of reasoning in these models.