Volume 16, Number 6
S-AI-Anti Hallucination: A Bio-Inspired and Confidence-aware Sparse AI Framework for Reliable Generative Systems
Authors
Said Slaoui, Mohammed V University, Morocco
Abstract
Large Language Models (LLMs) exhibit impressive generative capabilities but remain prone to hallucinations plausible yet false statements produced with high confidence. Such phenomena undermine trust and reliability in sensitive domains including health, law, and cybersecurity. Despite significant progress in retrieval-augmented generation (RAG), calibration methods, and mixture-of-experts (MoE) architectures, existing systems still lack a unified framework for veridiction, abstention, and energy-aware reasoning. This work introduces S-AI Against Hallucinations, a bio-inspired and parsimonious architecture derived from the Sparse Artificial Intelligence (S-AI) framework. The proposed model implements a symbolic-hormonal orchestration mechanism that enables generative agents to detect uncertainty, abstain when appropriate, and maintain citation integrity under ambiguous or adversarial conditions. The system relies on four hormonal variables — Hallucination Uncertainty (HU), Citation Integrity (CI), Contradiction Observer (CO), and Retrieval Evidence (RE) — dynamically regulated by a MetaAgent through hysteresis-based thresholds. Experiments performed on diverse scenarios, including factual question answering, scientific summarization, numerical reasoning, and out-of-distribution prompts, demonstrate stable abstention behavior, consistent citation tracking, and adaptive evidence retrieval. Evaluation follows a transparent and reproducible protocol inspired by PRISMA standards and Scopus-indexed benchmarking practices. S-AI Against Hallucinations provides a coherent, confidence-aware foundation for explainable and resource-efficient generative intelligence. It establishes a conceptual and operational bridge between statistical learning, symbolic reasoning, and biological homeostasis, paving the way for reliable and ethically governed AI systems.
Keywords
Sparse Artificial Intelligence (S-AI), Hormonal MetaAgent, Symbolic-Hormonal Orchestration, Hallucination Mitigation, Confidence-Aware Reasoning, Explainable and Frugal AI, Triadic Memory System, Governed Parsimony, Retrieval-Augmented Veridiction.