Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TriFusion-LLM: Prior-Guided Multimodal Fusion with LLM Arbitration for Fine-grained Code Clone Detection

Mengdi Li, Yuming Liu, He Wang, Zifeng Xu, Yuqing Zhang

Abstract

Code clone detection (CCD) supports software maintenance, refactoring, and security analysis. Although pre-trained models capture code semantics, most work reduces CCD to binary classification, overlooking the heterogeneity of clone types and the seven fine-grained categories in BigCloneBench. We present Full Model, a multimodal fusion framework that jointly integrates heuristic similarity priors from classical machine learning, structural signals from abstract syntax trees (ASTs), and deep semantic embeddings from CodeBERT into a single predictor. By fusing structural, statistical, and semantic representations, Full Model improves discrimination among fine-grained clone types while keeping inference cost practical. On the seven-class BigCloneBench benchmark, Full Model raises Macro-F1 from 0.695 to 0.875. Ablation studies show that using the primary model's probability distribution as a prior to guide selective arbitration by a large language model (LLM) substantially outperforms blind reclassification; arbitrating only ~0.2% of high-uncertainty samples yields an additional 0.3 absolute Macro-F1 gain. Overall, Full Model achieves an effective performance-cost trade-off for fine-grained CCD and offers a practical solution for large-scale industrial deployment.

TriFusion-LLM: Prior-Guided Multimodal Fusion with LLM Arbitration for Fine-grained Code Clone Detection

Abstract

Code clone detection (CCD) supports software maintenance, refactoring, and security analysis. Although pre-trained models capture code semantics, most work reduces CCD to binary classification, overlooking the heterogeneity of clone types and the seven fine-grained categories in BigCloneBench. We present Full Model, a multimodal fusion framework that jointly integrates heuristic similarity priors from classical machine learning, structural signals from abstract syntax trees (ASTs), and deep semantic embeddings from CodeBERT into a single predictor. By fusing structural, statistical, and semantic representations, Full Model improves discrimination among fine-grained clone types while keeping inference cost practical. On the seven-class BigCloneBench benchmark, Full Model raises Macro-F1 from 0.695 to 0.875. Ablation studies show that using the primary model's probability distribution as a prior to guide selective arbitration by a large language model (LLM) substantially outperforms blind reclassification; arbitrating only ~0.2% of high-uncertainty samples yields an additional 0.3 absolute Macro-F1 gain. Overall, Full Model achieves an effective performance-cost trade-off for fine-grained CCD and offers a practical solution for large-scale industrial deployment.
Paper Structure (33 sections, 2 equations, 5 figures, 7 tables)

This paper contains 33 sections, 2 equations, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Taxonomy and Benchmarking of Code Clones
  4. Conventional and Augmented Code Clone Detection
  5. Code Representation Learning and Semantic Analysis
  6. Hybrid Representations and Multi-modal Learning
  7. LLM-driven Code Clone Detection
  8. Methodology
  9. Task Definition and Rationale for Fine-grained 7-class Classification
  10. Dataset Construction and Preprocessing
  11. Data Cleansing and Metadata Indexing
  12. Project-level Isolation Protocol
  13. Goal-oriented Stratified Sampling
  14. Diversity Optimization for Type-4 Semantic Clones
  15. Heterogeneous Feature Space Extraction
  16. ...and 18 more sections

Figures (5)

  • Figure 1: Overall pipeline of the proposed multi-dimensional clone detection framework.
  • Figure 2: Data Processing Pipeline of the Proposed Multi‑Dimensional Clone Detection Framework
  • Figure 3: Confusion matrix analysis on the test set. (a) Normalized confusion matrix of the proposed model. (b) Difference matrix illustrating performance improvements over the baseline.
  • Figure 4: Performance behavior across confidence intervals on the validation set. The left figure shows prediction accuracy across bins, while the right figure illustrates the Macro-F1 variation used to determine the arbitration threshold.
  • Figure 5: Performance behavior across confidence intervals on the validation set. The left figure shows prediction accuracy across bins, while the right figure illustrates the Macro-F1 variation used to determine the arbitration threshold.