OThink-R1: A Dual-Mode Reasoning Framework to Cut Redundant Computation in LLMs

Recent advancements in large reasoning models (LRMs) highlight the inefficiency of static chain-of-thought (CoT) reasoning, which often results in unnecessarily lengthy outputs for simple tasks, thereby increasing computational costs. To address this, the proposed OThink-R1 framework introduces a dual-mode reasoning approach that dynamically adjusts the reasoning process based on task complexity, mimicking human intuition by employing fast, intuitive responses for easy problems and more detailed reasoning for complex ones. This adaptive reasoning strategy aims to overcome the limitations of existing fixed-pattern training and inference methods, which lack flexibility in balancing reasoning depth and efficiency. By