MultiPress: A Multi-Agent Framework for Interpretable Multimodal News Classification
Authors: Tailong Luo, Hao Li, Rong Fu, Xinyue Jiang, Huaxuan Ding, Yiduo Zhang, Zilin Zhao, Simon Fong, +2 more
Organizations: New York Institute of Technology · University of Arizona · University of Macau · Peking University · Chang’an University · Juniata College
With the growing prevalence of multimodal news content, effective news topic classification demands models capable of jointly understanding and reasoning over heterogeneous data such as text and images. Existing methods often process modalities independently or employ simplistic fusion strategies, limiting their ability to capture complex cross-modal interactions and leverage external knowledge. To overcome these limitations, we propose MultiPress, a novel three-stage multi-agent framework for multimodal news classification. MultiPress integrates specialized agents for multimodal perception, retrieval-augmented reasoning, and gated fusion scoring, followed by a reward-driven iterative optimization mechanism. We validate MultiPress on a newly constructed large-scale multimodal news dataset, demonstrating significant improvements over strong baselines and highlighting the effectiveness of modular multi-agent collaboration and retrieval-augmented reasoning in enhancing classification accuracy and interpretability.
Figures & tables
Fig. 1: Overview of our MultiPress Framework.
Model
Acc.
MP
MR
MF1
GLM-4V-9B
70.5
69.8
69.2
69.5
mPLUG-7B
68.9
68.1
67.5
67.8
Qwen2.5-VL-3B (zero-shot)
62.3
61.0
60.5
60.7
Qwen2.5-VL-3B (fine-tuned)
66.8
66.0
65.5
65.7
Qwen2.5-VL-7B (zero-shot)
77.4
76.8
76.1
76.4
Qwen2.5-VL-7B (fine-tuned)
80.2
79.5
79.0
79.2
TABLE I: Performance comparison on NewsMM test set.
Variant
Acc.
MF1
Full MultiPress (GPT-4o)
91.2
90.4
w/o Retrieval-Augmented Generation
86.5
85.7
w/o Online Reasoning Agent
87.1
86.3
w/o Gated Fusion
84.9
84.1
w/o Reward-Driven Optimization
88.3
87.5
Text-only input
79.4
78.6
TABLE II: Ablation study on NewsMM test set.
Model
Faithfulness
Groundedness
Usefulness
Qwen2.5-VL-7B
3.42
3.15
3.38
GPT-4o-20250513
4.10
3.85
4.45
MultiPress (GPT-4o)
4.65
4.72
4.58
TABLE III: Interpretability evaluation on 100 sampled test instances (higher is better).
Method
#Calls
Tokens
Latency
GPT-4o (direct)
1.0
1,450
1.8s
MultiPress (GPT-4o)
6.2
5,820
8.5s
TABLE IV: Efficiency comparison averaged per sample.