Decision-Rule-Controlled Multi-Agent Validation for Hungarian Banking Complaint Classification

Authors

  • Zsolt Krutilla Obuda University; University of Dunaújváros

DOI:

https://doi.org/10.32968/psaie.2026.1.1.X

Keywords:

Multi-Agent LLM, Decision-Controlled Validation, Banking NLP

Abstract

This paper presents a decision-rule-controlled multi-agent validation framework for Hungarian banking complaint classification. A previously trained DistilBERT model provides the initial category prediction, while Qwen2.5-7B, Qwen3-8B, and Mistral-7B are used as reasoning and validation agents. A rule-based decision agent determines whether the baseline label is accepted, revised, or routed to exception handling. Experiments were conducted on a 31-category random sample and a reduced Top6 + Exception balanced sample. The results show that unrestricted LLM-based relabeling may significantly degrade performance, while conservative decision rules better preserve baseline stability. The findings indicate that LLMs are more suitable as controlled validation components than as unrestricted classifiers in closed-label banking NLP tasks.

Author Biography

Zsolt Krutilla, Obuda University; University of Dunaújváros

ORCID: 0009-0000-3607-4309

Downloads

Published

2026-11-17