MAF-IE: Multi-Agent Finetuning for Zero-shot Information Extraction

ACL ARR 2025 May Submission2095 Authors

Abstract:

Large language models (LLMs) excel at text generation and reasoning but struggle to produce structured output while maintaining accuracy in zero-shot information extraction (IE).
Recent studies have explored multi-agent frameworks to enhance LLMs' capabilities. Still, these efforts primarily target general reasoning and fail to address key structured IE challenges such as boundary ambiguity and cross-type semantic conflicts.
In this work, we propose MAF-IE, a multi-agent finetuning framework that combines specialization and collaborative training to improve both the accuracy and efficiency of multi-agent systems for IE.
Specifically, we introduce a type-specified multi-agent collaboration framework to generate high-quality pseudo-labeled data.
Based on the generated data, we design a novel contrastive data selection strategy to finetune multiple LLMs on dialogue trajectories, enabling the model to better learn from both correct and incorrect predictions, enhancing task-specific feature learning. Combined with a simple majority voting strategy, the finetuned models achieve comparable performance to multi-agent large language models (LLMs) while significantly reducing inference costs.
Extensive experiments on seven datasets across six tasks, spanning coarse- and fine-grained settings at both sentence and document levels, demonstrate MAF-IE significantly outperforms zero-shot IE baselines.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: named entity recognition and relation extraction, event extraction, zero/few-shot extraction

Contribution Types: Model analysis & interpretability, Data analysis

Languages Studied: English

Submission Number: 2095