Structured Protocols Achieve Universal Hallucination Suppression Across Commercial LLM Families

Read full text Summarise with VIRAT
Authors: Blake Musselman
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We present a cross-family empirical study of hallucination at competence boundaries in large language models. Using 10 computationally incoherent formal operations as stimuli, we test 10 models across 4 architectural families (Anthropic Claude, OpenAI GPT, Google Gemini/Gemma, Meta Llama) under three intervention conditions. We find that blind hallucination is universal (93-100% on frontier models), that a single abstention token shows inverse scaling on Claude (more capable models comply less), and that a structured formal protocol (AOP) reverses this scaling and achieves 90-100% compliance across model families. A simplified protocol variant (AOP-Lite) addresses a capacity threshold effect on smaller models, informed by a Gemini model's self-diagnosis of its own failure mode.

Fields of Study & Subjects

Topic Modeling Natural Language Processing Techniques Adversarial Robustness in Machine Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
26 March 2026
Language
EN
Document type
Preprint