Saved in:
Bibliographic Details
Main Authors: Wei, Zihao, Pang, Liang, Liu, Jiahao, Shi, Wenjie, Deng, Jingcheng, Xu, Shicheng, Duan, Zenghao, Sun, Fei, Shen, Huawei, Cheng, Xueqi
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2508.17627
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908760926060544
author Wei, Zihao
Pang, Liang
Liu, Jiahao
Shi, Wenjie
Deng, Jingcheng
Xu, Shicheng
Duan, Zenghao
Sun, Fei
Shen, Huawei
Cheng, Xueqi
author_facet Wei, Zihao
Pang, Liang
Liu, Jiahao
Shi, Wenjie
Deng, Jingcheng
Xu, Shicheng
Duan, Zenghao
Sun, Fei
Shen, Huawei
Cheng, Xueqi
contents Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., </think>). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.
format Preprint
id arxiv_https___arxiv_org_abs_2508_17627
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis
Wei, Zihao
Pang, Liang
Liu, Jiahao
Shi, Wenjie
Deng, Jingcheng
Xu, Shicheng
Duan, Zenghao
Sun, Fei
Shen, Huawei
Cheng, Xueqi
Computation and Language
Artificial Intelligence
Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., </think>). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.
title The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2508.17627