HARP: Hierarchical Adaptive Ranking with Preference-Adaptive Fusion for Query-Based CVE Prioritization

Haochen Liu; Zhengzhang Chen; Haoyu Wang; Yanchi Liu; Jundong Li; Haifeng Chen

cs.IRAug 19, 2026

HARP: Hierarchical Adaptive Ranking with Preference-Adaptive Fusion for Query-Based CVE Prioritization

Authors: Haochen Liu, Zhengzhang Chen, Haoyu Wang, Yanchi Liu, Jundong Li, Haifeng Chen

Organizations: University of Virginia · NEC Laboratories America

Abstract

Vulnerability prioritization is inherently preference dependent, since the same CVE can receive different remediation priority under different operational preference scenarios. Existing scoring systems and ranking methods typically assume a fixed criterion. In practice, organizations already operate under a preference scenario, but this preference is often implicit and difficult to express as a written prompt instruction, while triage queries usually do not encode it. Past validated triage cases under the current scenario are more readily available. We study query-based CVE prioritization in this setting and propose HARP, a graph-grounded multi-view framework that ranks candidates from a natural-language query together with a support bank of historical labeled examples from the current preference scenario, without requiring an explicit textual summary of that scenario. HARP retrieves evidence from a vulnerability knowledge graph, scores candidates with policy-conditioned global, enterprise, and user views, and fits view-fusion weights from sampled supports. Experiments across three preference scenarios and multiple backbone LLMs show that HARP outperforms multiple baselines, expressing our method's effectiveness.

Figures & tables

Figure 1: Motivating example of preference-dependent CVE prioritization. The same query and candidate CVEs can require different rankings under different preference scenarios. Fixed scoring or single-prompt ranking produces one ordering and may fail to match the current scenario’s remediation preference.

Figure 2: Overview of the HARP framework for query-based CVE prioritization. Given a natural-language query and a support bank of past labeled examples from the current preference scenario, HARP retrieves candidate CVEs and evidence from a vulnerability knowledge graph, scores candidates through policy-conditioned global, enterprise, and user views, adaptively fuses the view-specific scores with support-bank guidance, and produces the final ranked CVE list with multi-view regularization and Top- $K$ consensus routing.

Setting	Method	Importance			Exploit_First			Patch_Order
Setting	Method	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k
Non-LLM	Industrial Signals	0.4945	0.3594	0.3304	0.5057	0.3672	0.3385	0.4966	0.3621	0.3327
Non-LLM	BM25	0.4882	0.3783	0.3319	0.4566	0.3487	0.3066	0.4755	0.3703	0.3277
GPT-OSS-20B	HARP	0.5685	0.4598	0.4194	0.5758	0.4534	0.4123	0.5921	0.4833	0.4414
	Global	0.4700	0.4208	0.3541	0.4422	0.3946	0.3228	0.4634	0.4168	0.3495
	Enterprise	0.4852	0.4357	0.3709	0.4730	0.4167	0.3505	0.5117	0.4522	0.3876

Table 1: Main ranking performance across non-LLM baselines, backbone LLMs, and preference scenarios. We report Precision@k, MAP@k, and MRR@k averaged over five runs. The best result for each metric within each backbone setting is highlighted in bold . HARP standard deviations are in Appendix E .

Backbone LLM	Method	Importance			Exploit_First			Patch_Order
Backbone LLM	Method	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k
GPT-OSS-20B	HARP	0.5685	0.4598	0.4194	0.5758	0.4534	0.4123	0.5921	0.4833	0.4414
	Mixed Support Bank	0.5340	0.4395	0.3985	0.5175	0.4180	0.3765	0.5585	0.4580	0.4175
	No Regularization	0.5601	0.4526	0.4128	0.5204	0.4051	0.3706	0.5605	0.4431	0.4076
	Equal Weight	0.5317	0.4380	0.3974	0.5192	0.4197	0.3776	0.5608	0.4604	0.4190
	Random Weight	0.4056	0.3109	0.2926	0.4115	0.3204	0.2963	0.4106	0.3071	0.2881

Table 2: Ablation study across backbone LLMs and preference scenarios. Mixed Support Bank pools supports from all preference scenarios without using the evaluated scenario identity. We report Precision@k, MAP@k, and MRR@k. The best result for each metric within each setting is highlighted in bold .

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Importance			Exploit_First			Patch_Order
Backbone	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k	P@k	MAP@k	MRR@k
GPT-OSS-20B	0.0082	0.0068	0.0061	0.0084	0.0069	0.0062	0.0079	0.0068	0.0062
Llama3-8B	0.0054	0.0058	0.0056	0.0055	0.0060	0.0057	0.0053	0.0059	0.0056
Qwen2.5-7B-Instruct	0.0038	0.0042	0.0040	0.0034	0.0039	0.0039	0.0030	0.0028	0.0036

Appendix

Table 3: Standard deviation of HARP over five independent runs.

Feature	Contribution Rule
exploit	$1.0$ if KNOWN_EXPLOITED ; $0.6$ if has_exploit is true; otherwise $0.0$
severity	If base_score exists, $\min(1,\texttt{base\_score}/10)$ ; otherwise use severity mapping
patch	$1.0$ if has_patch_in_references is true; otherwise $0.0$
attack_vector	NETWORK: $1.0$ ; ADJACENT: $0.5$ ; LOCAL: $0.2$ ; PHYSICAL: $0.1$ ; otherwise $0.0$
recency	$\max(0,1-\texttt{days\_since\_published}/365)$
version_info	$0.5\cdot\texttt{has\_version\_info}+0.5\cdot\texttt{affects\_legacy\_versions}$

Appendix

Table 4: Feature contribution rules used to guide the construction of policy-specific training supervision.

Policy	exploit	severity	patch	attack	recency	version	reference	cwe	platform	scope
global_balanced	2.5	2.0	0.5	0.8	0.5	0.3	0.3	0.2	0.1	0.2
global_exploit_first	3.5	1.0	0.3	0.5	0.6	0.2	0.2	0.1	0.0	0.2
global_severity_first	2.0	2.5	1.0	0.5	0.3	0.5	0.2	0.2	0.1	0.2
global_recent	1.5	1.0	0.5	0.3	2.5	0.3	0.5	0.2	0.1	0.2
global_version_aware	1.5	1.5	1.0	0.4	0.3	2.0	0.3	0.2	0.2	0.2
enterprise_patch	2.0	1.0	2.0	0.3	0.2	0.6	0.4	0.1	0.1	0.2

Appendix

Table 5: Policy weights of the three views global , enterprise and user .

Figure 3: Sensitivity analysis of the interpolation coefficient

\alpha

on Qwen2.5-7B-Instruct. We report Precision@K, MAP@K, and MRR@K under the importance, exploit-first, and patch-order preference scenarios. The dashed vertical line indicates the value used in the main experiments,

\alpha=0.3

Cybersecurity

Cybersecurity develops methods to protect digital and cyber-physical systems from attacks, vulnerabilities, data loss, and misuse.

See all 354 papers

Bridging the Gap Between Security Metrics and Key Risk Indicators: An Empirical Framework for Vulnerability Prioritization
Mar 12, 2026Emad Sherif, Iryna Yevseyeva, Vitor Basto-Fernandes +1Cybersecurity
Faculty of Technology, Arts and Culture, De Montfort University, Leicester, United Kingdom · Instituto Universitário De Lisboa (ISCTE-IUL), University Institute of Lisbon, ISTAR-IUL, Lisboa, Portugal · Sorbonne Université, CNRS, LIP6, 75005 Paris, France
Predicting CVSS and CWE from CVE Descriptions with Fine-Tuned Small Language Models
Oct 16, 2026Tara Dixit, Jessie Dong, Summer LiLLMs for Cybersecurity
Stanford University
Athena: Vulnerability-Affected Library Identification via Knowledge Graph Completion
Sep 1, 2026Phong Trinh Duy, Trang Dang Yen, Hung Nguyen-Huu +5Software Security KG Completion
Hanoi University of Science and Technology, Vietnam · The University of Sydney, Australia · Singapore Management University, Singapore +3

Mar 12, 2026J/K move · Enter open · S save

Bridging the Gap Between Security Metrics and Key Risk Indicators: An Empirical Framework for Vulnerability Prioritization

Emad Sherif, Iryna Yevseyeva, Vitor Basto-Fernandes, Allan Cook

Faculty of Technology, Arts and Culture, De Montfort University, Leicester, United Kingdom · Instituto Universitário De Lisboa (ISCTE-IUL), University Institute of Lisbon, ISTAR-IUL, Lisboa, Portugal · Sorbonne Université, CNRS, LIP6, 75005 Paris, France

Cybersecurity

Open paper AskPDF arXiv

Organisations overwhelmingly prioritize vulnerability remediation using Common Vulnerability Scoring System (CVSS) severity scores, yet CVSS classifiers achieve an Area Under the Precision-Recall Curve (AUPRC) of 0.011 on real-world exploitation data, near random chance. We propose a composite Key Risk Indicator grounded in expected-loss decomposition, integrating dimensions of threat, impact, and exposure. We evaluated the KRI framework against the Known Exploited Vulnerabilities (KEV) catalog using a comprehensive dataset of 280,694 Common Vulnerabilities and Exposures (CVEs). KRI achieves Receiver Operating Characteristic Area Under the Curve (ROC-AUC) 0.927 and AUPRC 0.223 versus 0.747 and 0.011 for CVSS (24 percents, 20). Ablation analysis shows Exploit Prediction Scoring System (EPSS) alone achieves AUPRC 0.365, higher than full KRI (0.223), confirming that EPSS and KRI serve distinct objectives: EPSS maximizes raw exploit detection, while KRI re-orders by impact and exposure, capturing 92.3 percents of impact-weighted remediation value at k=500 versus 82.6 percents for EPSS, and surfacing 1.75 more Critical-severity exploited CVEs. KRI's net benefit exceeds EPSS whenever the severity premium exceeds 2. While EPSS serves as a robust baseline for exploit detection, the KRI framework is the superior choice for organizations seeking to align remediation efforts with tangible risk reduction.

Abstract

Cybersecurity

Explore similar work

Bridging the Gap Between Security Metrics and Key Risk Indicators: An Empirical Framework for Vulnerability Prioritization

Predicting CVSS and CWE from CVE Descriptions with Fine-Tuned Small Language Models

Athena: Vulnerability-Affected Library Identification via Knowledge Graph Completion