Diverse Stacking Ensemble for Attributing LLM Outputs via Relational Reasoning

Attributing the origin of outputs generated by large language models (LLMs) is critical for ensuring transparency and accountability in AI systems. This paper presents LLM-OriginNet, a two-stage ensemble learning framework designed to identify the source model of a given response. In the first stage, multiple fine-tuned classifiers based on diverse LLM architectures are employed to capture model-specific generation patterns. The second stage introduces relational feature engineering, which extracts comparative signals from the outputs of these classifiers. A LightGBM-based meta-classifier then integrates these relational features to produce the final attribution decision. By leveraging architectural diversity and inter-model relational reasoning, LLM-OriginNet provides a scalable and interpretable solution for LLM attribution in multi-model environments. Furthermore, the framework can incorporate differential privacy mechanisms and support federated learning, enabling attribution on decentralized or noisy feature representations while preserving user privacy and data confidentiality.