The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis

TL;DR

This work identifies two intertwined Reasoning Dynamics in LLMs: Thinking-Content Compensation (length-based trade-offs) and Semantic Path Convergence (latent-space stabilization). It defines the instance-specific Reasoning Completion Point (RCP) marking when further thinking becomes redundant, and presents RCPD, an inference-time, rank-based detector that terminates thinking at the RCP to curb overthinking. Across AIME and GPQA benchmarks on Qwen3 and DeepSeek-R1, RCPD achieves substantial token reductions (up to 44%) with minimal accuracy loss, demonstrating a principled, training-free approach to efficient test-time scaling. The findings highlight the importance of adaptive early stopping in reasoning processes and offer a practical tool for robust, efficient deployment of advanced LLM reasoning systems.

Abstract

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.

Figures (8)

• Figure 1: Overview of Reasoning Dynamics and RCP. The top panel summarizes Reasoning Length Dynamics, where content length decreases as thinking length grows under a thinking-content compensation regime until reaching the RCP. The bottom panel summarizes Reasoning Semantic Dynamics, where the latent semantic trajectory transitions from broad exploration to a stable neighborhood with repetitive oscillations, with the onset of convergence aligning with the RCP. The top and bottom panels are defined in §\ref{['sec:length_dynamics']} and §\ref{['sec:semantic_dynamics']}, respectively.
• Figure 2: Two-stage reasoning dynamics separated by RCP: the Pre-RCP Active Reasoning Stage and the Post-RCP Converged Reasoning Stage. The vertical dashed line indicates the RCP boundary. Additional examples are provided in Appendix Figure \ref{['fig:appendix_three_stage_additional']}.
• Figure 3: Semantic trajectory showing the transition from Pre-RCP Active Exploration to Post-RCP Reasoning Convergence. The dashed line indicates the RCP boundary. Additional examples are provided in Appendix Figure \ref{['fig:appendix_semantic_trajectory_additional']}.
• Figure 4: Semantic convergence residual over thinking steps. $\mathcal{D}_{\text{global}}(k)$ declines and then approaches a low plateau. The vertical dashed line indicates the RCP boundary; the inset zooms into the late-step region for readability.
• Figure 5: Top panel: Accuracy stabilizes around answer emergence. Bottom panel: The rank of </think> ($R_k$) drops precipitously at answer emergence; this drop serves as a signature of convergence.