Length-Controlled Reasoning

• Length-Controlled Reasoning is the study of managing chain-of-thought tokens to balance correctness, latency, cost, and safety in diverse problem scenarios.
• It promotes selective concision by advocating brief outputs for simple problems and extended deliberation for complex challenges.
• Key experiments on GSM8K, MATH500, and AIME2024 demonstrate that dynamic output lengths can preserve or enhance accuracy while reducing token cost.

Length-controlled reasoning is the study of how LLMs and large reasoning models regulate the amount of explicit chain-of-thought they generate, typically measured in output or reasoning tokens, so as to balance correctness, latency, token cost, and, in some settings, safety. A central 2025 contribution in this area is "Fast on the Easy, Deep on the Hard: Efficient Reasoning via Powered Length Penalty" (Ling et al., 12 Jun 2025), which argues that reasoning should not be shortened uniformly: simple problems often benefit from concise solutions, whereas difficult problems may require longer deliberation. In that formulation, the goal is selective concision rather than global brevity, and the reported outcomes are shortened outputs on GSM8K and MATH500 while preserving or enhancing accuracy, together with improved accuracy on AIME2024 (Ling et al., 12 Jun 2025).

1. Problem formulation and motivation

The modern length-control literature begins from a shared empirical observation: RL-trained or CoT-p