UPAIR: Diagnosing Reasoning States via Uncertainty-Progress Alignment for Selective Intervention

Extracted source text · arXiv 2607.17188

% \documentclass[sigconf]{acmart}
% \documentclass[sigconf,anonymous,review]{acmart}
\documentclass[sigconf,nonacm]{acmart}
\usepackage{amsmath}
\usepackage{mathtools}
\usepackage{amsthm}
\usepackage{caption}
\usepackage{microtype}
\usepackage{graphicx}
\usepackage{subfigure}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{enumitem}
\usepackage{multirow}
\usepackage{multicol}
\usepackage{ulem}
\usepackage{algorithm}
\usepackage{algorithmic}

\usepackage{pifont}
\usepackage{xspace}

\usepackage{scalefnt}
\usepackage{wrapfig}
\usepackage{tabularx} 
\usepackage{colortbl}
\usepackage{xcolor,bm}
\usepackage{bm}
\usepackage{xspace}
\usepackage{pifont}
\usepackage{float}
\usepackage{caption}


\usepackage[table]{xcolor} % 颜色支持 (必须带 table 选项)
\usepackage{multirow}
\usepackage{booktabs}       % professional-quality tables
\usepackage{colortbl}
\usepackage{arydshln}

% 设置 1: 极淡的天蓝色 (PUMA_entropy)
\definecolor{MyColor1}{RGB}{215, 240, 245}
% 设置 2: 淡长春花蓝 (PUMA_latent)
\definecolor{MyColor2}{RGB}{220, 235, 250}
% 设置 3: 淡紫罗兰色 (PUMA)
\definecolor{MyColor3}{RGB}{225, 230, 255}

\newcommand{\cblack}[1]{\textcolor{black}{$_{#1}$}}
\newcommand{\cred}[1]{\textcolor{red}{$_{#1}$}}
\newcommand{\cgreen}[1]{\textcolor{darkgreen}{$_{#1}$}}

% \definecolor{newgreen}{rgb}{0.7, 0.9, 0.7}
% \definecolor{newblue}{rgb}{0.85, 0.85, 0.9}


\newcommand{\frameworkname}{PUMA\xspace}


%%
%% \BibTeX command to typeset BibTeX logo in the docs
\AtBeginDocument{%
  \providecommand\BibTeX{{%
    Bib\TeX}}}

%% Rights management information.  This information is sent to you
%% when you complete the rights form.  These commands have SAMPLE
%% values in them; it is your responsibility as an author to replace
%% the commands and values with those provided to you when you
%% complete the rights form.
% \setcopyright{acmlicensed}
% \copyrightyear{2018}
% \acmYear{2018}
% \acmDOI{XXXXXXX.XXXXXXX}
% %% These commands are for a PROCEEDINGS abstract or paper.
% \acmConference[Conference acronym 'XX]{Make sure to enter the correct
%   conference title from your rights confirmation email}{June 03--05,
%   2018}{Woodstock, NY}
% %%
% %%  Uncomment \acmBooktitle if the title of the proceedings is different
% %%  from ``Proceedings of ...''!
% %%
% %%\acmBooktitle{Woodstock '18: ACM Symposium on Neural Gaze Detection,
% %%  June 03--05, 2018, Woodstock, NY}
% \acmISBN{978-1-4503-XXXX-X/2018/06}


%%
%% Submission ID.
%% Use this when submitting an article to a sponsored event. You'll
%% receive a unique submission ID from the organizers
%% of the event, and this ID should be used as the parameter to this command.
%%\acmSubmissionID{123-A56-BU3}

%%
%% For managing citations, it is recommended to use bibliography
%% files in BibTeX format.
%%
%% You can then either use BibTeX with the ACM-Reference-Format style,
%% or BibLaTeX with the acmnumeric or acmauthoryear sytles, that include
%% support for advanced citation of software artefact from the
%% biblatex-software package, also separately available on CTAN.
%%
%% Look at the sample-*-biblatex.tex files for templates showcasing
%% the biblatex styles.
%%

%%
%% The majority of ACM publications use numbered citations and
%% references.  The command \citestyle{authoryear} switches to the
%% "author year" style.
%%
%% If you are preparing content for an event
%% sponsored by ACM SIGGRAPH, you must use the "author year" style of
%% citations and references.
%% Uncommenting
%% the next command will enable that style.
%%\citestyle{acmauthoryear}


%%
%% end of the preamble, start of the body of the document source.
\begin{document}

%%
%% The "title" command has an optional parameter,
%% allowing the author to define a "short title" to be used in page headers.
\title{Is Your Model Thinking or Just Stagnating? PUMA: Diagnosing Reasoning Pathology via Phase-Momentum Alignment}

% Thinking or Stagnating? Diagnosing Reasoning Pathology in Large Reasoning Models via Phase-Momentum Alignment Reasoning Pathology

%%
%% The "author" command and its associated commands are used to define
%% the authors and their affiliations.
%% Of note is the shared affiliation of the first two authors, and the
%% "authornote" and "authornotemark" commands
%% used to denote shared contribution to the research.
% Cheng Yan: co-first author
\author{Cheng Yan}
\authornote{These authors contributed equally to this work.}
\email{yc_sa22218099@mail.ustc.edu.cn}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

% Guangyang Ye: co-first author
\author{Guangyang Ye}
\authornotemark[1]
\email{yeguangyang@mail.ustc.edu.cn}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

% Wuyang Zhang: corresponding author
\author{Wuyang Zhang}
\authornote{Corresponding authors.}
\email{wuyangz@ustc.edu.cn}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

\author{Fan Xu}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

\author{Zhijun Fan}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

\author{Xiang Xia}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

% Yanyong Zhang: corresponding author
\author{Yanyong Zhang}
\authornotemark[2]
\email{yanyongz@ustc.edu.cn}
\affiliation{%
  \institution{University of Science and Technology of China}
  \city{Hefei}
  \country{China}
}

\renewcommand{\shortauthors}{Yan, Ye, et al.}

%%
%% By default, the full list of authors will be used in the page
%% headers. Often, this list is too long, and will overlap
%% other information printed in the page headers. This command allows
%% the author to define a more concise list
%% of authors' names for this purpose.
\renewcommand{\shortauthors}{Yan et al.}

%%
%% The abstract is a short summary of the work to be presented in the
%% article.
\begin{abstract}
\input{component/1_abstract}
\end{abstract}

%%
%% The code below is generated by the tool at http://dl.acm.org/ccs.cfm.
%% Please copy and paste the code instead of the example below.
%%
% \begin{CCSXML}
% <ccs2012>
%  <concept>
%   <concept_id>00000000.0000000.0000000</concept_id>
%   <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
%   <concept_significance>500</concept_significance>
%  </concept>
%  <concept>
%   <concept_id>00000000.00000000.00000000</concept_id>
%   <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
%   <concept_significance>300</concept_significance>
%  </concept>
%  <concept>
%   <concept_id>00000000.00000000.00000000</concept_id>
%   <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
%   <concept_significance>100</concept_significance>
%  </concept>
%  <concept>
%   <concept_id>00000000.00000000.00000000</concept_id>
%   <concept_desc>Do Not Use This Code, Generate the Correct Terms for Your Paper</concept_desc>
%   <concept_significance>100</concept_significance>
%  </concept>
% </ccs2012>
% \end{CCSXML}

% \ccsdesc[500]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc[300]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc{Do Not Use This Code~Generate the Correct Terms for Your Paper}
% \ccsdesc[100]{Do Not Use This Code~Generate the Correct Terms for Your Paper}
\begin{CCSXML}
<ccs2012>
   <concept>
       <concept_id>10010147.10010178.10010179</concept_id>
       <concept_desc>Computing methodologies~Natural language processing</concept_desc>
       <concept_significance>500</concept_significance>
   </concept>
   <concept>
       <concept_id>10010147.10010257.10010258.10010260.10010229</concept_id>
       <concept_desc>Computing methodologies~Learning latent representations</concept_desc>
       <concept_significance>500</concept_significance>
   </concept>
   <concept>
       <concept_id>10010147.10010178.10010179.10010182</concept_id>
       <concept_desc>Computing methodologies~Natural language generation</concept_desc>
       <concept_significance>500</concept_significance>
   </concept>
</ccs2012>
\end{CCSXML}

\ccsdesc[500]{Computing methodologies~Natural language processing}
\ccsdesc[500]{Computing methodologies~Learning latent representations}
% \ccsdesc[500]{Computing methodologies~Natural language generation}

% \begin{CCSXML}
% <ccs2012>
%    <concept>
%        <concept_id>10010405.10010432.10010437.10010438</concept_id>
%        <concept_desc>Applied computing~Environmental sciences</concept_desc>
%        <concept_significance>500</concept_significance>
%        </concept>
%  </ccs2012>
% \end{CCSXML}

% \ccsdesc[500]{Applied computing~Environmental sciences}

%%
%% Keywords. The author(s) should pick words that accurately describe
%% the work being presented. Separate the keywords with commas.

\keywords{Large Reasoning Models; Test-time Scaling; Efficient Reasoning; Latent Representation Analysis}%% A "teaser" image appears between the author and affiliation
%% information and the body of the document, and typically spans the
%% page.
% \begin{teaserfigure}
%   \includegraphics[width=\textwidth]{sampleteaser}
%   \caption{Seattle Mariners at Spring Training, 2010.}
%   \Description{Enjoying the baseball game from the third-base
%   seats. Ichiro Suzuki preparing to bat.}
%   \label{fig:teaser}
% \end{teaserfigure}

% \received{20 February 2007}
% \received[revised]{12 March 2009}
% \received[accepted]{5 June 2009}

%%
%% This command processes the author and affiliation and title
%% information and builds the first part of the formatted document.
\maketitle

\input{component/2_introduction}
\input{component/3_related_work}
\input{component/4_methods}
\input{component/5_experiment}
\input{component/6_conclusion}

\clearpage
\normalem 
\bibliographystyle{ACM-Reference-Format}
\bibliography{sample-sigconf}


\appendix
\input{component/7_appendix}


\end{document}
\endinput
%%
%% End of file `sample-sigconf.tex'.

Test-time scaling empowers Large Reasoning Models (LRMs) to tackle complex tasks via extensive Chain-of-Thought (CoT). However, this often induces the "overthinking" paradox, where redundant reasoning increases computational overhead without guaranteeing accuracy. Existing test-time efficiency optimization methods primarily fall into two categories: information-theoretic approaches, which are prone to "deceptive convergence" where low uncertainty masks hallucinations, and latent representation analyses, which are often post-hoc, lacking the real-time sensitivity for dynamic reasoning. To bridge this gap, we first posit the \textit{Phase-Momentum Alignment Hypothesis}, asserting that reasoning correctness hinges on the temporal synchronization between geometric momentum and uncertainty resolution. We then theoretically formulate the \textit{Cognitive-Energy Model} to characterize these dynamics through two orthogonal dimensions: \textit{Geometric Cognitive Effort} (quantified by latent velocity and tortuosity) and \textit{Entropic Cognitive Uncertainty}. To operationalize this, we introduce \textbf{\frameworkname} (\textbf{P}hase-\textbf{U}ncertainty \textbf{M}omentum \textbf{A}lignment), a training-free framework employing a tiered diagnostic architecture. By coupling lightweight phase monitoring with event-triggered geometric analysis, \frameworkname effectively distinguishes active exploration from passive stagnation, enabling precise interventions through adaptive truncation or corrective measures. Extensive experiments on LRMs spanning 1.5B to 32B demonstrate that \frameworkname consistently outperforms state-of-the-art baselines across diverse benchmarks, achieving a superior accuracy-efficiency trade-off and robust cross-domain generalization.
\section{Introduction}

%  A Draft Figure
\begin{figure}[t] % [t] 表示优先放在当前栏的顶部，也可以改成 [b] 放底部
  \centering
  % 这里的 width 设置为 \linewidth 或 \columnwidth 以适应单栏宽度
  \includegraphics[width=\linewidth]{figure/intro.pdf}
  \caption{\textbf{Motivation: Active Thinking vs. Stagnation.} 
Visualizing reasoning dynamics of QwQ-32B on MATH-500:
(a) Solvable cases exhibit geometric effort \textit{synchronized with} uncertainty resolution; 
(b) Unsolvable cases manifest persistent high entropy \textit{coupled with} vanishing cognitive momentum.}
\label{fig:motivation}
\vspace{-5pt}
\end{figure}

Large Reasoning Models (LRMs), such as OpenAI o1~\cite{jaech2024openai} and DeepSeek-R1~\cite{guo2025deepseek}, have achieved remarkable success in complex reasoning tasks. To further enhance these capabilities, recent approaches leverage \textit{test-time scaling}, allocating more inference-time computation to generate extensive Chain-of-Thought sequences~\cite{wei2022chain}. 
While this extended generation enables complex reasoning, it also introduces the risk of "overthinking", characterized by redundant and repetitive reasoning that escalate computational costs without guaranteeing accuracy~\cite{sui2025stop, feng2025efficient}. Consequently, developing mechanisms for \textit{efficient reasoning}, particularly through dynamic self-evaluation, has become a key priority to optimize the trade-off between computational cost and reasoning reliability.


Current efforts optimizing test-time computation generally fall into two primary streams: information-theoretic approaches and latent space representation analysis. From an information-theoretic perspective, methods like \textit{Think or Not}~\cite{yong2025think} propose an \textit{Adaptive Think} strategy that dynamically halts generation upon detecting low-entropy states, empirically validating that effective reasoning typically manifests as a progressive transition from high initial uncertainty to low-entropy certainty. However, observations in \textit{Entropy-Lens}~\cite{ali2025entropy} indicate that relying solely on entropy introduces significant ambiguity since high entropy may signify effective exploration rather than ineffective exploration while low entropy can deceptively mask "confident hallucinations," rendering it an unreliable proxy for correctness. In parallel, approaches grounded in latent representation analysis, such as \textit{Tracing the Traces}~\cite{vilas2025tracing}, evaluate reasoning quality by monitoring the temporal evolution of hidden states, demonstrating that the internal representational dynamics inherently encode signals predictive of output correctness. Nevertheless, these methods primarily operate in a post-hoc manner which renders them unsuitable for real-time monitoring. Although \textit{CoRE}~\cite{li2025core} attempts to address this by identifying local cyclical loops, it remains inefficient at detecting subtle error patterns emerging in the early stages of reasoning. In summary, neither paradigm alone suffices: entropy signals are prone to deceptive convergence patterns stemming from the intrinsic ambiguity of entropy, while latent analysis often lacks real-time sensitivity.

% observation
To bridge this gap, we draw inspiration from human cognitive processes, specifically the dynamic interplay between \textit{metacognitive monitoring} and \textit{cognitive effort}. Human reasoning typically evolves through structured phases from initial problem understanding and divergence to iteration and final convergence. Crucially, this mental effort is non-uniform across phases: intense exploration requires high cognitive load, while convergence is characterized by a reduction of effort. Hypothesizing that this phenomenon of phase-dependent thinking effort is also inherent in the generation dynamics of Large Reasoning Models (LRMs), we seek to decode these internal signals. To validate this intuition, we analyze the joint evolution of latent representation dynamics and entropy shifts. 

Our empirical observations, as visualized in Figure~\ref{fig:motivation}, reveal two distinct thinking patterns. In solvable cases (Figure~\ref{fig:motivation}(a)), the model exhibits \textit{synchronized dynamics}: a sharp reduction in entropic uncertainty is strictly coupled with high latent momentum, signaling a distinct transition from exploratory divergence to structural convergence. Conversely, unsolvable or incorrect cases (Figure~\ref{fig:motivation}(b)) manifest a significant \textit{asynchrony}: persistent high entropy coincides with vanishing latent momentum, reflecting a state of "cognitive stagnation" where the model continues to generate tokens despite having ceased constructive semantic processing.
Based on these insights, we formally propose the \textit{Phase-Momentum Alignment Hypothesis}: \textit{The correctness of a reasoning trajectory is closely correlated with the temporal synchronization between its epistemic uncertainty (phase) and its internal semantic Effort (momentum).} 

% The proposal of frameworks and metrics
Motivated by these insights, we formally define the \textit{Cognitive-Energy Model}, which decomposes the reasoning process into two orthogonal dimensions: (1) \textit{Geometric Cognitive Effort}, quantifying the geometric dynamics of the latent trajectory via \textit{Global Thinking Intensity} (tortuosity) and \textit{Local Semantic Gain} (velocity); and (2) \textit{Entropic Cognitive Uncertainty}, leveraging lookahead entropy as a proxy for the solution-space phase. By synthesizing these metrics, we derive the \textit{Net Reasoning Momentum} to operationalize the \textit{Phase-Momentum Alignment Hypothesis} for diagnosis of reasoning pathology.
To implement this model, we introduce \frameworkname, a training-free framework for adaptive cognitive control. It employs a tiered architecture where a lightweight entropy-based Finite State Machine (FSM) serves as a continuous monitor to track phase transitions with minimal computational overhead. The system triggers the Geometric Cognitive Effort calculation strictly upon detecting critical shifts to jointly diagnose the reasoning pattern. Based on the diagnostic outcomes, \frameworkname executes extensible intervention policies, such as effectively truncating generation or applying corrective measures to ensure robust reasoning integrity.

% Main contributions
Our contributions are summarized as follows: 
(1) We formulate the \textit{Cognitive-Energy Model} and the \textit{Phase-Momentum Alignment Hypothesis}. By introducing \textit{Net Reasoning Momentum} derived from \textit{Global Thinking Intensity} and \textit{Local Semantic Gain}, we provide a formal interpretation interpretation of "thinking effort" that effectively distinguishes valid reasoning from cognitive stagnation. 
(2) To operationalize this, we propose \frameworkname, a training-free inference framework that integrates a lightweight entropy-based FSM with event-triggered cognitive diagnosis, enabling adaptive supervision of the reasoning process without parameter updates. 
(3) Extensive experiments validate that \frameworkname achieves a superior accuracy-cost trade-off. Crucially, our framework demonstrates robust generalization across diverse models and reasoning domains, consistently outperforming distinct optimization paradigms.
\section{Related Work}

\subsection{Adaptive and Efficient Reasoning}

To mitigate the ``overthinking'' phenomenon and computational overhead in LRMs, recent research focuses on optimizing reasoning length and eliminating redundant computation ~\cite{sui2025stop, qu2025survey, feng2025efficient, wang2025harnessing}. We categorize these efforts into two primary streams. One major category is Training-based methods, which internalize efficiency via parameter updates, primarily enhancing performance through Supervised Fine-Tuning (SFT) on compressed CoT traces~\cite{luo2025o1, shen2025dast, kang2025c3ot, xia2025tokenskip, ma2025cot}, Reinforcement Learning (RL) with length-aware penalties~\cite{arora2025training, team2025kimi, aggarwal2025l1, yangdemystifying, zhang2025adaptthink}, or implicit reasoning within continuous latent spaces~\cite{shen2025codi, shen2025efficient, su2025token, xu2025softcot}. Alternatively, Training-free methods offer a model-agnostic approach, optimizing inference through prompt constraints for brevity~\cite{lee2025well, chen2024unlocking, ma2025reasoning, han2025token} or dynamic interventions such as early exiting and speculative decoding~\cite{liao2025reward, wu2025more, aytes2025sketch, yong2025think}.

\subsection{Information Theory in Efficient Reasoning.}
Information theory \cite{cover1999elements} provides a valuable theoretical basis for analyzing the behavior of language models, offering rigorous frameworks to characterize generalization, uncertainty and hallucination phenomena \cite{farquhar2024detecting, ren2025revisiting, jeon2025information, chen20252, kossen2024semantic}. Building on these foundations, recent work extends information-theoretic perspectives to Large Reasoning Models (LRMs) to optimize and validate complex reasoning trajectories \cite{gan2025rethinking, chen2024learning}.
\citet{yong2025think} propose an \textit{Adaptive Think} strategy that truncates reasoning when semantic entropy indicates sufficient confidence, though this heuristic remains vulnerable to high-confidence hallucinations where models are confidently incorrect.
Complementary to this, \citet{ali2025entropy} introduce \textit{Entropy-Lens} to analyze layer-wise token distributions, but find that distinguishing between healthy exploratory states and erroneous confusion remains challenging and often yields limited accuracy.
Similarly, \citet{ton2024understanding} quantify stepwise information gain to evaluate reasoning chains, but this verification process necessitates the training of an external supervisor model and relies heavily on well-defined final outcomes, limiting its flexibility in open-ended tasks.

\subsection{Latent Representation Analysis.}
Previous studies have shown that probing the hidden states of Large Language Models (LLMs) reveals informative signals regarding model reliability~\cite{yuksekgonul2023attention}, safety~\cite{turner2023steering}, and generation performance~\cite{wang2024latent}. Building on these insights, recent works have attempted to utilize latent space dynamics to enhance the efficiency of reasoning tasks. Chain-of-Embedding (CoE)~\cite{wang2024latent} estimates response correctness by analyzing the geometric trajectory of hidden states. However, CoE depends on the relative displacement between the input and final output states for normalization, rendering it unsuitable for real-time step-level verification.  ~\cite{li2025core} utilizes geometric analysis of latent trajectories to identify redundant patterns in Chain-of-Thought (CoT) generation.
Latent-Trajectory signals~\cite{vilas2025tracing} characterize the temporal evolution of representations, such as cumulative change, to guide sample selection. However, these metrics rely on coarse-grained segment averaging, which fails to pinpoint fine-grained logical inconsistencies within individual steps.

\section{Preliminaries and Problem Formulation}
\label{sec:preliminaries}

\subsection{Reasoning as a Structured Cognitive Process}
Consider an LRM $\mathcal{M}$ parameterized by $\theta$. We formally model the inference process as the generation of a semantic triplet $\mathcal{T}$, comprising the input query $x$, the final answer $y$, and the intermediate latent reasoning chain $\mathcal{R}$. Aligning with established Chain-of-Thought formalisms~\cite{wei2022chain}, we explicitly structure $\mathcal{R}$ within this triplet by decomposing the continuous token stream into an ordered sequence of discrete \textit{Reasoning Steps}:
\begin{equation}
    \mathcal{T} = \{x, \mathcal{R}, y\}, \quad \text{where } \mathcal{R} = \{s_1, s_2, \ldots, s_T\}.
\end{equation}
Here, $T$ denotes the total number of reasoning steps, which varies dynamically based on the complexity of the query $x$, and each step $s_t$ constitutes a coherent semantic unit, typically delimited by textual separators (e.g., \texttt{\textbackslash n\textbackslash n}).

Beyond this discrete segmentation, we hypothesize that the reasoning generation process aligns with human cognitive dynamics. Specifically, an effective reasoning chain evolves through a specific ordered sequence of cognitive phases $\mathbb{P}$:
\begin{equation}
\label{phases}
    \mathcal{R} \sim \mathcal{P}_{\text{Und}} \to \mathcal{P}_{\text{Div}} \to \mathcal{P}_{\text{Iter}} \to \mathcal{P}_{\text{Conv}} \to \mathcal{P}_{\text{Ref}},
\end{equation}
where the cognitive progression consists of \textit{Problem Understanding} ($\mathcal{P}_{\text{Und}}$), \textit{Cognitive Divergence} ($\mathcal{P}_{\text{Div}}$), optional \textit{Sub-problem Iteration} ($\mathcal{P}_{\text{Iter}}$), \textit{Convergence} ($\mathcal{P}_{\text{Conv}}$), and optional \textit{Reflection} ($\mathcal{P}_{\text{Ref}}$).

\subsection{Problem Formulation}
\label{sec:problem_formulation}

We frame the validation of the reasoning chain $\mathcal{R}$ as an Online Cognitive Diagnosis problem. Specifically, our goal is to monitor the validity of the inference process, distinguishing valid reasoning steps from pathological behaviors such as semantic stagnation or redundancy. To achieve this, we identify the current \textit{Cognitive State} $S_t$ by decoding the temporal synchronization between the model's epistemic uncertainty and its semantic progression.

Guided by the \textit{Phase-Momentum Alignment} hypothesis, we operationalize this diagnosis by deriving two complementary time-series signals from the discrete sequence $\mathcal{R} = \{s_t\}_{t=1}^T$:

\noindent\textbf{Latent Trajectory ($\mathcal{Z}$).} To capture semantic evolution, we project each step $s_t$ into the continuous latent space via a feature extraction function $\psi$:
\begin{equation}
    \mathbf{z}_t = \psi(s_t, \mathcal{M}) \in \mathbb{R}^{L \times d}, \quad \mathcal{Z} = \{\mathbf{z}_1, \dots, \mathbf{z}_t\}.
\end{equation}
This yields a continuous trajectory representing the geometric "momentum" of thought within the semantic manifold.

\noindent\textbf{Uncertainty Profile ($\mathcal{U}$).} Simultaneously, we quantify the generation uncertainty from an information-theoretic perspective to monitor the "phase" of exploration, forming the sequence $\mathcal{U} = \{u_1, \dots, u_t\}$.

\noindent\textbf{Diagnostic Objective.} We define the task as mapping the joint history of these spatiotemporal signals to the discrete state $S_t$. Formally, we seek a diagnostic function $\mathcal{D}$ such that:
\begin{equation}
\label{eq:Diagnostic_Objective}
    S_t = \mathcal{D}(\mathcal{Z}_{\le t}, \mathcal{U}_{\le t}) \in \mathcal{S}_{\text{cognitive}},
\end{equation}
where $\mathcal{S}_{\text{cognitive}}$ encompasses both healthy reasoning states (e.g., active exploration) and pathological modes (e.g., stagnation) that necessitate adaptive intervention.

% ===============================================================================
% 

\section{The Cognitive-Energy Model}
\label{sec:theory}

To formally characterize CoT efficiency, we propose the \textit{Cognitive-Energy Model}, which conceptualizes reasoning as a dynamic trajectory within the transformer's latent space, governed by processing intensity and epistemic uncertainty. We decompose this trajectory into two orthogonal dimensions: \textbf{Geometric Cognitive Effort} (Section~\ref{sec:effort}), quantifying the kinematic momentum of latent evolution, and \textbf{Entropic Cognitive Uncertainty} (Section~\ref{sec:dispersion}), reflecting the "phase" of solution exploration. Synthesizing these quantities, we formalize the \textbf{Phase-Momentum Alignment Hypothesis} (Section~\ref{sec:patterns}) to diagnose the critical temporal synchronization between cognitive expenditure and information resolution.

\subsection{Geometric Cognitive Effort}
\label{sec:effort}

To quantify the semantic dynamics during reasoning, we analyze the geometric evolution of hidden states within the final $m$ layers, as deeper layers encode the most abstract semantics. Let $C_t = \{x, s_1, \dots, s_t\}$ denote the full context sequence up to step $t$. We introduce two complementary metrics to characterize the latent geometry, based on which we subsequently formulate a unified indicator of cognitive effort.

\noindent\textbf{Global Thinking Intensity ($\mathcal{I}_t$).}
First, we define $\mathcal{I}_t$ as the ratio of cumulative layer-wise transformation to the net semantic displacement, serving as a proxy for the geometric complexity of the reasoning process:
\begin{equation}
    \mathcal{I}_t = \frac{\sum_{l=L-m+1}^{L-1} \delta(\mathbf{h}_t^{l}, \mathbf{h}_t^{l+1})}{\delta(\mathbf{h}_t^{L-m+1}, \mathbf{h}_t^{L})},
\end{equation}
where $\mathbf{h}_t^l \in \mathbb{R}^d$ is defined as the mean-pooled representation of the full context sequence $C_t$ at the $l$-th transformer layer. The function $\delta(\cdot, \cdot)$ denotes a geometric distance metric. We instantiate this metric in two forms to capture distinct properties of the latent space—Magnitude (Euclidean) and Angle (Cosine):
\begin{equation}
    \label{distance}
    \delta_{\text{mag}}(\mathbf{a}, \mathbf{b}) = \|\mathbf{a} - \mathbf{b}\|_2, \quad 
    \delta_{\text{ang}}(\mathbf{a}, \mathbf{b}) = \arccos\left(\frac{\mathbf{a}^\top \mathbf{b}}{\|\mathbf{a}\|_2 \|\mathbf{b}\|_2}\right).
\end{equation}
While $\delta_{\text{mag}}$ captures the absolute scale of semantic updates, we explicitly utilize the Angular Distance ($\delta_{\text{ang}}$) in our primary analysis to ensure robustness against scale variations inherent in high-dimensional latent spaces.

Physically, the ratio $\mathcal{I}_t$ quantifies the \textit{tortuosity} of the signal propagation path traversed on the semantic manifold. A high $\mathcal{I}_t$ indicates that the model engages in extensive internal processing, characterized by a circuitous trajectory to refine latent representations rather than a direct mapping.

\noindent\textbf{Local Semantic Gain ($\mathcal{G}_t$).}
while $\mathcal{I}_t$ captures intra-step processing depth, we introduce $\mathcal{G}_t$ to measure the inter-step semantic transition, serving as the \textit{semantic velocity} of the trajectory:
\begin{equation}
    \mathcal{G}_t = \delta(\mathbf{z}_t, \mathbf{z}_{t-1}), \quad \text{where } \mathbf{z}_t = \frac{1}{m} \sum_{l=L-m+1}^{L} \mathbf{h}_t^l.
\end{equation}
Here, $\mathbf{z}_t$ represents the \textit{Semantic Centroid} of the processing unit. A diminishing $\mathcal{G}_t \to 0$ indicates that the reasoning chain has ceased to generate new semantic content.

\noindent\textbf{Unified Metric: Net Reasoning Momentum ($\Phi_t$).}
Valid reasoning requires not just velocity ($\mathcal{G}_t$) but also sustained cognitive intensity. we formulate this comprehensive metric as:
\begin{equation}
\label{eq: Net Reasoning Momentum}
    \Phi_t = \underbrace{\mathcal{G}_t}_{\text{Semantic Velocity}} + \lambda \cdot \underbrace{|\mathcal{I}_t - 2\mathcal{I}_{t-1} + \mathcal{I}_{t-2}|}_{\text{Intensity Fluctuation}},
\end{equation}
where the second term approximates the \textit{Intensity Curvature} ($\kappa_t \approx |\Delta^2 \mathcal{I}_t|$), representing the acceleration of cognitive load. Crucially, as both terms rely on angular metrics, $\Phi_t$ serves as a robust consistent energy function.

\textit{Convergence Mechanics:} When the model reaches a solution fixed point $y^*$, the latent representations stabilize such that $\mathbf{h}_t^l \approx \mathbf{h}_{t-1}^l$. Geometrically, this creates a stationary state on the manifold: both the semantic displacement vanishes ($\mathcal{G}_t \to 0$) and the internal processing path stabilizes ($\Delta^2 \mathcal{I}_t \to 0$), driving $\Phi_t \to 0$. This mathematically characterizes the transition from active reasoning to \textit{cognitive stagnation} $\mathcal{P}_{\text{Conv}}$.


\subsection{Entropic Cognitive Uncertainty}
\label{sec:dispersion}

Complementary to geometric analysis, we quantify the epistemic uncertainty of the reasoning process. We define the Proxy-based Lookahead Entropy ($\mathcal{U}_t$) via a lightweight verifier $\mathcal{M}_{\phi}$:
\begin{equation}
    \mathcal{U}_t = -\sum_{y \in \mathcal{Y}} P_{\phi}(y | C_t) \log_2 P_{\phi}(y | C_t),
\end{equation}
where, the set of candidate final answers $\mathcal{Y} = \{y_1, \dots, y_k\}$ is generated by prompting a lightweight proxy model $\mathcal{M}_{\phi}$ (with parameters $|\phi| \ll |\theta|$) conditioned on the current context $C_t$. The term $P_{\phi}(y | C_t)$ denotes the normalized generation probability of candidate $y$.

This decoupled design significantly reduces overhead and mitigates \textit{Self-Reinforcement Bias}. Crucially, by functioning as a \textit{conditional verifier} rather than an independent generator, the proxy leverages the pre-generated context to effectively discriminate between valid and erroneous reasoning despite the capacity gap. This ensures diagnostic precision, establishing $\mathcal{U}_t$ as a reliable indicator of the model's evolutionary state.

\noindent\textbf{Theoretical Phases and Probability Landscapes.}
To rigorously diagnose the reasoning phases, we first establish the boundary conditions of the uncertainty metric. The theoretical upper bound $\mathcal{U}_{max}$ is defined by the maximum entropy (uniform) distribution over the discrete output space $\mathcal{Y}$:
\begin{equation}
    \mathcal{U}_{max} \coloneqq \max_{P} \mathcal{U}(P) = -\sum_{y \in \mathcal{Y}} \frac{1}{|\mathcal{Y}|} \log_2 \frac{1}{|\mathcal{Y}|} = \log_2 |\mathcal{Y}|.
\end{equation}
We then define the \textit{Initial Uncertainty} of the Understanding Phase ($\mathcal{P}_{\text{Und}}$) as $\mathcal{U}_{init} = \frac{1}{k} \sum_{t=1}^{k} \mathcal{U}_t$ (where $k \le 5$), capturing the model's \textit{intuitive prior} before deep reasoning branches out.

We map the step $s_t$ to the theoretical phases defined in Eq.~\ref{phases} by monitoring the persistence of uncertainty patterns over a temporal window $w$. Formally, the phase assignment is determined as:

\begin{equation}
\label{eq:reasoning phase}
    s_t \in
    \begin{cases} 
    \mathcal{P}_{\text{Conv}} & \text{if } \forall \tau \in [t-w, t], \ \mathcal{U}_\tau < \alpha \cdot \mathcal{U}_{max} \\
    \mathcal{P}_{\text{Div}} & \text{if } \forall \tau \in [t-w, t], \ \mathcal{U}_\tau > \mathcal{U}_{init} 
    \end{cases},
\end{equation}
where $w$ denotes the window size for stability verification, and $\alpha \in (0,1)$ is a sensitivity coefficient.

From a time-series perspective, the evolution of $\mathcal{U}_t$ reflects the dynamic reshaping of the probability landscape over $\mathcal{Y}$. Persistence in the $\mathcal{P}_{\text{Div}}$ phase signifies a \textit{flattened distribution} where probability mass is dispersed across multiple hypotheses. Conversely, the $\mathcal{P}_{\text{Conv}}$ phase implies that the distribution asymptotically approaches a \textit{Dirac delta function} centered on a consensus candidate $y^*$. This confirms the effective collapse of the solution space, signaling the final elimination of semantic ambiguity.


\begin{figure*}[t] % [t] 表示强制置顶，双栏图片通常只能置顶或置底
  \centering
  \includegraphics[width=0.95\textwidth]{figure/framework_draft1.pdf} 
  
  \vspace{-0.2cm} % 可选：如果觉得图片和标题间距太大，可以用负值微调
  
 \caption{\textbf{Overview of the \frameworkname{} Framework.} 
    (1) \textbf{Real-time Monitoring} via an entropy-based FSM to track phase transitions; 
    (2) \textbf{Event-triggered Diagnosis} utilizing Geometric Cognitive Effort to verify Phase-Momentum Alignment; 
    and (3) \textbf{Adaptive Intervention} executing dynamic truncation or remedial actions to rectify cognitive pathology.}
\label{fig:framework}
  
  \vspace{-0.2cm} % 可选：减少图注下方的空白，节省版面
\end{figure*}


\subsection{Cognitive Pattern Recognition}
\label{sec:patterns}

With geometric momentum ($\Phi_t$) and entropic uncertainty ($\mathcal{U}_t$) formally established, we synthesize these metrics to characterize the \textit{Phase-Momentum Alignment Hypothesis}, focusing specifically on the temporal synchronization between effort relaxation and uncertainty collapse. 
To operationalize the cognitive state $S_t$ (defined in Eq.~\ref{eq:Diagnostic_Objective}), we categorize the joint dynamics of momentum and reasoning phases into four distinct thinking patterns:

\begin{equation}
\label{eq:thinking_pattern}
    S_t = 
    \begin{cases} 
     \mathcal{S}_{\text{Stagnant}} & \text{if } s_t \in \mathcal{P}_{\text{Div}} \land \Phi_t < \epsilon \\
     \mathcal{S}_{\text{Healthy}} & \text{if } s_t \in \mathcal{P}_{\text{Conv}} \land \Phi_t < \epsilon \land (\nexists \tau < t, s_\tau \in \mathcal{S}_{\text{Stagnant}}) \\
    \mathcal{S}_{\text{Redundant}} & \text{if } s_t \in \mathcal{P}_{\text{Div}} \land (\exists \tau < t, S_\tau = \mathcal{S}_{\text{Healthy}})
    \end{cases},
\end{equation}
where $\epsilon$ is the relaxation threshold. To facilitate this analysis, we introduce two critical timestamps: $T_{\text{trans}}$, denoting the transition point from Divergence ($\mathcal{P}_{\text{Div}}$) to Convergence ($\mathcal{P}_{\text{Conv}}$); and $T_{\text{relax}}$, denoting the moment when Net Reasoning Momentum drops below a relaxation threshold $\epsilon$ (i.e., $\Phi_t < \epsilon$). 

\noindent\textit{\textbf{Pattern I: Synchronized Alignment ($\mathcal{S}_{\text{Healthy}}$).}}
In this ideal state, the model maintains high cognitive momentum throughout the exploratory phase, and the relaxation of effort coincides with or slightly follows the resolution of uncertainty ($T_{\text{relax}} \ge T_{\text{trans}}$).
Physically, this indicates that the model is actively traversing the semantic manifold to resolve ambiguity. The eventual drop in $\Phi_t$ is a natural consequence of successfully locating the solution fixed point.

\noindent\textit{\textbf{Pattern II: Energy-Entropy Misalignment ($\mathcal{S}_{\text{Stagnant}}$).}}
Here, the cognitive momentum dissipates \textit{before} the uncertainty is resolved ($T_{\text{relax}} \ll T_{\text{trans}}$).
Despite remaining in a high-entropy state, the model ceases to update its semantic representation. This pattern characterizes a "giving up" behavior, where the model generates empty tokens without performing actual geometric work, inevitably drifting towards a hallucinated or incorrect conclusion.

\noindent\textit{\textbf{Pattern III:  Post-Convergence Repetition($\mathcal{S}_{\text{Redundant}}$).}}
This pattern is defined as the superfluous continuation of generation after a successful Pattern I sequence has been established.
This pattern signifies a failure of the internal stop mechanism, where the model manifests spurious uncertainty despite having solved the problem, resulting in repetitive verification loops.

\section{\frameworkname: Adaptive Cognitive Control}
\label{sec:method}

To operationalize the \textit{Cognitive-Energy Model} defined in Section~\ref{sec:theory}, we propose a training-free framework, as shown in Figure~\ref{fig:framework}), integrating lightweight monitoring, event-triggered cognitive diagnosis, and extensible dynamic intervention. The system orchestrates the inference process through three stages:
(1) A lightweight entropy-based \textit{Finite State Machine (FSM)} monitors phase transitions in real-time;
(2) Upon detecting critical phase shifts, the system triggers the \textit{Geometric Cognitive Effort} calculation to jointly diagnose the reasoning pattern;
(3) Based on the diagnosis, the system adaptively executes corresponding intervention measures, effectively truncating generation or applying corrective measures.

\subsection{Entropy-based FSM Monitoring}
To capture valid reasoning signals, we initiate monitoring strictly after the \textit{Problem Understanding} phase $\mathcal{P}_{\text{Und}}$ (as defined in Section~\ref{sec:dispersion}). 
We formalize the FSM over a ternary state space $\mathcal{S} = \{\mathcal{P}_{\text{Conv}}, \mathcal{P}_{\text{Div}}, \mathcal{P}_{\text{Undet}}\}$. 
 where any interval failing to satisfy the strict thresholds for either convergence or divergence is categorized as the indeterminate state $\mathcal{P}_{\text{Undet}}$. To ensure robust detection, the system tracks phase stability using a sliding window of size $w_{\text{fixed}}$ according to Eq.~\ref{eq:reasoning phase}.

To balance diagnostic precision with runtime efficiency, we employ an event-driven protocol. The Diagnostic Trigger is activated strictly upon the detection of critical state transitions: 
\ding{182} \textbf{Convergence Alert:} The entropy consistently drops ($s_t \to \mathcal{P}_{\text{Conv}}$), signaling a potential solution candidate. 
\ding{183} \textbf{Divergence Persistence Alert:} The entropy persists in high-uncertainty ($s_t \in \mathcal{P}_{\text{Div}}$), suggesting potential infinite looping or cognitive stagnation.
Once triggered, the system computes the Net Reasoning Momentum $\Phi_t$ (Eq.~\ref{eq: Net Reasoning Momentum}) and synthesizes it with the current phase state to identify the underlying cognitive pattern as defined in Eq.~\ref{eq:thinking_pattern}.

\subsection{Dynamic Intervention Mechanism}
\label{sec:intervention}
Based on the cognitive state $S_t$ identified by the diagnostic module, we define a mapping function $\pi: \mathcal{S}_{\text{cognitive}} \to \mathcal{A}$ that translates reasoning patterns into actionable control policies:
\begin{equation}
\pi(S_t) = 
\begin{cases} 
\text{TRUNCATE}_{\text{success}}, & \text{if } S_t = \mathcal{S}_{\text{Healthy}} \\
\text{TRUNCATE}_{\text{fail}}, & \text{if } S_t = \mathcal{S}_{\text{Stagnant}} \\
\text{CONTINUE}, & \text{otherwise}
\end{cases}
\label{eq:intervention_policy}
\end{equation}
This mechanism unifies efficiency optimization and error mitigation through the following targeted protocols:

\noindent\textbf{Adaptive Success Truncation (Targeting $\mathcal{S}_{\text{Healthy}}$).}
Standard stopping criteria often cut off reasoning prematurely or fail to detect redundancy. Our protocol ensures the completeness of the "Thought Chain" by acting strictly upon the detection of \textit{Synchronized Alignment}. By executing $\text{TRUNCATE}_{\text{success}}$ at the exact moment the model satisfies healthy convergence (low entropy coupled with relaxed momentum), we guarantee that the solution fixed point $y^*$ has been reached. Crucially, this preemptive intervention implicitly manages Post-Convergence Repetition ($\mathcal{S}_{\text{Redundant}}$) by eliminating the temporal window required for the model to drift into a redundant state, thereby maximizing efficiency without a separate suppression mechanism.

\noindent\textbf{Cognitive Circuit Breaking \& Restart (Targeting $\mathcal{S}_{\text{Stagnant}}$).}
To prevent hallucinations induced by reasoning failure, we employ a "Circuit Breaker" mechanism. When the diagnosis identifies \textit{Energy-Entropy Misalignment} ($S_t = \mathcal{S}_{\text{Stagnant}}$), the system executes $\text{TRUNCATE}_{\text{fail}}$. This action halts the generation of vacuous tokens where the model has "given up" (vanishing momentum) despite unresolved ambiguity. This dynamic intervention specifically targets the pathology of "reasoning collapse," ensuring computational resources are not wasted on logically invalid content and preventing error propagation at an early stage.

Furthermore, \textit{Circuit Breaker} activation signals critical cognitive barriers, likely stemming from missing domain information or intrinsic reasoning insufficiency. We preliminarily explore a \textit{Retrieval-Augmented Restart} mechanism on mathematical datasets, where external context (e.g., theorems) is injected to reset the reasoning trajectory and rectify potential hallucinations. Given the high extensibility of our intervention mechanism, we reserve comprehensive evaluation of such advanced strategies for future work.
\section{Experiments}
\label{sec:experiments}

\begin{table*}[t]
% \vspace{2pt}
\centering
\small
\setlength{\tabcolsep}{1.2mm}
\caption{\textbf{Main Results on Math, Science, and General Reasoning Benchmarks.} We report Accuracy ($\uparrow$) and Average Token Cost ($\downarrow$) across DeepSeek-R1-Distill (7B/14B) and QwQ-32B models. We compare \frameworkname{} against baselines from three distinct paradigms (Information-Theoretic, Latent-based, and Training-based) alongside our two ablated variants.}
\label{tab:main_results}
\resizebox{0.9\textwidth}{!}{
\begin{tabular}{@{} l cccccc cc cc @{}}
\toprule
% === 第一层表头 ===
\multirow{3}{*}{\textbf{Methods}} 
& \multicolumn{6}{c}{\textbf{MATH}} 
& \multicolumn{2}{c}{\textbf{SCIENCE}} 
& \multicolumn{2}{c}{\textbf{GENERAL}} \\ 
\cmidrule(lr{1em}){2-7} \cmidrule(lr{1em}){8-9} \cmidrule(lr{0em}){10-11}

% === 第二层表头 ===
& \multicolumn{2}{c}{\textbf{GSM8K}} 
& \multicolumn{2}{c}{\textbf{MATH-500}} 
& \multicolumn{2}{c}{\textbf{AIME2025}}
& \multicolumn{2}{c}{\textbf{GPQA-D}}
& \multicolumn{2}{c}{\textbf{MMLU-Pro}} \\ 
% \cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11}

% === 第三层表头 ===
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ \\ 
\midrule

% ==================== 7B Models ====================
\multicolumn{11}{c}{\textit{DeepSeek-R1-Distill-Qwen-7B}} \\ 
% \cmidrule{1-11}
\hdashline[1pt/2pt] \addlinespace[2pt]
Vanilla    & 80.21 & 507.16 & 84.80 & 2345.30 & 36.67 & 7243.87 & 25.76 & 5336.08 & 55.71 & 2157.76 \\
TracingtheTraces & 80.60 & 1406.21  & 84.20 & 6126.10 & 33.33 & 11433.15 & 29.61 & 9136.77 & 57.30 & 5668.90 \\
ThinkorNot  & 79.15 & 410.31  & 81.80 & 1516.59  & 30.00 & 6497.57  & 25.76 & 3882.31 & 48.00 & 1638.92 \\
AdaptThink        & \textbf{87.18} & 392.94  & 86.00 & \textbf{1484.91} & 33.33 & \textbf{6445.47}  & 37.88 & 4206.66 & 60.00 & 1773.70 \\
\midrule
% \rowcolor{MyColor1} 
\textbf{$\text{\frameworkname}_{\text{entropy}}$} 
& 79.83 & 239.88 
& 82.40 & 1708.52 
& 30.00 & 6681.00 
& 25.76 & 3957.54 & 51.41 & 1758.77 \\
% \rowcolor{MyColor2} 
\textbf{$\text{\frameworkname}_{\text{latent}}$} 
& 78.99 & 241.44 
& 81.80 & 1712.98 
& 36.67 & 6801.60 
& 38.39 &4463.95 & 58.57 & \textbf{1047.59} \\
\rowcolor{MyColor2} \textbf{$\text{\frameworkname}$} 
& 84.08 & \textbf{235.48} 
& \textbf{87.40} & 1847.71 
& \textbf{36.67} & 6620.17 
& \textbf{39.40} & \textbf{2861.08} & \textbf{62.86} & 1552.76 \\
\rowcolor{MyColor3} 
\textbf{$\Delta$ vs. Vanilla}
& +4.82\% & -53.57\%
& +3.07\% & -21.22\% 
& +0.0\% & -8.61\% 
& +52.95\% & -46.38\% & +12.83\% & -28.04\% \\

\midrule
% ==================== 14B Models ====================
\multicolumn{11}{c}{\textit{DeepSeek-R1-Distill-Qwen-14B}} \\ 
% \cmidrule{1-11}
\hdashline[1pt/2pt] \addlinespace[2pt]
Vanilla    & 88.63 & 710.49  & 86.60 & 2397.04 & 43.33 & 8101.67 & 53.53 & 3650.61 & 74.29 & 1728.84 \\
TracingtheTraces & 88.86 & 2097.41  & 84.80 & 6433.80 & 36.67 & 11966.90 & 54.02 & 7418.03 & 74.16 & 4722.90 \\
ThinkorNot  & 87.18 & 525.30  & 83.40 & 1781.60  & 33.33 & 7076.65 & 48.89 & 2822.94 & 68.41 & 1313.26 \\

\midrule
% \rowcolor{MyColor1} 
\textbf{$\text{\frameworkname}_{\text{entropy}}$} 
& 88.02 & 395.76 
& 84.00 & \textbf{1720.23} 
& 33.33 & \textbf{7073.80} 
& 53.03 & 3038.69 & 72.86 & 1360.91 \\
% \rowcolor{MyColor2} 
\textbf{$\text{\frameworkname}_{\text{latent}}$} 
& 88.40 & 423.00 
& 82.80 & 1760.22 
& 40.00 & 7511.53 
& 51.52 & \textbf{2805.70} & 72.86 & \textbf{995.56} \\
\rowcolor{MyColor2} \textbf{$\text{\frameworkname}$}
& \textbf{90.14} & \textbf{394.99} 
& \textbf{89.60} & 1822.31 
& \textbf{40.00} & 7349.97 
& \textbf{56.06} & 2885.09 & \textbf{75.71} & 1259.39 \\
\rowcolor{MyColor3} 
\textbf{$\Delta$ vs. Vanilla}
& +1.70\% & -44.41\% 
& +3.46\% & -23.98\% 
& -7.69\% & -9.28\% 
& +4.73\% & -20.97\% & +1.91\% & -27.15\% \\

\midrule
% ==================== 32B Models ====================
\multicolumn{11}{c}{\textit{Qwen-QWQ-32B}} \\ 
% \cmidrule{1-11}
\hdashline[1pt/2pt] \addlinespace[2pt]
Vanilla    & \textbf{95.98} & 1240.32  & 87.80 & 3590.62 & 43.33 & 8749.20 & 51.51 & 6609.05 & 80.00 & 3120.86 \\
TracingtheTraces & 95.82 & 3406.74  & 86.66 & 7331.60 & 36.67 & 12347.89 & 53.61 & 9873.06 & 76.44 & 7081.94 \\
ThinkorNot  & 93.63 & 819.11  & 84.40 & 2649.45  & 36.67 & 7648.86 & 49.90 & 5127.31 & 73.83 & 2343.89 \\
\midrule
% \rowcolor{MyColor1} 
\textbf{$\text{\frameworkname}_{\text{entropy}}$} 
& 93.86 & 814.75 
& 85.40 & 2755.13 
& 36.67 & \textbf{7596.27} 
& 50.00 & \textbf{5032.51} & 74.16 & 2492.94 \\
% \rowcolor{MyColor2} 
\textbf{$\text{\frameworkname}_{\text{latent}}$} 
& 95.15 & 931.80 
& 84.20 & 2655.70 
& 40.00 & 7986.43 
& 59.09 & 5388.70 & 78.57 & \textbf{1977.53} \\
\rowcolor{MyColor2} \textbf{$\text{\frameworkname}$} 
& 95.91 & \textbf{812.48} 
& \textbf{91.40} & \textbf{2616.89} 
& \textbf{46.67} & 8083.70 
& \textbf{63.13} & 5414.01 & \textbf{81.41} & 2301.93 \\
\rowcolor{MyColor3} 
\textbf{$\Delta$ vs. Vanilla}
& -0.49\% & -34.49\% 
& +4.1\% & -27.12\% 
& +7.08\% & -7.6\% 
& +22.56 & -18.08\% & +1.76\% & -26.24\% \\

\bottomrule
\end{tabular}
}
\vspace{1mm}
\end{table*}

\subsection{Experimental Setup}
\label{sec:setup}

\noindent\textbf{Models and Datasets.} We evaluate \frameworkname on four Large Reasoning Models (LRMs) spanning parameter scales from 1.5B to 32B. Specifically, we employ the \textbf{DeepSeek-R1-Distill}~\cite{guo2025deepseek} series (1.5B, 7B, 14B) and \textbf{QwQ-32B}~\cite{team2025qwq}. To ensure a comprehensive evaluation, we select five benchmarks across three domains: \textbf{Mathematical Reasoning} (GSM8k~\cite{cobbe2021training}, MATH-500~\cite{hendrycks2021measuring}, and AIME 2025), \textbf{Scientific Reasoning} (GPQA-Diamond~\cite{rein2024gpqa}), and \textbf{General Complex Reasoning} (MMLU-Pro~\cite{wang2024mmlu}).

\vspace{2mm}
\noindent\textbf{Baselines.} We benchmark against representative efficiency optimization methods across three distinct paradigms:
\begin{itemize}[leftmargin=*]
    \item \textbf{Information-Theoretic Methods:} \textbf{Think or Not}~\cite{yong2025think} adopts an entropy-based halting mechanism, utilizing step-level entropy shifts and low-uncertainty thresholds to adaptively truncate generation when the model exhibits sufficient confidence.
    
    \item \textbf{Latent Representation Analysis:} \textbf{Tracing the Traces}(LT)~\cite{vilas2025tracing} leverages \textit{Latent-Trajectory Signals} (e.g., cumulative trajectory changes) to enable the early selection of high-quality traces, estimating the correctness of reasoning chains in a label-free manner.
    
    \item \textbf{Training-based Control:} \textbf{AdaptThink}~\cite{zhang2025adaptthink} employs a multi-stage Reinforcement Learning (RL) framework to train a policy that dynamically switches between elaborate reasoning and direct answering. Due to the substantial computational overhead of RL training, we restrict the comparison of this baseline to the 1.5B and 7B parameter scales.
\end{itemize}

% \vspace{2mm}
\noindent\textbf{Variants for Hypothesis Verification.} To explicitly verify the Phase-Momentum Alignment and disentangle the contributions of information-theoretic and geometric signals, we construct two ablated variants of our framework:
\begin{itemize}[leftmargin=*]
    \item $\text{\frameworkname}_{\text{entropy}}$: This variant relies solely on the Information-Theo\-retic dimension, executing truncation strictly based on the low-entropy convergence signal derived from our refined Lookahead Entropy (Section 4.2), ignoring geometric momentum.
    \item $\text{\frameworkname}_{\text{latent}}$: This variant operates exclusively on the geometric analysis of hidden states, triggering truncation solely when the Net Reasoning Momentum drops below the relaxation threshold (i.e., $\Phi_t < \epsilon$), without considering phase alignment.
\end{itemize}

% \vspace{2mm}
\noindent\textbf{Implementation Details.}
We perform all experiments on NVIDIA A100 (80GB) GPUs. For \textit{Entropic Cognitive Dispersion}, we employ DeepSeek-R1-Distill-Qwen-1.5B as the default proxy model $M_{\phi}$. To balance monitoring efficiency and accuracy, we define the reasoning step granularity as every three newline separators (`\verb|\n\n|'). Regarding hyperparameters, we set the phase sensitivity $\alpha=0.05$ and momentum threshold $\epsilon=0.1$ with a sliding window $w=3$; sensitivity analyses for these parameters and the proxy model size are provided in Section~\ref{sec:sensitivity}. For \textit{Geometric Cognitive Effort}, we analyze the hidden states of the final 25\% of layers using Angular Distance ($\delta_{ang}$). All reported results are averaged over three independent runs to mitigate generation randomness.

\vspace{-3pt}
\subsection{Main Results: Efficiency-Accuracy Trade-off}
\label{sec:main_results}
% 对应：Math & Other benchmarks
% 展示 PUMA 在不同尺寸模型上，既能保持高 Acc，又能大幅减少 Token。

% \noindent\textbf{Comparative Efficiency Analysis.} 
% We present the comprehensive performance of \frameworkname{} across mathematical (GSM8k, MATH-500) and general reasoning benchmarks (MMLU-Pro). Table~\ref{tab:main_results} reports the Accuracy and \textit{Token Saving Ratio (TSR)} compared to the baselines.

% % 1. PUMA 在所有模型尺寸上都优于 Entropy-based 方法（因为它们容易被 hallucination 欺骗）。
% % 2. PUMA 优于 Latent-based 方法（如 Tracing the Traces），证明了 step-level 监控优于 coarse-grained segment 分析。
% % 3. 说明我们就算单一指标也优于等于对应baseline
% % 4. 强调在非数学任务（General Reasoning）上的泛化性。

% \noindent\textbf{Pareto Frontier Analysis.} 
% To rigorous evaluate the trade-off, Figure~\ref{fig:pareto} illustrates the accuracy-efficiency Pareto frontier. \frameworkname{} consistently pushes the frontier boundary, achieving higher accuracy for the same computational budget compared to Training-based (AdaptThink) and Training-free baselines.

\noindent\textbf{Overall Efficiency Analysis.} 
We present the comprehensive performance of \frameworkname{} across mathematical (GSM8k, MATH-500, AIME 2025), scientific (GPQA-Diamond), and general reasoning benchmarks (MMLU-Pro). As shown in Table~\ref{tab:main_results} and the accuracy-efficiency Pareto frontier in Figure~\ref{fig:main_results}(b), \frameworkname{} achieves a superior balance between reasoning accuracy and computational cost. It consistently outperforms baselines across varying parameter scales from 7B to 32B (with 1.5B results detailed in the Appendix). To understand the sources of these improvements, we provide a detailed comparison against three representative baseline paradigms below.

\vspace{2mm}
\noindent\textbf{Entropy Ambiguity and Calculation Bias (vs. Think or Not).} 
While \textit{Think or Not}~\cite{yong2025think} reduces token consumption via its adaptive halting mechanism, it yields lower accuracy compared to our ablated variant $\text{\frameworkname}_{\text{entropy}}$. For instance, on the GPQA-D benchmark with the DeepSeek-R1-Distill-14B model, \textit{Think or Not} achieves an accuracy of 53.53\% $\to$ 48.89\%, whereas $\text{\frameworkname}_{\text{entropy}}$ reaches \textbf{53.03\%}. This performance difference stems from two structural limitations in the baseline:
(1) \textbf{Length Bias:} It relies on beam search and sequence-based self-calculated entropy, inducing a bias toward shorter answers to minimize uncertainty.
(2) \textbf{Coarse Granularity:} Its segmentation into fixed $n$-token intervals disrupts semantic continuity, causing the mechanism to miss the optimal halting timing.

\begin{figure*}[t]
  \centering
  \includegraphics[width=\textwidth]{figure/exp.pdf} % Replace with your actual filename
  \caption{\textbf{Comprehensive Performance Analysis of \frameworkname{}.} 
    (a) \textbf{Reasoning Quadrant Distribution} of QwQ-32B on MATH-500, defined in \S\ref{sec:ablation}.
    (b) \textbf{Efficiency-Accuracy Pareto Frontier} comparing \frameworkname{} with baselines across various on MATH-500 dataset.
    (c) \textbf{Parameter Sensitivity} of phase ($\alpha$) and momentum ($\epsilon$) thresholds regarding average accuracy and token cost.}
  \label{fig:main_results}
  \vspace{-5pt} % Optional: Adjust spacing to save vertical space
\end{figure*}

Furthermore, \frameworkname{} demonstrates higher overall efficiency compared to \textit{Think or Not}. On the MATH-500 benchmark using the 32B model, \frameworkname{} attains 91.40\% accuracy compared to 84.40\% for the baseline (and 87.80\% for Vanilla). This improvement arises because \frameworkname{} addresses the \textit{"deceptive convergence"} of single-modal entropy. By cross-referencing with \textit{Net Reasoning Momentum}, PUMA identifies high-momentum states as "Active Exploration," preventing the premature truncation of reasoning paths.

\vspace{2mm}
\noindent\textbf{Representation Granularity and Dynamic Diagnosis (vs. Tracing the Traces).}
Compared to our latent-only variant ($\text{\frameworkname}_{\text{latent}}$), \textit{Tracing the Traces} (LT)~\cite{vilas2025tracing} may achieve marginal accuracy gains in specific configurations (e.g., 88.86\% vs. 88.40\% on GSM8k with the 14B model). However, this comes at an exorbitant computational cost. This inefficiency stems from LT's reliance on a \textit{generative selection paradigm}, which necessitates generating multiple reasoning paths of fixed length (set to $k=3$, 2000 tokens each) before selection. Consequently, on MATH-500 with the 7B model, LT consumes 6126.10 tokens on average—over 3.5$\times$ higher than $\text{\frameworkname}_{\text{latent}}$ (1712.98 tokens). $\text{\frameworkname}_{\text{latent}}$ avoids this redundancy by performing real-time diagnosis on a single path.

Furthermore, regarding reasoning robustness, \frameworkname{} exhibits superior adaptability on challenging benchmarks. As shown in Table~\ref{tab:main_results}, on AIME 2025 with the 32B model, LT accuracy stagnates at 36.67\%, significantly underperforming \frameworkname{} (46.67\%). This performance gap stems from LT's reliance on coarse-grained metrics that aggregate signals over broad segments and the inductive bias that correct reasoning exhibits faster semantic evolution. Such heuristics fail on complex tasks where valid reasoning often requires "Active Exploration" (characterized by slower initial progress). In contrast, \frameworkname{} captures both \textit{fine-grained layer-wise variations} and \textit{dynamic step-level evolution}, enabling precise real-time intervention.

\vspace{2mm}
\noindent\textbf{Distribution Shift(vs. AdaptThink).} 
\textit{AdaptThink}~\cite{zhang2025adaptthink} demonstrates impressive efficacy on in-distribution mathematical tasks, notably achieving 87.18\% on GSM8k with the 7B model via its RL-optimized policy. Nevertheless, its performance is less robust under distribution shifts compared to \frameworkname{}. When evaluated on diverse out-of-distribution benchmarks using the same 7B backbone, \frameworkname{} consistently exhibits superior generalization. Specifically, on the scientific benchmark GPQA-D, \frameworkname{} attains 39.40\% against AdaptThink's 37.88\%. This advantage extends to general reasoning tasks like MMLU-Pro, where it scores 62.86\% compared to 60.00\%, and challenging competitions like AIME 2025, leading with 36.67\% over 33.33\%. This confirms that while AdaptThink excels within its training distribution, \frameworkname{} captures intrinsic cognitive patterns, offering a more generalizable training-free solution.

\vspace{-5pt}
\subsection{Ablation and In-Depth Analysis}
\label{sec:ablation}

\noindent\textbf{Ablation Study.} 
Table~\ref{tab:main_results} demonstrates that \frameworkname{} significantly outperforms its decoupled variants, confirming the necessity of the Phase-Momentum Alignment. For instance, on the MATH-500 benchmark with the 32B model, $\text{\frameworkname}_{\text{entropy}}$ and $\text{\frameworkname}_{\text{latent}}$ achieve accuracies of 85.40\% and 84.20\%, respectively. In contrast, the full \frameworkname{} effectively synergizes these signals to reach 91.40\%, achieving a substantial gain of over 6\% compared to the single-modal baselines.
To further analyze the source of this improvement, we define four reasoning quadrants formed by two orthogonal dimensions: the correctness of the model's original reasoning ($y_{ori}$) and the correctness after intervention ($y_{trunc}$), these quadrants are categorized as: (1) \textbf{Efficiency Gain} ($y_{ori}\checkmark, y_{trunc}\checkmark$), where truncation preserves correctness; (2) \textbf{Resource Saving} ($y_{ori}\times, y_{trunc}\times$), where tokens are saved on inevitable errors; (3) \textbf{Accuracy Gain} ($y_{ori}\times, y_{trunc}\checkmark$), representing rectified truncation from overthinking; and (4) \textbf{Accuracy Loss} ($y_{ori}\checkmark, y_{trunc}\times$), denoting premature failure. As illustrated in Figure~\ref{fig:main_results}(a). Empirically, while $\text{\frameworkname}_{\text{entropy}}$ and $\text{\frameworkname}_{\text{latent}}$ achieve 85.4\% and 84.2\% Efficiency Gain respectively, they suffer from 2.0\% and 4.0\% Accuracy Loss. In contrast, \frameworkname{} achieves the highest Efficiency Gain (87.8\%) with 0\% Accuracy Loss and a notable 3.6\% Accuracy Gain.
\label{sec:ablation}

\vspace{2mm}
\vspace{2mm}
\noindent\textbf{Fine-grained Mechanism Analysis.}
The performance disparity stems from the intrinsic limitations of uncoupled diagnostic signals. First, $\text{\frameworkname}_{\text{latent}}$ exhibits significant Accuracy Loss (4.0\%) due to the \textit{sensitivity dilemma} of the momentum threshold $\epsilon$: a loose $\epsilon$ triggers premature truncation before semantic convergence, severing valid reasoning paths essential for correct deduction, whereas a strictly tight $\epsilon$ fails to arrest "overthinking" in time. Second, $\text{\frameworkname}_{\text{entropy}}$ suffers from non-negligible loss (2.0\%) as low entropy can be deceptive, , thereby masking "confident hallucinations" where the model stabilizes superficially despite unsettled internal momentum. In contrast, the full \frameworkname{} resolves these trade-offs by subjecting $\epsilon$ to \textit{Phase constraints}. By functionally inhibiting truncation during active exploration (even if momentum fluctuates), \frameworkname{} acts as a semantic gatekeeper that guarantees intervention only upon dual-signal synchronization between geometric momentum and entropic phase. Consequently, this alignment eliminates premature interventions (0\% Accuracy Loss) while enabling the system to precisely rectify overthinking errors exactly when a valid intermediate thought is formed.

\vspace{-1.4mm}
\subsection{Sensitivity Analysis and Robustness}
\label{sec:sensitivity}

\noindent\textbf{Hyperparameter Robustness ($\alpha$ and $\epsilon$).} 
We investigate the system's sensitivity to the phase threshold $\alpha$ and momentum threshold $\epsilon$ within the interval $[0, 0.1]$ using the 14B model averaged across all benchmarks. As illustrated in Figure~\ref{fig:main_results}(c), both accuracy and token consumption remain largely invariant to parameter variations. We attribute this stability to the \textit{mutual constraint} mechanism inherent in \frameworkname{}. 
For $\epsilon$, consistent with \S\ref{sec:ablation}, the phase alignment requirement relaxes the precision needed for the momentum threshold, allowing a flexible range. 
For $\alpha$, although larger values might theoretically accelerate the detection of low-entropy states, the system avoids premature truncation because it is dually constrained by the momentum check ($\epsilon$). Since \frameworkname{} targets distinct high/low entropy regimes rather than transitional uncertainty, this orthogonal verification ensures the framework remains robust to hyperparameter fluctuations, eliminating the need for meticulous tuning.

\vspace{2mm}
\noindent\textbf{Impact of Proxy Model Scale ($M_{\phi}$).} 
We analyze the performance consistency when varying the proxy model size (from 1.5B to 14B). As illustrated in Figure~\ref{fig:main_results}(a), we observe a counter-intuitive trend where increasing proxy capacity exacerbates Accuracy Loss (rising to 2\% with 7B and 3\% with 14B). We attribute this to the \textit{contextual determinism} of smaller models: they rely strictly on the generated context for verification, yielding sharper entropy signals for correctness. In contrast, larger models, driven by vast parametric knowledge, may introduce unnecessary semantic ambiguity (associating plausibility with incorrect steps), thereby diluting the diagnostic signal. This finding validates the architectural advantage of \frameworkname{}: a lightweight 1.5B proxy is not only cost-effective but also intrinsically more accurate for monitoring large-scale reasoners.
\vspace{2mm}
\section{Conclusion}
In this work, we first posit the \textit{Phase-Momentum Alignment Hypothesis}, identifying that valid reasoning hinges on the temporal synchronization between entropic uncertainty and geometric cognitive effort. We rigorously formalize this insight via the \textit{Cognitive-Energy Model} to precisely distinguish active exploration from passive stagnation. To operationalize this, we propose \frameworkname, a training-free framework that optimizes test-time reasoning through dynamic diagnosis and intervention mechanism. Extensive experiments demonstrate that \frameworkname{} consistently outperforms representative baselines in information-theoretic, latent-based, and training-based paradigms, resulting in a superior accuracy-efficiency trade-off and robust cross-domain generalization. By establishing a transparent metric for reasoning integrity, our approach offers a robust and theoretically principled path for reliable inference scaling.
\clearpage
\appendix
% 1.5B results
\begin{table*}[t]
\centering
\small
\setlength{\tabcolsep}{1.2mm}
\caption{\textbf{Detailed Performance Comparison on Mathematical and General Reasoning Benchmarks.} This table provides the complete numerical results for DeepSeek-R1-Distill-Qwen-1.5B. We report Accuracy ($\uparrow$) and Average Token Cost ($\downarrow$). \frameworkname{} is compared against baselines from three distinct paradigms (Information-Theoretic, Latent-based, and Training-based) alongside our ablated variants. Note that $\Delta$ denotes the relative improvement of our full method over the Vanilla baseline.}
\label{tab:appendix_full_results}
\resizebox{0.95\textwidth}{!}{
\begin{tabular}{@{} l cccccc cc cc @{}}
\toprule
% === Header Level 1 ===
\multirow{3}{*}{\textbf{Methods}} 
& \multicolumn{6}{c}{\textbf{MATH}} 
& \multicolumn{2}{c}{\textbf{SCIENCE}} 
& \multicolumn{2}{c}{\textbf{GENERAL}} \\ 
\cmidrule(lr{1em}){2-7} \cmidrule(lr{1em}){8-9} \cmidrule(lr{0em}){10-11}

% === Header Level 2 ===
& \multicolumn{2}{c}{\textbf{GSM8K}} 
& \multicolumn{2}{c}{\textbf{MATH-500}} 
& \multicolumn{2}{c}{\textbf{AIME2025}}
& \multicolumn{2}{c}{\textbf{GPQA-D}}
& \multicolumn{2}{c}{\textbf{MMLU-Pro}} \\ 

% === Header Level 3 ===
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ 
& Acc.$\uparrow$ & Tokens$\downarrow$ \\ 
\midrule

% === DATA SECTION ===
% === DATA SECTION ===
\multicolumn{11}{c}{\textit{DeepSeek-R1-Distill-Qwen-1.5B}} \\ 
\hdashline[1pt/2pt] \addlinespace[2pt]

Vanilla      
& 64.22 & 532.99 
& 72.80 & 2759.63 
& 20.00 & 8166.60 
& 12.12 & 5002.78 
& 27.14 & 2706.99 \\

TracingtheTraces
& 63.85 & 1495.34
& 73.10 & 5620.15
& 16.67 & 5598.40
& 12.63 & 5605.22
& 26.50 & 5580.11 \\

ThinkorNot   
& 54.22 & 425.59 
& 57.36 & 1728.82 
& 14.00 & 5170.39 
& 12.12 & 3578.75 
& 27.14 & 1917.98 \\

AdaptThink   
& \textbf{70.56} & 506.43 
& 72.20 & \textbf{1285.72} 
& 23.33 & \textbf{5673.40} 
& 13.64 & 2842.57 
& 27.14 & 1632.61 \\

\midrule

% === OTHERS & ABLATIONS ===
\textbf{$\text{\frameworkname{}}_{\text{entropy}}$} 
& 55.88 & \textbf{249.55} 
& 60.80 & 2030.53 
& 13.33 & 7581.37 
& 13.13 & 4257.55 
& 28.57 & 2340.60 \\

\textbf{$\text{\frameworkname{}}_{\text{latent}}$} 
& 64.29 & 257.59 
& 72.20 & 2042.55 
& 20.00 & 7362.90 
& 25.25 & 3824.96 
& 32.86 & \textbf{904.57} \\

\rowcolor{MyColor2} 
\textbf{\text{\frameworkname{}}} 
& 67.15 & 250.87 
& \textbf{74.60} & 2106.88 
& \textbf{23.33} & 7581.37 
& \textbf{26.26} & \textbf{2452.99} 
& \textbf{37.14} & 2233.13 \\

% === DELTA ANALYSIS ===
\rowcolor{MyColor3} 
$\Delta$ vs. Vanilla
& +4.56\% & -52.93\%
& +2.47\% & -23.65\%
& +16.65\% & -7.17\%
& +116.66\% & -50.97\%
& +36.85\% & -17.51\% \\

\bottomrule
\end{tabular}
}
\vspace{-1.4mm}
\end{table*}





\section{Extended Experimental Main Results}
\label{sec:appendix_main_results}

\subsection{Scalability on Lightweight Models}
\label{sec:appendix_1.5b}

\noindent\textbf{Superior Pareto Frontier and Hallucination Mitigation.} 
Table~\ref{tab:appendix_full_results} presents the detailed performance of the DeepSeek-R1-Distill-Qwen-1.5B model across five benchmarks, demonstrating that \frameworkname{} establishes a superior accuracy-efficiency Pareto frontier even under limited model capacity. Compared to the Vanilla baseline, our framework significantly reduces computational overhead while maintaining or enhancing reasoning accuracy. Specifically, on mathematical tasks such as GSM8K and MATH-500, \frameworkname{} reduces token consumption by 52.93\% and 23.65\% respectively, with simultaneous accuracy gains of 4.56\% and 2.47\%. This efficiency advantage is even more pronounced on the challenging GPQA-Diamond benchmark, where the small model typically struggles with severe hallucinations (Vanilla Acc: 12.12\%). Here, \frameworkname{} achieves a remarkable relative improvement of 116.66\% (reaching 26.26\%), suggesting that our mechanism effectively diagnoses "stagnation" in incorrect reasoning paths and applies corrective interventions to convert wasted computation into valid reasoning attempts.

\vspace{2mm}
\noindent\textbf{Necessity of Phase-Momentum Alignment.} 
Furthermore, a comparative analysis against the ablated variants, $\text{\frameworkname{}}_{\text{entropy}}$ and $\text{\frameworkname{}}_{\text{latent}}$, highlights the critical necessity of the dual-signal Phase-Momentum Alignment for lightweight models. The entropy-only variant ($\text{\frameworkname{}}_{\text{entropy}}$) suffers from substantial performance degradation on hard tasks, such as AIME 2025, where its accuracy drops to 13.33\%—significantly below the Vanilla baseline (20.00\%). This failure stems from "confident hallucinations," where the 1.5B model exhibits low entropy despite flawed reasoning. Similarly, the latent-only variant ($\text{\frameworkname{}}_{\text{latent}}$) incurs higher token costs due to the lack of phase-aware triggering. In contrast, the full \frameworkname{} framework integrates these signals: it uses geometric momentum to filter out deceptive low-entropy states and uses phase monitoring to trigger geometric checks only when necessary. This coupling recovers the accuracy loss on AIME2025, which reaches 23.33\%, and achieves the lowest token consumption across most benchmarks, validating that the \textit{Cognitive-Energy Model} provides a robust diagnostic guardrail essential for trustworthy reasoning in small-scale models under resource constraints.

\subsection{Generalization vs. Training-based Methods}
\label{sec:appendix_training_baseline}

We benchmark \frameworkname{} against \textit{AdaptThink}, a representative method employing Reinforcement Learning for reasoning optimization. The empirical results reveal a distinct trade-off between in-distribution specialization and out-of-distribution (OOD) robustness. While \textit{AdaptThink} achieves peak accuracy on GSM8K (70.56\%), surpassing both Vanilla and our method, this performance gain appears highly dependent on the training distribution, essentially overfitting the policy to mathematical word problems. This hypothesis is substantiated by the sharp performance decline observed under distributional shifts: on the scientific reasoning benchmark GPQA-Diamond and the general reasoning benchmark MMLU-Pro, \textit{AdaptThink} significantly underperforms \frameworkname{} (13.64\% vs. 26.26\% and 27.14\% vs. 37.14\%, respectively). This contrast underscores the fundamental advantage of our training-free paradigm. Unlike policy-based methods that bias the model toward specific generation patterns learned during training, \frameworkname{} leverages the intrinsic \textit{Phase-Momentum Alignment}—a universal cognitive signature of the reasoning process. Consequently, our framework acts as a domain-agnostic supervisor, ensuring robust generalization across diverse reasoning tasks without the fragility associated with task-specific parameter updates.

\begin{figure*}[t]
  \centering
  \includegraphics[width=\textwidth]{figure/exp_more.pdf} % Replace with your actual filename
 \caption{\textbf{Hyperparameter Robustness Analysis.} Performance metrics (averaged across five benchmarks) for 1.5B, 7B, and 32B models under independent parameter sweeps. The x-axis denotes the value for $\alpha$ (blue) and $\epsilon$ (red).}
  \label{fig:appendix_alpha_sensitivity}
  \vspace{-5pt} % Optional: Adjust spacing to save vertical space
\end{figure*}

\subsection{Reliability vs. Entropy-based Methods}
\label{sec:appendix_entropy_baseline}

To isolate the contribution of geometric signals, we contrast \frameworkname{} with \textit{Think or Not} and the entropy-only ablation $\text{\frameworkname{}}_{\text{entropy}}$. The experimental data highlights a critical pathology in relying exclusively on information-theoretic metrics: "Deceptive Convergence." On challenging tasks like AIME 2025, the 1.5B model frequently enters states of "confident hallucination," characterized by low semantic entropy but incorrect logic. Entropy-based methods misinterpret this stability as valid convergence and prematurely truncate generation, causing accuracy to plummet to 13.33\%, well below the Vanilla baseline of 20.00\%. \frameworkname{} effectively resolves this issue by introducing \textit{Geometric Cognitive Effort} as an orthogonal verification dimension. By demanding that low-entropy states also exhibit relaxed latent momentum (signifying a true semantic fixed point), our framework filters out spurious confidence. This dual-signal validation successfully recovers the accuracy to 23.33\%, demonstrating that geometric dynamics serve as an indispensable verification mechanism for preserving reasoning integrity in lightweight models.

\subsection{Efficiency vs. Latent-based Methods}
\label{sec:appendix_latent_baseline}

Finally, we compare the computational efficiency of our online diagnostic protocol against \textit{Tracing the Traces}, a prominent latent-space selection method. The results expose the inherent inefficiency of the "generate-then-select" paradigm. Although \textit{Tracing the Traces} maintains competitive accuracy on benchmarks like MATH-500 (73.10\%), it imposes a prohibitive computational burden, consuming an average of 5620.15 tokens per query due to the necessity of generating multiple complete candidate paths. In sharp contrast, \frameworkname{} operates as a step-level "Cognitive Circuit Breaker," monitoring phase transitions in real-time and intervening only when stagnation is detected. This preemptive mechanism allows \frameworkname{} to achieve superior accuracy (74.60\%) while reducing token consumption to 2106.88 tokens—a reduction of over 60\%. This empirical evidence validates that event-triggered online diagnosis offers a far more scalable solution for test-time scaling than post-hoc selection, delivering high-precision control with minimal redundancy.

\section{Extended Sensitivity and Robustness Analysis}
\label{sec:appendix_sensitivity}

To rigorously validate the scalability and architectural universality of \frameworkname{}, we extend the sensitivity analysis to cover the spectrum from edge to high-capacity models (1.5B, 7B, and 32B). This multi-scale investigation aims to determine if the Phase-Momentum alignment captures intrinsic reasoning dynamics independent of model size. Figure~\ref{fig:appendix_alpha_sensitivity} presents performance metrics averaged across all five benchmarks. By mitigating task-specific variance through aggregation, we confirm that the observed robustness is intrinsic to the framework rather than dataset-specific.

\subsection{Robustness to Phase Sensitivity}
\label{sec:appendix_alpha_sensitivity}

The Phase Sensitivity $\alpha$ governs the strictness of the Entropy-based FSM. As evidenced by the \textbf{blue metrics} in Figure~\ref{fig:appendix_alpha_sensitivity}, \frameworkname{} demonstrates \textbf{invariant stability} across the sweep range $[0.00, 0.10]$.
\noindent\textbf{Dual-Signal Robustness:} Both average accuracy (blue line) and token consumption (blue bars) remain virtually constant across all model scales. This counter-intuitive stability validates the architectural superiority of our \textit{Cognitive-Energy Model}. Even if a looser $\alpha$ (e.g., $0.10$) prematurely flags a "Convergence Phase" based on entropy, the subsequent \textit{Geometric Cognitive Effort} check acts as an orthogonal gatekeeper. It prevents truncation unless the latent momentum strictly dissipates. Consequently, \frameworkname{} is resilient to $\alpha$ variations, eliminating the need for scale-specific tuning of the entropy threshold.

\subsection{Impact of Momentum Relaxation Threshold}
\label{sec:appendix_epsilon_sensitivity}

The Momentum Relaxation threshold $\epsilon$ primarily regulates the "thinking budget." As $\epsilon$ increases (relaxing strictness), token consumption consistently decreases (red bars). However, accuracy reveals distinct scale-dependent dynamics:

\noindent\textbf{Small Models (1.5B \& 7B): Efficiency without Penalty.}
Accuracy remains flat even as $\epsilon$ increases to $0.1$. This implies that additional tokens under tighter thresholds ($\epsilon < 0.04$) consist largely of redundant generation rather than constructive reasoning. \frameworkname{} effectively prunes this stagnation, converting computational waste into savings without compromising correctness.

\noindent\textbf{Large Models (32B): The Depth Trade-off.}
For QwQ-32B, substantial cost reduction comes with a marginal accuracy decline. This reflects deeper reasoning depth, where valid reasoning may exhibit \textit{transient low momentum} (e.g., restructuring logic) before semantic resolution. A relaxed $\epsilon$ might aggressively prune these pauses. Crucially, however, this trade-off is favorable: a negligible accuracy dip yields disproportionate savings, offering a adjustable parameter to significantly lower latency while maintaining reasoning fidelity.

\section{Case Study}

\begin{center}
    \begin{minipage}{\textwidth}
    \centering
    \nopagebreak
  \includegraphics[width=0.8\textwidth]{figure/case_study.pdf} % Replace with your actual filename
  \captionof{figure}{\textbf{Comparison on a MATH500 geometry task. QwQ-32B generates a lengthy but flawed reasoning chain based on incorrect angle properties. In contrast, PUMA leverages retrieved geometric principles to restart the reasoning process, correctly identifying the angle relationships and achieving the correct answer.}}
  \label{fig:appendix_case}
  \end{minipage}
\end{center}

% QwQ-32B/MATH500/PUMA(Retrieval-Augmented Restart)/Accuracy Gain




%%% -*-BibTeX-*-
%%% Do NOT edit. File created by BibTeX with style
%%% ACM-Reference-Format-Journals [18-Jan-2012].

\begin{thebibliography}{49}

%%% ====================================================================
%%% NOTE TO THE USER: you can override these defaults by providing
%%% customized versions of any of these macros before the \bibliography
%%% command.  Each of them MUST provide its own final punctuation,
%%% except for \shownote{}, \showDOI{}, and \showURL{}.  The latter two
%%% do not use final punctuation, in order to avoid confusing it with
%%% the Web address.
%%%
%%% To suppress output of a particular field, define its macro to expand
%%% to an empty string, or better, \unskip, like this:
%%%
%%% \newcommand{\showDOI}[1]{\unskip}   % LaTeX syntax
%%%
%%% \def \showDOI #1{\unskip}           % plain TeX syntax
%%%
%%% ====================================================================

\ifx \showCODEN    \undefined \def \showCODEN     #1{\unskip}     \fi
\ifx \showDOI      \undefined \def \showDOI       #1{#1}\fi
\ifx \showISBNx    \undefined \def \showISBNx     #1{\unskip}     \fi
\ifx \showISBNxiii \undefined \def \showISBNxiii  #1{\unskip}     \fi
\ifx \showISSN     \undefined \def \showISSN      #1{\unskip}     \fi
\ifx \showLCCN     \undefined \def \showLCCN      #1{\unskip}     \fi
\ifx \shownote     \undefined \def \shownote      #1{#1}          \fi
\ifx \showarticletitle \undefined \def \showarticletitle #1{#1}   \fi
\ifx \showURL      \undefined \def \showURL       {\relax}        \fi
% The following commands are used for tagged output and should be
% invisible to TeX
\providecommand\bibfield[2]{#2}
\providecommand\bibinfo[2]{#2}
\providecommand\natexlab[1]{#1}
\providecommand\showeprint[2][]{arXiv:#2}

\bibitem[Aggarwal and Welleck(2025)]%
        {aggarwal2025l1}
\bibfield{author}{\bibinfo{person}{Pranjal Aggarwal} {and} \bibinfo{person}{Sean Welleck}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{L1: Controlling how long a reasoning model thinks with reinforcement learning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.04697}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Ali et~al\mbox{.}(2025)]%
        {ali2025entropy}
\bibfield{author}{\bibinfo{person}{Riccardo Ali}, \bibinfo{person}{Francesco Caso}, \bibinfo{person}{Christopher Irwin}, {and} \bibinfo{person}{Pietro Li{\`o}}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Entropy-lens: The information signature of transformer computations}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.16570}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Arora and Zanette(2025)]%
        {arora2025training}
\bibfield{author}{\bibinfo{person}{Daman Arora} {and} \bibinfo{person}{Andrea Zanette}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Training language models to reason efficiently}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.04463}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Aytes et~al\mbox{.}(2025)]%
        {aytes2025sketch}
\bibfield{author}{\bibinfo{person}{Simon~A Aytes}, \bibinfo{person}{Jinheon Baek}, {and} \bibinfo{person}{Sung~Ju Hwang}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.05179}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Chen et~al\mbox{.}(2024b)]%
        {chen2024unlocking}
\bibfield{author}{\bibinfo{person}{Qiguang Chen}, \bibinfo{person}{Libo Qin}, \bibinfo{person}{Jiaqi Wang}, \bibinfo{person}{Jingxuan Zhou}, {and} \bibinfo{person}{Wanxiang Che}.} \bibinfo{year}{2024}\natexlab{b}.
\newblock \showarticletitle{Unlocking the capabilities of thought: A reasoning boundary framework to quantify and optimize chain-of-thought}.
\newblock \bibinfo{journal}{\emph{Advances in Neural Information Processing Systems}}  \bibinfo{volume}{37} (\bibinfo{year}{2024}), \bibinfo{pages}{54872--54904}.
\newblock


\bibitem[Chen et~al\mbox{.}(2024a)]%
        {chen2024learning}
\bibfield{author}{\bibinfo{person}{Xin Chen}, \bibinfo{person}{Hanxian Huang}, \bibinfo{person}{Yanjun Gao}, \bibinfo{person}{Yi Wang}, \bibinfo{person}{Jishen Zhao}, {and} \bibinfo{person}{Ke Ding}.} \bibinfo{year}{2024}\natexlab{a}.
\newblock \showarticletitle{Learning to maximize mutual information for chain-of-thought distillation}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2403.03348}} (\bibinfo{year}{2024}).
\newblock


\bibitem[Chen et~al\mbox{.}(2025)]%
        {chen20252}
\bibfield{author}{\bibinfo{person}{Zhuo Chen}, \bibinfo{person}{Zhuotao Jin}, \bibinfo{person}{Di Luo}, \bibinfo{person}{Marin Solja{\v{c}}i{\'c}}, {et~al\mbox{.}}} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{{L$^2$M}: Mutual Information Scaling Law for Long-Context Language Modeling}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.04725}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Cobbe et~al\mbox{.}(2021)]%
        {cobbe2021training}
\bibfield{author}{\bibinfo{person}{Karl Cobbe}, \bibinfo{person}{Vineet Kosaraju}, \bibinfo{person}{Mohammad Bavarian}, \bibinfo{person}{Mark Chen}, \bibinfo{person}{Heewoo Jun}, \bibinfo{person}{Lukasz Kaiser}, \bibinfo{person}{Matthias Plappert}, \bibinfo{person}{Jerry Tworek}, \bibinfo{person}{Jacob Hilton}, \bibinfo{person}{Reiichiro Nakano}, {et~al\mbox{.}}} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Training verifiers to solve math word problems}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2110.14168}} (\bibinfo{year}{2021}).
\newblock


\bibitem[Cover(1999)]%
        {cover1999elements}
\bibfield{author}{\bibinfo{person}{Thomas~M Cover}.} \bibinfo{year}{1999}\natexlab{}.
\newblock \bibinfo{booktitle}{\emph{Elements of information theory}}.
\newblock \bibinfo{publisher}{John Wiley \& Sons}.
\newblock


\bibitem[Farquhar et~al\mbox{.}(2024)]%
        {farquhar2024detecting}
\bibfield{author}{\bibinfo{person}{Sebastian Farquhar}, \bibinfo{person}{Jannik Kossen}, \bibinfo{person}{Lorenz Kuhn}, {and} \bibinfo{person}{Yarin Gal}.} \bibinfo{year}{2024}\natexlab{}.
\newblock \showarticletitle{Detecting hallucinations in large language models using semantic entropy}.
\newblock \bibinfo{journal}{\emph{Nature}} \bibinfo{volume}{630}, \bibinfo{number}{8017} (\bibinfo{year}{2024}), \bibinfo{pages}{625--630}.
\newblock


\bibitem[Feng et~al\mbox{.}(2025)]%
        {feng2025efficient}
\bibfield{author}{\bibinfo{person}{Sicheng Feng}, \bibinfo{person}{Gongfan Fang}, \bibinfo{person}{Xinyin Ma}, {and} \bibinfo{person}{Xinchao Wang}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Efficient reasoning models: A survey}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2504.10903}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Gan et~al\mbox{.}(2025)]%
        {gan2025rethinking}
\bibfield{author}{\bibinfo{person}{Zeyu Gan}, \bibinfo{person}{Yun Liao}, {and} \bibinfo{person}{Yong Liu}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Rethinking external slow-thinking: From snowball errors to probability of correct reasoning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.15602}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Guo et~al\mbox{.}(2025)]%
        {guo2025deepseek}
\bibfield{author}{\bibinfo{person}{Daya Guo}, \bibinfo{person}{Dejian Yang}, \bibinfo{person}{Haowei Zhang}, \bibinfo{person}{Junxiao Song}, \bibinfo{person}{Ruoyu Zhang}, \bibinfo{person}{Runxin Xu}, \bibinfo{person}{Qihao Zhu}, \bibinfo{person}{Shirong Ma}, \bibinfo{person}{Peiyi Wang}, \bibinfo{person}{Xiao Bi}, {et~al\mbox{.}}} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.12948}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Han et~al\mbox{.}(2025)]%
        {han2025token}
\bibfield{author}{\bibinfo{person}{Tingxu Han}, \bibinfo{person}{Zhenting Wang}, \bibinfo{person}{Chunrong Fang}, \bibinfo{person}{Shiyu Zhao}, \bibinfo{person}{Shiqing Ma}, {and} \bibinfo{person}{Zhenyu Chen}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Token-budget-aware llm reasoning}. In \bibinfo{booktitle}{\emph{Findings of the Association for Computational Linguistics: ACL 2025}}. \bibinfo{pages}{24842--24855}.
\newblock


\bibitem[Hendrycks et~al\mbox{.}(2021)]%
        {hendrycks2021measuring}
\bibfield{author}{\bibinfo{person}{Dan Hendrycks}, \bibinfo{person}{Collin Burns}, \bibinfo{person}{Saurav Kadavath}, \bibinfo{person}{Akul Arora}, \bibinfo{person}{Steven Basart}, \bibinfo{person}{Eric Tang}, \bibinfo{person}{Dawn Song}, {and} \bibinfo{person}{Jacob Steinhardt}.} \bibinfo{year}{2021}\natexlab{}.
\newblock \showarticletitle{Measuring mathematical problem solving with the math dataset}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2103.03874}} (\bibinfo{year}{2021}).
\newblock


\bibitem[Jaech et~al\mbox{.}(2024)]%
        {jaech2024openai}
\bibfield{author}{\bibinfo{person}{Aaron Jaech}, \bibinfo{person}{Adam Kalai}, \bibinfo{person}{Adam Lerer}, \bibinfo{person}{Adam Richardson}, \bibinfo{person}{Ahmed El-Kishky}, \bibinfo{person}{Aiden Low}, \bibinfo{person}{Alec Helyar}, \bibinfo{person}{Aleksander Madry}, \bibinfo{person}{Alex Beutel}, \bibinfo{person}{Alex Carney}, {et~al\mbox{.}}} \bibinfo{year}{2024}\natexlab{}.
\newblock \showarticletitle{Openai o1 system card}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2412.16720}} (\bibinfo{year}{2024}).
\newblock


\bibitem[Jeon(2025)]%
        {jeon2025information}
\bibfield{author}{\bibinfo{person}{Hong~Jun Jeon}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \emph{\bibinfo{title}{Information-theoretic foundations for machine learning}}.
\newblock \bibinfo{thesistype}{Ph.\,D. Dissertation}. \bibinfo{school}{Stanford University}.
\newblock


\bibitem[Kang et~al\mbox{.}(2025)]%
        {kang2025c3ot}
\bibfield{author}{\bibinfo{person}{Yu Kang}, \bibinfo{person}{Xianghui Sun}, \bibinfo{person}{Liangyu Chen}, {and} \bibinfo{person}{Wei Zou}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{C3ot: Generating shorter chain-of-thought without compromising effectiveness}. In \bibinfo{booktitle}{\emph{Proceedings of the AAAI Conference on Artificial Intelligence}}, Vol.~\bibinfo{volume}{39}. \bibinfo{pages}{24312--24320}.
\newblock


\bibitem[Kossen et~al\mbox{.}(2024)]%
        {kossen2024semantic}
\bibfield{author}{\bibinfo{person}{Jannik Kossen}, \bibinfo{person}{Jiatong Han}, \bibinfo{person}{Muhammed Razzak}, \bibinfo{person}{Lisa Schut}, \bibinfo{person}{Shreshth Malik}, {and} \bibinfo{person}{Yarin Gal}.} \bibinfo{year}{2024}\natexlab{}.
\newblock \showarticletitle{Semantic entropy probes: Robust and cheap hallucination detection in llms}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2406.15927}} (\bibinfo{year}{2024}).
\newblock


\bibitem[Lee et~al\mbox{.}(2025)]%
        {lee2025well}
\bibfield{author}{\bibinfo{person}{Ayeong Lee}, \bibinfo{person}{Ethan Che}, {and} \bibinfo{person}{Tianyi Peng}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{How well do llms compress their own chain-of-thought? a token complexity approach}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.01141}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Li et~al\mbox{.}(2025)]%
        {li2025core}
\bibfield{author}{\bibinfo{person}{Haoxi Li}, \bibinfo{person}{Sikai Bai}, \bibinfo{person}{Jie Zhang}, {and} \bibinfo{person}{Song Guo}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{CoRE: Enhancing Metacognition with Label-free Self-evaluation in LRMs}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2507.06087}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Liao et~al\mbox{.}(2025)]%
        {liao2025reward}
\bibfield{author}{\bibinfo{person}{Baohao Liao}, \bibinfo{person}{Yuhui Xu}, \bibinfo{person}{Hanze Dong}, \bibinfo{person}{Junnan Li}, \bibinfo{person}{Christof Monz}, \bibinfo{person}{Silvio Savarese}, \bibinfo{person}{Doyen Sahoo}, {and} \bibinfo{person}{Caiming Xiong}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Reward-guided speculative decoding for efficient llm reasoning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.19324}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Luo et~al\mbox{.}(2025)]%
        {luo2025o1}
\bibfield{author}{\bibinfo{person}{Haotian Luo}, \bibinfo{person}{Li Shen}, \bibinfo{person}{Haiying He}, \bibinfo{person}{Yibo Wang}, \bibinfo{person}{Shiwei Liu}, \bibinfo{person}{Wei Li}, \bibinfo{person}{Naiqiang Tan}, \bibinfo{person}{Xiaochun Cao}, {and} \bibinfo{person}{Dacheng Tao}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.12570}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Ma et~al\mbox{.}(2025a)]%
        {ma2025reasoning}
\bibfield{author}{\bibinfo{person}{Wenjie Ma}, \bibinfo{person}{Jingxuan He}, \bibinfo{person}{Charlie Snell}, \bibinfo{person}{Tyler Griggs}, \bibinfo{person}{Sewon Min}, {and} \bibinfo{person}{Matei Zaharia}.} \bibinfo{year}{2025}\natexlab{a}.
\newblock \showarticletitle{Reasoning models can be effective without thinking}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2504.09858}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Ma et~al\mbox{.}(2025b)]%
        {ma2025cot}
\bibfield{author}{\bibinfo{person}{Xinyin Ma}, \bibinfo{person}{Guangnian Wan}, \bibinfo{person}{Runpeng Yu}, \bibinfo{person}{Gongfan Fang}, {and} \bibinfo{person}{Xinchao Wang}.} \bibinfo{year}{2025}\natexlab{b}.
\newblock \showarticletitle{Cot-valve: Length-compressible chain-of-thought tuning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.09601}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Qu et~al\mbox{.}(2025)]%
        {qu2025survey}
\bibfield{author}{\bibinfo{person}{Xiaoye Qu}, \bibinfo{person}{Yafu Li}, \bibinfo{person}{Zhaochen Su}, \bibinfo{person}{Weigao Sun}, \bibinfo{person}{Jianhao Yan}, \bibinfo{person}{Dongrui Liu}, \bibinfo{person}{Ganqu Cui}, \bibinfo{person}{Daizong Liu}, \bibinfo{person}{Shuxian Liang}, \bibinfo{person}{Junxian He}, {et~al\mbox{.}}} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{A survey of efficient reasoning for large reasoning models: Language, multimodality, and beyond}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.21614}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Rein et~al\mbox{.}(2024)]%
        {rein2024gpqa}
\bibfield{author}{\bibinfo{person}{David Rein}, \bibinfo{person}{Betty~Li Hou}, \bibinfo{person}{Asa~Cooper Stickland}, \bibinfo{person}{Jackson Petty}, \bibinfo{person}{Richard~Yuanzhe Pang}, \bibinfo{person}{Julien Dirani}, \bibinfo{person}{Julian Michael}, {and} \bibinfo{person}{Samuel~R Bowman}.} \bibinfo{year}{2024}\natexlab{}.
\newblock \showarticletitle{Gpqa: A graduate-level google-proof q\&a benchmark}. In \bibinfo{booktitle}{\emph{First Conference on Language Modeling}}.
\newblock


\bibitem[Ren and Liu(2025)]%
        {ren2025revisiting}
\bibfield{author}{\bibinfo{person}{Ruifeng Ren} {and} \bibinfo{person}{Yong Liu}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Revisiting transformers through the lens of low entropy and dynamic sparsity}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2504.18929}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Shen et~al\mbox{.}(2025a)]%
        {shen2025efficient}
\bibfield{author}{\bibinfo{person}{Xuan Shen}, \bibinfo{person}{Yizhou Wang}, \bibinfo{person}{Xiangxi Shi}, \bibinfo{person}{Yanzhi Wang}, \bibinfo{person}{Pu Zhao}, {and} \bibinfo{person}{Jiuxiang Gu}.} \bibinfo{year}{2025}\natexlab{a}.
\newblock \showarticletitle{Efficient reasoning with hidden thinking}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.19201}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Shen et~al\mbox{.}(2025c)]%
        {shen2025dast}
\bibfield{author}{\bibinfo{person}{Yi Shen}, \bibinfo{person}{Jian Zhang}, \bibinfo{person}{Jieyun Huang}, \bibinfo{person}{Shuming Shi}, \bibinfo{person}{Wenjing Zhang}, \bibinfo{person}{Jiangze Yan}, \bibinfo{person}{Ning Wang}, \bibinfo{person}{Kai Wang}, \bibinfo{person}{Zhaoxiang Liu}, {and} \bibinfo{person}{Shiguo Lian}.} \bibinfo{year}{2025}\natexlab{c}.
\newblock \showarticletitle{Dast: Difficulty-adaptive slow-thinking for large reasoning models}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.04472}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Shen et~al\mbox{.}(2025b)]%
        {shen2025codi}
\bibfield{author}{\bibinfo{person}{Zhenyi Shen}, \bibinfo{person}{Hanqi Yan}, \bibinfo{person}{Linhai Zhang}, \bibinfo{person}{Zhanghao Hu}, \bibinfo{person}{Yali Du}, {and} \bibinfo{person}{Yulan He}.} \bibinfo{year}{2025}\natexlab{b}.
\newblock \showarticletitle{Codi: Compressing chain-of-thought into continuous space via self-distillation}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.21074}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Su et~al\mbox{.}(2025)]%
        {su2025token}
\bibfield{author}{\bibinfo{person}{DiJia Su}, \bibinfo{person}{Hanlin Zhu}, \bibinfo{person}{Yingchen Xu}, \bibinfo{person}{Jiantao Jiao}, \bibinfo{person}{Yuandong Tian}, {and} \bibinfo{person}{Qinqing Zheng}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Token assorted: Mixing latent and text tokens for improved language model reasoning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.03275}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Sui et~al\mbox{.}(2025)]%
        {sui2025stop}
\bibfield{author}{\bibinfo{person}{Yang Sui}, \bibinfo{person}{Yu-Neng Chuang}, \bibinfo{person}{Guanchu Wang}, \bibinfo{person}{Jiamu Zhang}, \bibinfo{person}{Tianyi Zhang}, \bibinfo{person}{Jiayi Yuan}, \bibinfo{person}{Hongyi Liu}, \bibinfo{person}{Andrew Wen}, \bibinfo{person}{Shaochen Zhong}, \bibinfo{person}{Na Zou}, {et~al\mbox{.}}} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Stop overthinking: A survey on efficient reasoning for large language models}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.16419}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Team et~al\mbox{.}(2025)]%
        {team2025kimi}
\bibfield{author}{\bibinfo{person}{Kimi Team}, \bibinfo{person}{Angang Du}, \bibinfo{person}{Bofei Gao}, \bibinfo{person}{Bowei Xing}, \bibinfo{person}{Changjiu Jiang}, \bibinfo{person}{Cheng Chen}, \bibinfo{person}{Cheng Li}, \bibinfo{person}{Chenjun Xiao}, \bibinfo{person}{Chenzhuang Du}, \bibinfo{person}{Chonghua Liao}, {et~al\mbox{.}}} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Kimi k1. 5: Scaling reinforcement learning with llms}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2501.12599}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Team(2025)]%
        {team2025qwq}
\bibfield{author}{\bibinfo{person}{Qwen Team}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \bibinfo{title}{Qwq-32b: Embracing the power of reinforcement learning}.
\newblock
\newblock


\bibitem[Ton et~al\mbox{.}(2024)]%
        {ton2024understanding}
\bibfield{author}{\bibinfo{person}{Jean-Francois Ton}, \bibinfo{person}{Muhammad~Faaiz Taufiq}, {and} \bibinfo{person}{Yang Liu}.} \bibinfo{year}{2024}\natexlab{}.
\newblock \showarticletitle{Understanding chain-of-thought in llms through information theory}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2411.11984}} (\bibinfo{year}{2024}).
\newblock


\bibitem[Turner et~al\mbox{.}(2023)]%
        {turner2023steering}
\bibfield{author}{\bibinfo{person}{Alexander~Matt Turner}, \bibinfo{person}{Lisa Thiergart}, \bibinfo{person}{Gavin Leech}, \bibinfo{person}{David Udell}, \bibinfo{person}{Juan~J Vazquez}, \bibinfo{person}{Ulisse Mini}, {and} \bibinfo{person}{Monte MacDiarmid}.} \bibinfo{year}{2023}\natexlab{}.
\newblock \showarticletitle{Steering language models with activation engineering}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2308.10248}} (\bibinfo{year}{2023}).
\newblock


\bibitem[Vilas et~al\mbox{.}(2025)]%
        {vilas2025tracing}
\bibfield{author}{\bibinfo{person}{Martina~G Vilas}, \bibinfo{person}{Safoora Yousefi}, \bibinfo{person}{Besmira Nushi}, \bibinfo{person}{Eric Horvitz}, {and} \bibinfo{person}{Vidhisha Balachandran}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2510.10494}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Wang et~al\mbox{.}(2025)]%
        {wang2025harnessing}
\bibfield{author}{\bibinfo{person}{Rui Wang}, \bibinfo{person}{Hongru Wang}, \bibinfo{person}{Boyang Xue}, \bibinfo{person}{Jianhui Pang}, \bibinfo{person}{Shudong Liu}, \bibinfo{person}{Yi Chen}, \bibinfo{person}{Jiahao Qiu}, \bibinfo{person}{Derek~Fai Wong}, \bibinfo{person}{Heng Ji}, {and} \bibinfo{person}{Kam-Fai Wong}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Harnessing the reasoning economy: A survey of efficient reasoning for large language models}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2503.24377}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Wang et~al\mbox{.}(2024a)]%
        {wang2024mmlu}
\bibfield{author}{\bibinfo{person}{Yubo Wang}, \bibinfo{person}{Xueguang Ma}, \bibinfo{person}{Ge Zhang}, \bibinfo{person}{Yuansheng Ni}, \bibinfo{person}{Abhranil Chandra}, \bibinfo{person}{Shiguang Guo}, \bibinfo{person}{Weiming Ren}, \bibinfo{person}{Aaran Arulraj}, \bibinfo{person}{Xuan He}, \bibinfo{person}{Ziyan Jiang}, {et~al\mbox{.}}} \bibinfo{year}{2024}\natexlab{a}.
\newblock \showarticletitle{Mmlu-pro: A more robust and challenging multi-task language understanding benchmark}.
\newblock \bibinfo{journal}{\emph{Advances in Neural Information Processing Systems}}  \bibinfo{volume}{37} (\bibinfo{year}{2024}), \bibinfo{pages}{95266--95290}.
\newblock


\bibitem[Wang et~al\mbox{.}(2024b)]%
        {wang2024latent}
\bibfield{author}{\bibinfo{person}{Yiming Wang}, \bibinfo{person}{Pei Zhang}, \bibinfo{person}{Baosong Yang}, \bibinfo{person}{Derek~F Wong}, {and} \bibinfo{person}{Rui Wang}.} \bibinfo{year}{2024}\natexlab{b}.
\newblock \showarticletitle{Latent space chain-of-embedding enables output-free llm self-evaluation}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2410.13640}} (\bibinfo{year}{2024}).
\newblock


\bibitem[Wei et~al\mbox{.}(2022)]%
        {wei2022chain}
\bibfield{author}{\bibinfo{person}{Jason Wei}, \bibinfo{person}{Xuezhi Wang}, \bibinfo{person}{Dale Schuurmans}, \bibinfo{person}{Maarten Bosma}, \bibinfo{person}{Fei Xia}, \bibinfo{person}{Ed Chi}, \bibinfo{person}{Quoc~V Le}, \bibinfo{person}{Denny Zhou}, {et~al\mbox{.}}} \bibinfo{year}{2022}\natexlab{}.
\newblock \showarticletitle{Chain-of-thought prompting elicits reasoning in large language models}.
\newblock \bibinfo{journal}{\emph{Advances in neural information processing systems}}  \bibinfo{volume}{35} (\bibinfo{year}{2022}), \bibinfo{pages}{24824--24837}.
\newblock


\bibitem[Wu et~al\mbox{.}(2025)]%
        {wu2025more}
\bibfield{author}{\bibinfo{person}{Yuyang Wu}, \bibinfo{person}{Yifei Wang}, \bibinfo{person}{Ziyu Ye}, \bibinfo{person}{Tianqi Du}, \bibinfo{person}{Stefanie Jegelka}, {and} \bibinfo{person}{Yisen Wang}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{When more is less: Understanding chain-of-thought length in llms}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.07266}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Xia et~al\mbox{.}(2025)]%
        {xia2025tokenskip}
\bibfield{author}{\bibinfo{person}{Heming Xia}, \bibinfo{person}{Chak~Tou Leong}, \bibinfo{person}{Wenjie Wang}, \bibinfo{person}{Yongqi Li}, {and} \bibinfo{person}{Wenjie Li}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Tokenskip: Controllable chain-of-thought compression in llms}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.12067}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Xu et~al\mbox{.}(2025)]%
        {xu2025softcot}
\bibfield{author}{\bibinfo{person}{Yige Xu}, \bibinfo{person}{Xu Guo}, \bibinfo{person}{Zhiwei Zeng}, {and} \bibinfo{person}{Chunyan Miao}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Softcot: Soft chain-of-thought for efficient reasoning with llms}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2502.12134}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Yang et~al\mbox{.}({[n.\,d.]})]%
        {yangdemystifying}
\bibfield{author}{\bibinfo{person}{Shiming Yang}, \bibinfo{person}{Yuxuan Tong}, \bibinfo{person}{Xinyao Niu}, \bibinfo{person}{Graham Neubig}, {and} \bibinfo{person}{Xiang Yue}.} \bibinfo{year}{[n.\,d.]}\natexlab{}.
\newblock \showarticletitle{Demystifying Long Chain-of-Thought Reasoning}. In \bibinfo{booktitle}{\emph{Forty-second International Conference on Machine Learning}}.
\newblock


\bibitem[Yong et~al\mbox{.}(2025)]%
        {yong2025think}
\bibfield{author}{\bibinfo{person}{Xixian Yong}, \bibinfo{person}{Xiao Zhou}, \bibinfo{person}{Yingying Zhang}, \bibinfo{person}{Jinlin Li}, \bibinfo{person}{Yefeng Zheng}, {and} \bibinfo{person}{Xian Wu}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2505.18237}} (\bibinfo{year}{2025}).
\newblock


\bibitem[Yuksekgonul et~al\mbox{.}(2023)]%
        {yuksekgonul2023attention}
\bibfield{author}{\bibinfo{person}{Mert Yuksekgonul}, \bibinfo{person}{Varun Chandrasekaran}, \bibinfo{person}{Erik Jones}, \bibinfo{person}{Suriya Gunasekar}, \bibinfo{person}{Ranjita Naik}, \bibinfo{person}{Hamid Palangi}, \bibinfo{person}{Ece Kamar}, {and} \bibinfo{person}{Besmira Nushi}.} \bibinfo{year}{2023}\natexlab{}.
\newblock \showarticletitle{Attention satisfies: A constraint-satisfaction lens on factual errors of language models}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2309.15098}} (\bibinfo{year}{2023}).
\newblock


\bibitem[Zhang et~al\mbox{.}(2025)]%
        {zhang2025adaptthink}
\bibfield{author}{\bibinfo{person}{Jiajie Zhang}, \bibinfo{person}{Nianyi Lin}, \bibinfo{person}{Lei Hou}, \bibinfo{person}{Ling Feng}, {and} \bibinfo{person}{Juanzi Li}.} \bibinfo{year}{2025}\natexlab{}.
\newblock \showarticletitle{Adaptthink: Reasoning models can learn when to think}.
\newblock \bibinfo{journal}{\emph{arXiv preprint arXiv:2505.13417}} (\bibinfo{year}{2025}).
\newblock


\end{thebibliography}