📄 中文摘要
大型推理模型（LRMs）展现了卓越的推理能力，但常常面临过度思考和思考不足的问题。过度思考导致在简单问题上耗费冗余的计算步骤，而思考不足则未能充分探索推理路径。这些问题造成了效率低下和潜在的不准确性，限制了在资源受限环境中的实际应用。现有的缓解过度思考的方法，如抑制反思关键词或调整推理长度，可能会无意中引发思考不足，损害准确性。因此，提出了一种名为ReBalance的无训练框架，旨在实现高效推理与平衡思维。ReBalance利用信心作为连续指标，优化推理过程。
高效推理与平衡思维
出处: Efficient Reasoning with Balanced Thinking
发布: 2026年3月16日
📄 English Summary
Efficient Reasoning with Balanced Thinking
Large Reasoning Models (LRMs) exhibit remarkable reasoning capabilities but often encounter issues of overthinking and underthinking. Overthinking leads to redundant computational steps on simple problems, while underthinking fails to explore sufficient reasoning paths. These challenges result in inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained environments. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. To address these issues, a training-free framework named ReBalance is proposed, which aims to achieve efficient reasoning through balanced thinking. ReBalance leverages confidence as a continuous indicator to optimize the reasoning process.