Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs

By Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang and 4 more. arXiv.org, 2024

536 citations recorded by the source corpus.

The remarkable performance of models like the OpenAI o1 can be attributed to their ability to emulate human-like long-time thinking during inference. These models employ extended chain-of-thought (CoT) processes, exploring multiple strategies to enhance problem-solving capabilities. However, a critical question remains: How to intelligently and efficiently scale computational resources during testing. This paper presents the first comprehensive study on the prevalent issue of overthinking in these models, where excessive computational resources are allocated for simple problems with minimal benefit. We introduce novel efficiency metrics from both outcome and process perspectives to evaluate the rational use of computational resources by o1-like models. Using a self-training paradigm, we propose strategies to mitigate overthinking, streamlining reasoning processes without compromising accuracy. Experimental results show that our approach successfully reduces computational overhead while preserving model performance across a range of testsets with varying difficulty levels, such as GSM8K, MATH500, GPQA, and AIME.

Subfields: Computational Theory and Mathematics

View Paper: https://www.semanticscholar.org/paper/84c149537dee8ef3a7ff54b6bb01e9a4c8eaa17c

Citations in the corpus

• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (2025, unknown): “…between thought traces and final answers [28, 29] as well as efficiency concerns through what researchers term the “overthinking phenomenon” [30, 31, 32, 33], where models produce verbose, redundant outputs, even after finding the solution, creating significant inference computational…”
• From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge (2024, unknown): “In the multimodal domain, Ahn et al. (2024) propose iterative self-retrospective judgment (i-SRT), which employs self-reflection to improve both response generation and preference modeling.”
• Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models (2025, unknown): “For instance, the "overthinking problem" arises when answering a simple question [10] like, " what is the answer of 2 plus 3?”
• SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild (2025, unknown): “As described in §2.1, we use GPT-4o to track shifts in reasoning behaviors on OlympiadBench.”
• L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning (2025, unknown): “Prior works have shown that reasoning models often underthink or overthink, both hurting performance and wasting compute (Wang et al., 2025; Chen et al., 2025).”
• Understanding R1-Zero-Like Training: A Critical Perspective (2025, unknown): “Moreover, on evaluation benchmarks, the length of incorrect responses is substantially reduced by Dr. GRPO compared to the baseline (Plot 4), suggesting that an unbiased optimizer also mitigates overthinking (Chen et al., 2024).”
• Search-o1: Agentic Search-Enhanced Large Reasoning Models (2025, unknown): “While this characteristic enhances logical coherence and interpretability of reasoning, an extended chain of thought may cause overthinking (Chen et al., 2024b) and increased risks of knowledge insufficiency (Wong et al., 2021; Raffel et al., 2020), where any knowledge gap can propagate errors and…”