Running summarizing tools on a new article
Check the paper below, then start its synthesis.
MORPHEUS v3.0 A Neuro-Symbolic Multi-Agent Architecture for Coherence Auditing and Hallucination Mitigation in Large Language Models
Belkacem Mansouri (École Nationale Polytechnique d'Oran)
Abstract: AbstractLarge Language Models (LLMs) generate hallucinations that undermine their reliabilityin critical domains. Existing mitigation strategies rely on costly post-generation verificationor LLM-as-a-judge approaches introducing significant latency. We previously introducedMORPHEUS v2.0, a deterministic pre-generation protocol based on three O(N )-complexityalgorithms, achieving 80% accuracy at 45 ms latency. However, this purely symbolic ap-proach exhibited a 20% failure rate on technical confabulations, attributable to the rigidityof the static knowledge base.This paper presents MORPHEUS v3.0, a neuro-symbolic evolution that transcendsthe algorithmic determinism of v2.0 through the integration of a dynamic RAG pipeline andthe orchestration of three independent cognitive judges: Judge Alpha (contextual support),Judge Beta (factual knowledge augmented by RAG), and Judge Gamma (formal logical con-sistency). Each judge operates in isolation, producing a traceable partial verdict aggregatedinto an auditable final decision.We demonstrate that this neuro-symbolic hybridization—where formal logic constrainsneural reasoning—resolves technical confabulation cases while preserving the computationalparsimony and epistemic traceability that characterize the MORPHEUS DNA. The protocolis open-source (MIT license) and designed for native integration into RAG pipelines andeducational systems.Keywords: Neuro-symbolic AI, Multi-agent systems, Anti-hallucination, RAG, Determin-istic verification, Epistemic traceability, Large Language Models, MORPHEUS