arXiv:2602.22787v2 Announce Type: replace 
Abstract: Large language model (LLM) hallucinations, meaning fluent but factually incorrect generations, fall into two types: faithfulness violations, where the model misuses provided context, and factuality violations, where answers reflect errors in internal knowledge. Proper mitigation depends on knowing which source drives each answer. We study contributive attribution, i.e. the classification of the dominant knowledge source behind each output, and show that a simple linear probe trained on hidden representations can reliably identify it. We introduce AttriWiki, a self-supervised pipeline that automatically generates labelled training data by prompting models to recall withheld entities from memory or read them from context without relying on knowledge conflicts. Probes trained on AttriWiki achieve up to 0.96 Macro-$F_1$ on Llama-3.1-8B, Mistral-7B, and Qwen-7B, transfer to SQuAD and WebQuestions with 0.94-0.99 Macro-$F_1$, and generalise zero-shot to Tighidet et al. (2024)'s benchmark, outperforming their probe on conflicting settings without retraining. Furthermore, attribution mismatches raise error rates by up to 70%, though correct attribution does not guarantee correct answers, pointing to the need for broader detection frameworks.

تركزت الأبحاث الأخيرة على مسألة نسب المعرفة في نماذج اللغة الكبيرة (LLMs)، مع التركيز بشكل خاص على ظاهرة الهلوسة—الإنتاجات غير الصحيحة من الناحية الواقعية التي تنشأ عن انتهاكات الولاء والحقائق. تقدم الدراسة AttriWiki، وهو خط أنابيب ذاتي الإشراف مصمم لتصنيف المصدر المعرفي السائد وراء كل إنتاج، محققًا أداءً عاليًا مع عدة LLMs بما في ذلك Llama-3.1-8B وMistral-7B.

Investigaciones recientes se han centrado en la cuestión de la atribución de conocimiento en los grandes modelos de lenguaje (LLMs), abordando particularmente el fenómeno de las alucinaciones—resultados factualmente incorrectos que surgen de violaciones de fidelidad y factualidad. El estudio introduce AttriWiki, un pipeline auto-supervisado diseñado para clasificar la fuente de conocimiento dominante detrás de cada salida, logrando un alto rendimiento con varios LLMs, incluyendo Llama-3.1-8B y Mistral-7B.

Des recherches récentes se sont concentrées sur la question de l'attribution des connaissances dans les grands modèles de langage (LLMs), en particulier sur le phénomène des hallucinations—des sorties factuellement incorrectes résultant de violations de fidélité et de factualité. L'étude introduit AttriWiki, un pipeline auto-supervisé conçu pour classifier la source de connaissance dominante derrière chaque sortie, atteignant de hautes performances avec divers LLMs, y compris Llama-3.1-8B et Mistral-7B.

Recent research has focused on the issue of knowledge attribution in large language models (LLMs), particularly addressing the phenomenon of hallucinations—factually incorrect outputs that arise from faithfulness and factuality violations. The study introduces AttriWiki, a self-supervised pipeline designed to classify the dominant knowledge source behind each output, achieving high performance with various LLMs including Llama-3.1-8B and Mistral-7B.