When Small Models Team Up: A Weak‑Expert Ensemble Surpassing LLMs for Automated Intellectual‑Property Audits

Xinji Mai, Haobo Chen, Mingze Wang, Sixu Li, Ziyao Xu, HonghaoWang, Ziheng Wu, Tao Wang

Keywords: LLM, VLM, Intellectual Property Rights, Small Models Team Up

Abstract: Intellectual Property Rights (IPR) enforcement on e-commerce platforms is increasingly challenged by the scale and complexity of modern online marketplaces, where counterfeit goods and brand infringements evolve rapidly to evade detection. Existing approaches rely heavily on human review or unimodal AI models, limiting their scalability and adaptability. To bridge this gap, we propose IPR-GPT, a novel multimodal framework that redefines automated IPR auditing by integrating a Multi-Agent Audit Framework and an Audit Reversal Data Augmentation mechanism. MAAF leverages a structured multi-agent collaboration strategy, combining specialized experts in visual analysis, textual reasoning, legal compliance, and exemption handling. By systematically exploring the trade-offs between Vision-Language Models and lightweight vision models, MAAF achieves a performance boost of up to 24.26% in key IPR audit tasks. AR further enhances model robustness by synthesizing hard-to-classify audit cases, improving generalization in real-world scenarios. Extensive experiments demonstrate that IPR-GPT consistently outperforms state-of-the-art models, setting a new benchmark for multimodal IPR enforcement. Our results challenge the prevailing belief that larger multimodal LLMs are always superior, showing instead that a purpose‑built ensemble of weak experts can deliver both higher accuracy and lower cost.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 7637