Strengthening Human-Centric Chain-of-Thought Reasoning Integrity in LLMs via a Structured Prompt Framework

Jiling Zhou; Aisvarya Adeseye; Seppo Virtanen; Antti Hakkala; Jouni Isoaho

cs.CRApr 6, 2026

Strengthening Human-Centric Chain-of-Thought Reasoning Integrity in LLMs via a Structured Prompt Framework

Authors: Jiling Zhou, Aisvarya Adeseye, Seppo Virtanen, Antti Hakkala, Jouni Isoaho

Organizations: Department of Computing, University of Turku, 20014 Turku, Finland

Abstract

Chain-of-Thought (CoT) prompting has been used to enhance the reasoning capability of LLMs. However, its reliability in security-sensitive analytical tasks remains insufficiently examined, particularly under structured human evaluation. Alternative approaches, such as model scaling and fine-tuning can be used to help improve performance. These methods are also often costly, computationally intensive, or difficult to audit. In contrast, prompt engineering provides a lightweight, transparent, and controllable mechanism for guiding LLM reasoning. This study proposes a structured prompt engineering framework designed to strengthen CoT reasoning integrity while improving security threat and attack detection reliability in local LLM deployments. The framework includes 16 factors grouped into four core dimensions: (1) Context and Scope Control, (2) Evidence Grounding and Traceability, (3) Reasoning Structure and Cognitive Control, and (4) Security-Specific Analytical Constraints. Rather than optimizing the wording of the prompt heuristically, the framework introduces explicit reasoning controls to mitigate hallucination and prevent reasoning drift, as well as strengthening interpretability in security-sensitive contexts. Using DDoS attack detection in SDN traffic as a case study, multiple model families were evaluated under structured and unstructured prompting conditions. Pareto frontier analysis and ablation experiments demonstrate consistent reasoning improvements (up to 40% in smaller models) and stable accuracy gains across scales. Human evaluation with strong inter-rater agreement (Cohen's k > 0.80) confirms robustness. The results establish structured prompting as an effective and practical approach for reliable and explainable AI-driven cybersecurity analysis.

Figures & tables

Figure 1: Three types of CoT prompts: (1)Free CoT Prompt : The step-by-step reasoning approach with minimal structural constraints emphasizes flexibility; (2)Evidence-Locked CoT Prompt : Incorporating consistency constraints into the reasoning process, requiring conclusions to be based on given information, increasing credibility and reducing hallucinations; (3)Structured Security Reasoning Prompt : The reasoning process follows the actual workflow of security analysis (e.g., threat detection → risk analysis → action recommendation) to ensure the model output is consistent with the decision-making logic of human security experts.

Figure 2: The Prompt Engineering Framework for CoT Reasoning: organizes 16 reasoning-control factors (F1–F16) across 4 dimensions to reduce hallucination, prevent reasoning drift, and strengthen interpretability in security-sensitive tasks; Output Control: System-level factors (S) guide global reasoning, structure, uncertainty calibration, and verification; User-level factors (U) ensure feature grounding, taxonomy alignment, and analytical discipline.

ID	Factor	P	Purpose	Literature Grounding
F1	Role specification (cybersecurity analyst / SOC expert)	S	Defines expert reasoning perspective	Role prompting literature
F2	Explicit task scope constraints	S	Prevents task drift beyond dataset	Prompt constraint research
F3	Dataset grounding (DDoS dataset only)	S	Restricts reasoning to provided data	Hallucination mitigation
F4	Avoid unstated assumptions	S	Reduces unsupported inference	Faithfulness studies
F5	Negative instruction (no inference beyond data)	S	Blocks external knowledge leakage	Prompt safety research
F6	Evidence citation requirement	U	Enforces feature $\rightarrow$ inference mapping	Explainable AI (XAI)

Table 1: Structured Prompt Framework Factors and Prompt Placement

Figure 3: Experimental Methodology and Prompting Workflow. This figure shows the experimental workflow. The accuracy of the detection was directly compared with the dataset, while the reasoning part was manually verified by two researchers under different prompt strategies.

Model	P	Accuracy	Precision	Recall	F1-score
gemma-2B	M	69.8 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.0\%)}}}{\rightarrow}$ 72.6	0.67 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctD}\scriptsize(+6.0\%)}}}{\rightarrow}$ 0.71	0.65 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctG}\scriptsize(+12.3\%)}}}{\rightarrow}$ 0.73	0.66 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctE}\scriptsize(+9.1\%)}}}{\rightarrow}$ 0.72
gemma-2B	C	71.6 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.8\%)}}}{\rightarrow}$ 75.0	0.69 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+5.8\%)}}}{\rightarrow}$ 0.73	0.69 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctF}\scriptsize(+10.1\%)}}}{\rightarrow}$ 0.76	0.69 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctD}\scriptsize(+7.2\%)}}}{\rightarrow}$ 0.74
Llama-3B	M	73.1 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.1\%)}}}{\rightarrow}$ 76.1	0.72 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.2\%)}}}{\rightarrow}$ 0.75	0.69 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctF}\scriptsize(+10.1\%)}}}{\rightarrow}$ 0.76	0.70 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctD}\scriptsize(+7.1\%)}}}{\rightarrow}$ 0.75
Llama-3B	C	75.0 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.1\%)}}}{\rightarrow}$ 78.1	0.74 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+5.4\%)}}}{\rightarrow}$ 0.78	0.72 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctE}\scriptsize(+9.7\%)}}}{\rightarrow}$ 0.79	0.73 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctD}\scriptsize(+6.8\%)}}}{\rightarrow}$ 0.78
Qwen3-4B	M	72.6 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctB}\scriptsize(+3.9\%)}}}{\rightarrow}$ 75.4	0.71 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.2\%)}}}{\rightarrow}$ 0.74	0.68 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctF}\scriptsize(+10.3\%)}}}{\rightarrow}$ 0.75	0.69 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctD}\scriptsize(+7.2\%)}}}{\rightarrow}$ 0.74
Qwen3-4B	C	74.4 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctB}\scriptsize(+3.8\%)}}}{\rightarrow}$ 77.2	0.73 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+4.1\%)}}}{\rightarrow}$ 0.76	0.71 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctE}\scriptsize(+8.5\%)}}}{\rightarrow}$ 0.77	0.72 $\overset{\ignorespaces\text{\hbox{\pagecolor{pctC}\scriptsize(+5.6\%)}}}{\rightarrow}$ 0.76

Table 2: Classification Performance Comparison Across Models and Prompting Strategies (NoFW

\rightarrow

FW). Note: P-Prompt, M-Manual and C-ChatGPT.

Models	P	Evidence	Faithfulness	Structure	Taxonomy
gemma-2b	M	0.72 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+45.8\%)}}}{\rightarrow}$ 1.05	0.70 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+45.7\%)}}}{\rightarrow}$ 1.02	0.76 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+47.4\%)}}}{\rightarrow}$ 1.12	0.68 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+44.1\%)}}}{\rightarrow}$ 0.98
gemma-2b	C	0.86 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+37.2\%)}}}{\rightarrow}$ 1.18	0.84 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+36.9\%)}}}{\rightarrow}$ 1.15	0.91 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+37.4\%)}}}{\rightarrow}$ 1.25	0.82 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+34.1\%)}}}{\rightarrow}$ 1.10
Llama-3B	M	0.80 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+40.0\%)}}}{\rightarrow}$ 1.12	0.78 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+38.5\%)}}}{\rightarrow}$ 1.08	0.84 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+42.9\%)}}}{\rightarrow}$ 1.20	0.76 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+38.2\%)}}}{\rightarrow}$ 1.05
Llama-3B	C	0.94 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+34.0\%)}}}{\rightarrow}$ 1.26	0.92 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+33.7\%)}}}{\rightarrow}$ 1.23	0.98 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+34.7\%)}}}{\rightarrow}$ 1.32	0.90 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+31.1\%)}}}{\rightarrow}$ 1.18
Qwen3-4B	M	0.78 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+38.5\%)}}}{\rightarrow}$ 1.08	0.76 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+36.8\%)}}}{\rightarrow}$ 1.04	0.82 $\overset{\text{\hbox{\pagecolor{gainExtreme}\scriptsize(+43.9\%)}}}{\rightarrow}$ 1.18	0.74 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+37.8\%)}}}{\rightarrow}$ 1.02
Qwen3-4B	C	0.92 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+32.6\%)}}}{\rightarrow}$ 1.22	0.90 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+31.1\%)}}}{\rightarrow}$ 1.18	0.96 $\overset{\text{\hbox{\pagecolor{gainHigh}\scriptsize(+33.3\%)}}}{\rightarrow}$ 1.28	0.88 $\overset{\text{\hbox{\pagecolor{gainModerate}\scriptsize(+29.5\%)}}}{\rightarrow}$ 1.14

Table 3: Human-Evaluated Reasoning Quality Scores Across Models and Prompting Strategies (NoFW

\rightarrow

FW). Note: P-Prompt, M-Manual and C-ChatGPT.

Figure 4: Model Size vs. Relative Performance Gain (Improvement) (FW vs. NoFW) in Detection Accuracy and Reasoning Quality

Figure 5: Pareto Frontier Comparison of Security Detection Accuracy and Human-Evaluated Reasoning Dimensions (Evidence, Faithfulness, Structure, Taxonomy) Across Large Language Models with and without the Structured Prompt Framework

Explore similar work

Reasoning Topology Matters: A Controlled Study of LLM-Based Cybersecurity Analysis
Sep 21, 2026Jiling Zhou, Aisvarya Adeseye, Antti Hakkala +2LLMs for Cybersecurity LLM Reasoning with Graphs
Department of Computing University of Turku Turku, Finland
CAP-CoT: Cycle Adversarial Prompt for Improving Chain of Thoughts in LLM Reasoning
Apr 25, 2026Shuxu Chen, Yitian Zhou, Jiaquan Zhang +6Adversarial Prompt Generation LLM Prompting
Department of Electronics and Information Convergence Engineering, Kyung Hee University, Republic of Korea · University of Electronic Science and Technology of China, Chengdu, China · East China Jiaotong University, Nanchang, China +1
Structured but Fragile: On the Limits of LLMs in Cybersecurity Decision-Making
Aug 21, 2026Pasquale Malacaria, Yunxiao ZhangLLMs for Cybersecurity LLM Decision-Making
School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK. · Department of Computer Science, University of Exeter, Exeter, UK.

Sep 21, 2026J/K move · Enter open · S save

Reasoning Topology Matters: A Controlled Study of LLM-Based Cybersecurity Analysis

Jiling Zhou, Aisvarya Adeseye, Antti Hakkala, Seppo Virtanen+1

Department of Computing University of Turku Turku, Finland

LLMs for Cybersecurity LLM Reasoning with Graphs Cybersecurity

Open paper AskPDF arXiv

Large Language Models (LLMs) are increasingly used in cybersecurity, where accurate analysis often requires multi-step and context-dependent reasoning over complex and heterogeneous data. However, existing prompting approaches typically focus on eliciting reasoning without explicitly considering how intermediate reasoning steps are structurally organized. We introduce Security Reasoning Topology, which models reasoning through three representative structures: Linear, Branching, and Graph. To evaluate their effects, we conduct controlled experiments on three cybersecurity datasets covering MITRE ATT&CK network traffic, cyber threat intelligence (CTI), and CVE vulnerability analysis. We evaluate multiple LLMs, including Llama 2 (7B, 13B, 70B), GPT-5.1, and Mistral Large 3, while keeping task inputs consistent and controlling reasoning structure through system-level prompting. Results show that reasoning topology substantially affects performance: Graph reasoning achieves the highest overall accuracy, improving over few-shot prompting by 9.8-12.2 percentage points across datasets, while Branching provides a strong intermediate solution. The results further show that the effect of reasoning topology remains consistent across model families and scales, highlighting reasoning topology as an important design factor for LLM-based cybersecurity analysis.

Abstract

Explore similar work

Reasoning Topology Matters: A Controlled Study of LLM-Based Cybersecurity Analysis

CAP-CoT: Cycle Adversarial Prompt for Improving Chain of Thoughts in LLM Reasoning

Structured but Fragile: On the Limits of LLMs in Cybersecurity Decision-Making