EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Multi-Granularity Distillation of Intervention Sensitivity Representations for Multimodal Large Language Models

Zhiguo Zhang, Wei Sun

Sensors · 04.11.2026 · 26 (21) · 6963

DOI: 10.3390/s26216963 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Sensors, online erschienen am 04.11.2026. Verfasst von Zhiguo Zhang und 1 weiteren Autor:innen. Beteiligte Einrichtung: School of Aerospace Science and Technology, Xidian University, Xi'an 710071, China. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Shaanxi “Hundred Teams” Project for University-Enterprise Collaboration 2025, Xi’an Science and Technology Plan Project.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Existing knowledge distillation methods for multimodal large language models (MLLMs) mainly align output distributions or intermediate representations, but rarely supervise how predictions depend on local visual evidence. This limitation may hinder the transfer of fine-grained perception and compositional reasoning. We propose Causal and Multi-source Visual Attribution Distillation (CMVAD), which transfers teacher-derived visual evidence during distillation. Region-level interventions first estimate token-level causal visual attributions, identifying image regions that influence each generated token. The Causal Evidence Adapter Fetch (CEAF) module jointly models token importance and semantic interactions to adaptively aggregate multi-source causal evidence into a unified representation. A Multi-granularity Attribution Distillation (MAD) objective transfers this representation via global evidence distribution alignment, critical-region alignment, and evidence-ranking consistency. Experiments on general visual question answering and compositional reasoning benchmarks show that CMVAD outperforms existing MLLM distillation methods, achieving larger gains in compositional reasoning while retaining competitive visual question answering performance. Results across teacher–student architectures demonstrate its generalizability. We further evaluate VisDrone UAV object detection through overall comparisons, providing supplementary evidence of CMVAD’s applicability to object recognition and localization in UAV imagery.

Schlagwörter

Stichwörter der Autor:innen

UAV object detectioncausal attributionknowledge distillationmodel compressionmultimodal large language modelsvision-language models

Autor:innen und Institutionen

  • School of Aerospace Science and Technology, Xidian University, Xi'an 710071, China.

Förderung

  • National Natural Science Foundation of China
  • Shaanxi “Hundred Teams” Project for University-Enterprise Collaboration 2025
  • Xi’an Science and Technology Plan Project

Zitieren

APA

Zhang, Z., & Sun, W. (2026). Multi-Granularity Distillation of Intervention Sensitivity Representations for Multimodal Large Language Models. Sensors, 26(21), 6963. https://doi.org/10.3390/s26216963

Vancouver

Zhang Z, Sun W. Multi-Granularity Distillation of Intervention Sensitivity Representations for Multimodal Large Language Models. Sensors. 2026;26(21):6963. doi:10.3390/s26216963

BibTeX
@article{Zhang2026,
  author = {Zhang, Zhiguo and Sun, Wei},
  title = {Multi-Granularity Distillation of Intervention Sensitivity Representations for Multimodal Large Language Models},
  journal = {Sensors},
  year = {2026},
  volume = {26},
  number = {21},
  pages = {6963},
  doi = {10.3390/s26216963}
}
Kostenlos starten →