arXiv:2606.04612v1 Announce Type: new 
Abstract: Large Language Models (LLMs) are vulnerable both to hallucination and adversarial manipulation. Although these problems are closely related, existing defences typically address them separately. We investigate a hybrid defence framework that combines entropy-based models, designed to reduce hallucinations, with uncertainty-based models and geometric-based models, designed to reduce vulnerability. Under in-domain tests on Natural Language Understanding datasets (FEVER, HotpotQA, CSQA, SIQA) we find our hybrid model improves both clean-task performance (up to 43.34\% increase in accuracy) and adversarial robustness (up to 64.92\% improvement in accuracy and 62.27\% reduction in attack success rate). For out-of-distribution datasets (AeroEngQA, CPIQA) we see similar adversarial robustness from our hybrid model (up to 57.14\% improvement in accuracy). For prompt injection (SafeGuard) and jailbreak detection (AdvBench, DAN) datasets our hybrid model is also very strong (up to 51\% reduction in attack success rate compared to state of the art baseline models). Overall, our results show that combining entropy, uncertainty and geometric features provides a more effective defence strategy than using any single feature alone for both in-domain and out-of-distribution tasks.

تم تطوير إطار دفاع هجين جديد لنماذج اللغة الكبيرة (LLMs) لمعالجة الثغرات المتعلقة بالهلوسة والتلاعب العدائي. يجمع هذا الإطار بين نماذج قائمة على الإنتروبيا، ونماذج قائمة على عدم اليقين، ونماذج هندسية، مما يؤدي إلى تحسينات كبيرة في الدقة والصلابة عبر مجموعات بيانات مختلفة لفهم اللغة الطبيعية، بما في ذلك FEVER وHotpotQA.

Se ha desarrollado un nuevo marco de defensa híbrido para los Modelos de Lenguaje Grande (LLMs) para abordar las vulnerabilidades relacionadas con la alucinación y la manipulación adversarial. Este marco combina modelos basados en la entropía, la incertidumbre y la geometría, lo que resulta en mejoras significativas en la precisión y la robustez en varios conjuntos de datos de comprensión del lenguaje natural, incluidos FEVER y HotpotQA.

Un nouveau cadre de défense hybride a été développé pour les modèles de langage de grande taille (LLMs) afin de traiter les vulnérabilités liées à l'hallucination et à la manipulation adversariale. Ce cadre combine des modèles basés sur l'entropie, l'incertitude et la géométrie, entraînant des améliorations significatives de la précision et de la robustesse sur divers ensembles de données de compréhension du langage naturel, y compris FEVER et HotpotQA.

A new hybrid defense framework has been developed for Large Language Models (LLMs) to address vulnerabilities related to hallucination and adversarial manipulation. This framework combines entropy-based, uncertainty-based, and geometric-based models, resulting in significant improvements in accuracy and robustness across various Natural Language Understanding datasets, including FEVER and HotpotQA.