PaperlayerView exact revision on arXiv

Source-linked AI summary

CoT-Kinetics: A Theoretical Modeling Assessing LRM Reasoning Process

Jinhe Bi, Danqi Yan, Yifan Wang, Wenke Huang, Haokun Chen, Guancheng Wan, Mang Ye, Xun Xiao, Hinrich Schuetze, Volker Tresp, Yunpu Ma

arXiv:2505.13408v1cs.AIcs.CL

TL;DR

Evaluating LRM outputs requires distinguishing answer correctness from the soundness of the reasoning trajectory. The paper introduces CoT-Kinetics, a mechanics-inspired energy score built from internal token-state dynamics and uncertainty, and reports that it better reflects the relationship between reasoning trajectories and final answers. Its broader validation is constrained by limited availability of open-source extremely large LRMs.

  • Problem

    Correct final answers can arise from flawed reasoning, while existing methods do not adequately reflect the causal relationship between reasoning trajectories and concluded answers.

  • Method

    CoT-Kinetics models layer-wise reasoning-token state transitions as particle dynamics and formulates their semantic kinetic energy with query-induced uncertainty.

  • Results

    Across seven open-source LRMs and six challenging benchmarks, CoT-Kinetics is reported to be a better modeling approach than state-of-the-art methods for assessing reasoning soundness and output quality.

  • Takeaways & Limitations

    The scalar CoT-Kinetics score provides an internal way to assess reasoning soundness and overall LRM output quality without external information.

  • Takeaways & Limitations

    Validation is constrained by the limited availability of open-source models, particularly extremely large-scale LRMs.

Abstract

from arXiv · show

Recent Large Reasoning Models significantly improve the reasoning ability of Large Language Models by learning to reason, exhibiting the promising performance in solving complex tasks. LRMs solve tasks that require complex reasoning by explicitly generating reasoning trajectories together with answers. Nevertheless, judging the quality of such an output answer is not easy because only considering the correctness of the answer is not enough and the soundness of the reasoning trajectory part matters as well. Logically, if the soundness of the reasoning part is poor, even if the answer is correct, the confidence of the derived answer should be low. Existing methods did consider jointly assessing the overall output answer by taking into account the reasoning part, however, their capability is still not satisfactory as the causal relationship of the reasoning to the concluded answer cannot properly reflected. In this paper, inspired by classical mechanics, we present a novel approach towards establishing a CoT-Kinetics energy equation. Specifically, our CoT-Kinetics energy equation formulates the token state transformation process, which is regulated by LRM internal transformer layers, as like a particle kinetics dynamics governed in a mechanical field. Our CoT-Kinetics energy assigns a scalar score to evaluate specifically the soundness of the reasoning phase, telling how confident the derived answer could be given the evaluated reasoning. As such, the LRM's overall output quality can be accurately measured, rather than a coarse judgment (e.g., correct or incorrect) anymore.

1 Introduction

LRMs generate reasoning trajectories alongside answers, but answer correctness alone can overlook flawed or unstable reasoning. CoT-Kinetics is proposed as an internal, mechanics-inspired score to assess reasoning soundness and better reflect its relationship to answer quality.

  • Motivation: A correct final answer can coexist with flawed assumptions or unstable reasoning, making answer-only evaluation insufficient.The reasoning trajectory must be assessed separately because its soundness affects confidence in the derived answer.
  • Existing Methods: Existing assessment methods use either external supervision or internal uncertainty and semantic-consistency signals.External approaches include human annotations, probing, and judge models, while internal approaches quantify uncertainty or hidden-state consistency.
  • Existing Methods: External methods have limited generality when supervision is unavailable or query distributions differ from training environments.Internal methods avoid external information but are described as theoretically inadequate and insufficiently causal in their assessment.
  • Proposed Approach: CoT-Kinetics models reasoning-token state transitions as particle dynamics under a force field regulated by transformer layers.Its energy equation characterizes intrinsic semantic dynamics while incorporating uncertainty induced by the query task.
  • Evaluation: CoT-Kinetics uses internal reasoning states to score reasoning soundness and assess overall output quality without external supervision.The paper reports evaluations across seven open-source LRMs, six challenging benchmarks, and AUROC, AUPR, and FPR@95 metrics.

2 CoT-Kinetics Energy: Modeling LRM CoT Reasoning Process

The method represents layer-wise reasoning-token evolution using pooled hidden states and internal uncertainty. It combines semantic progression, curvature-based correction, and entropy into a scalar CoT-Kinetics energy score for reasoning soundness.

  • 2.1 Formulation of LRM CoT Reasoning: LRMs generate reasoning tokens between designated special tokens, and their hidden states are updated layer by layer through transformer computations.The reasoning sequence contains K tokens, with each token state represented in a d-dimensional hidden space across L transformer layers.
  • 2.2 CoT-Kinetics Energy: The method collects a (K, L, d) tensor of reasoning-token states together with output-logit entropy at each inference step.The entropy signal provides internal uncertainty information alongside the hidden-state dynamics.
  • 2.2.1 Pre-processing: Mean pooling across the K reasoning tokens at each layer produces a compact semantic summary of the trajectory.This reduces token-level variance and highlights global reasoning progression across layers while excluding final-answer tokens.
  • 2.2.2 Formulation of CoT-Kinetics Energy: Semantic Momentum Energy τ_i measures displacement between consecutive layers, while Semantic Curvature Energy κ_i captures second-order variation across layers.The paper interprets τ_i as semantic progression and κ_i as semantic correction or dynamic adjustment of reasoning steps.
  • 2.2.2 Formulation of CoT-Kinetics Energy: The CoT-Kinetics equation combines τ_i, κ_i, and query-induced entropy into a scalar score of reasoning soundness.The entropy scaling factor γ aligns entropy magnitude with kinetic-energy components, while τ_i and κ_i are normalized for scale-invariant dynamics.
  • 2.3 Remarks: Higher E_CoT is associated with rapid semantic exploration and lower entropy, which the paper hypothesizes indicates sounder reasoning and greater answer correctness likelihood.The method is training-free, uses internal representational dynamics, and requires no architectural modifications.

3 Experiments

The experiments evaluate whether CoT-Kinetics energy reflects reasoning soundness and answer correctness across diverse models, benchmarks, scales, languages, and aggregation strategies. Results show broad generalization, lower false-positive rates at matched recall, and strongest performance from mean pooling over reasoning tokens.

  • Experiment Setup: The evaluation uses AUROC, AUPR, and FPR@95 to assess how well energy scores connect reasoning soundness with final-answer correctness.The setup compares CoT-Kinetics with Random, softmax, perplexity, entropy, and representation-based baselines across seven LRMs and six benchmarks.
  • Generalization Ability: CoT-Kinetics improves mathematical reasoning scores by +8.5%, +2.5%, and +10.8% across GSM8K and TheoremQA, while remaining strongest for commonsense reasoning across model scales.Table 1 covers reasoning tasks with varying difficulty and domain specificity.
  • Multi-task Reasoning: CoT-Kinetics shows +3.3%, +5.7%, and +2.3% generalization across model scales on MMLU and MMLU-Pro, with AUPR increasing as task difficulty decreases or model scale increases.Reported AUPR values range from 33.39 to 48.57 on MMLU-Pro and from 76.47 to 97.84 on MMLU.
  • ROC Comparison: ROC analysis shows CoT-Kinetics consistently achieves lower FPR at the same TPR than selected baselines, rejecting faulty outputs more effectively while maintaining correct-reasoning identification.Figure 2 compares ROC curves against baselines and uses random scoring as a lower bound.
  • Scalability and Multilingual Applicability: CoT-Kinetics achieves the best or second-best scoring performance on three models exceeding 30B parameters and the best performance across English, German, French, and Chinese MGSM subsets.The multilingual results indicate stronger improvements in some languages, including Chinese, than in English.
  • Ablation Study: Mean pooling over reasoning-token hidden representations is the most effective aggregation strategy for capturing reasoning dynamics in the ablation study.Alternatives include averaging all reasoning tokens per layer, using the last reasoning token, and averaging reasoning and answer tokens together.

4 Conclusion

The conclusion presents CoT-Kinetics energy as a principled framework for intrinsically evaluating reasoning soundness within LRMs and improving assessment of overall answers.

  • Table 6 reports an ablation study of CoT-Kinetics energy components.
  • CoT-Kinetics energy intrinsically evaluates the soundness of reasoning processes within LRMs.
  • The proposed score is described as accurately reflecting the relationship between reasoning trajectories and derived final answers.

5.1 Limitation

The paper identifies limited availability of open-sourced models as a limitation, especially for extremely large-scale LRMs. It calls for broader validation when more such models become available.

  • Open-sourced model availability remains limited, particularly for extremely large-scale LRMs.
  • The authors evaluated CoT-Kinetics across a wide range of state-of-the-art LRMs despite this availability constraint.
  • Future validation is intended to benefit from more high-quality open-sourced LRMs at larger model scales.

5.2 Related Work

The related work distinguishes external and internal approaches to evaluating LRM outputs. The paper positions CoT-Kinetics as an internal framework that independently assesses reasoning soundness.

  • Current LRM output-quality assessment is broadly divided into external and internal methods.
  • External methods use supervision signals such as annotations, knowledge bases, proxy models, or linear probing techniques.
  • Internal methods infer output reliability from hidden representations or uncertainty-based signals without external supervision.
  • Existing internal approaches often assess outputs holistically, allowing final-answer correctness to mask deficiencies in reasoning.
  • CoT-Kinetics extracts and independently assesses the reasoning part, offering a more principled scoring framework for overall output quality.

5.3 Benchmark Dataset Details

The evaluation uses six reasoning benchmarks spanning mathematics, commonsense reasoning, theorem proving, general knowledge, and multilingual mathematics. Dataset specifications and experiment configurations are documented in supplementary tables.

  • The study evaluates CoT-Kinetics across six widely adopted reasoning benchmarks.
  • GSM8K measures mathematical word-problem solving, while TheoremQA evaluates mathematical theorem reasoning and deduction.
  • CommonsenseQA tests commonsense reasoning with five-option multiple-choice questions, and MMLU covers diverse knowledge domains.
  • MMLU-Pro provides more challenging multidisciplinary questions for robust multitask reasoning evaluation.
  • MGSM evaluates arithmetic reasoning across English, German, French, and Chinese using numeric-answer accuracy.
  • Detailed dataset statistics, task specifications, prompting templates, examples, and model configurations are provided in Tables 7–9.

5.4 Model Details

The experiments cover diverse model configurations, varying architectures, scales, layers, hidden dimensions, and special reasoning-phase tokens to assess CoT-Kinetics robustness and generality.

  • 5.4 Model Details: The experiments span model scales and architectures to validate CoT-Kinetics across diverse LLM backbones.The configurations vary in base architectures, number of layers, hidden dimensions, and reserved special tokens.
  • 5.4 Model Details: Reserved <think> and </think> tokens delineate the explicit reasoning phase during inference.

5.5 Detailed Mathematical Derivation of Evaluation Metrics

The paper defines AUROC, AUPR, and FPR@95 to assess how CoT-Kinetics energy scores align with answer correctness. These metrics respectively evaluate ranking, precision-recall trade-offs, and false positives at high recall.

  • 5.5 Detailed Mathematical Derivation of Evaluation Metrics: The evaluation uses AUROC, AUPR, and FPR@95 to assess alignment between energy scores and answer correctness.
  • 5.5.1 Problem Setup and Notations: Each sample is represented by a query, an energy score, and a binary ground-truth correctness label.Correct samples form the positive set P, while incorrect samples form the negative set N.
  • 5.5.2 AUROC: AUROC measures the probability that a randomly chosen correct sample receives a higher energy score than an incorrect sample.Higher AUROC, closer to 1.0, indicates better global ranking capability.
  • 5.5.3 AUPR: AUPR quantifies the precision-recall trade-off, particularly when correct and incorrect cases are imbalanced.It is computed as the area under the precision-recall curve using precision and recall defined at threshold t.
  • 5.5.4 FPR@95: FPR@95 measures the false positive rate at the threshold where recall reaches exactly 95%.Lower FPR@95 indicates fewer false positives while identifying correct reasoning trajectories at high recall.
Loading 2505.13408v1…