Bridging the Gap: A Neuro-Symbolic Evaluation Framework for LLM-Driven Consensus Building
Year: 2026
Type: conference-paper
Abstract: The growing adoption of Large Language Models (LLMs) in decision-making processes is radically reshaping the practices of digital democracy. However, despite the technology's promise to scale deliberative procedures, an exclusive reliance on semantic evaluation metrics creates serious methodological gaps, as it tends to obscure “hallucination” phenomena and algorithmic instability. To address this issue, this paper introduces a hybrid Neuro-Symbolic Evaluation Framework, experimentally applied to the prototype “Dimokratia” platform with the participation of 30 students and three models (ChatGPT-4, Mistral Large 2, AI21 Jamba). The analysis revealed a systemic “Consensus Hallucination”: while the models achieved high semantic similarity, they did so at the expense of users’ terminology; at the same time, they adopted verbosity strategies (≈60% increase) and exhibited high instability (drift > 0.70) during feedback. In light of these findings, the study concludes that evaluation strategies are needed that ensure not only user satisfaction, but above all the faithfulness and stability of digital intermediaries.
Institution: Hellenic Open University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7223057836