Title: SMADE-IE: A Sparse Multi-Agent Framework Utilizing Evidence-Driven Debate for Zero-Shot Information Extraction

Abstract:

The flexibility of large language models (LLMs) in adapting to novel schemas and domains without the need for task-specific training has propelled zero-shot information extraction (IE) into the spotlight. Current methodologies predominantly utilize monolithic prompting, individual-type prompting, or multi-agent debate systems. However, monolithic prompting is frequently prone to errors related to boundaries and types, whereas each-type prompting and multi-agent debates often result in cross-type conflicts, superfluous agent interactions, and significant token consumption.

To mitigate these issues, we introduce SMADE-IE, a novel framework for zero-shot IE that is both sparse and driven by evidence. This approach begins with an Adaptive Mode Selector, which dynamically directs inputs toward either a lightweight Global Extraction Mode or a Type-Centric Extraction Mode. This mechanism effectively minimizes unnecessary type selection processes and reduces reasoning noise. Furthermore, to resolve conflicting predictions, we implement an Evidence-Driven Debate mechanism. This component organizes arguments into Toulmin-style structures and facilitates confidence aggregation by leveraging external evidence scoring alongside Bayesian updates.

Our experimental evaluations across nine benchmark datasets covering NER, RE, and JERE tasks demonstrate that SMADE-IE consistently surpasses existing zero-shot IE baselines. Additionally, the framework enhances token efficiency through its strategy of sparse agent selection and an early-stopping debate protocol.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC