Exécution de notre outil de synthèse sur un nouvel article

Résumé : The analysis of hallucinations in Large Language Models (LLMs), sequences that are coherent but factually false and semantically conflicting, has implications for the credibility and practical value of such systems. To address this concerning issue, the current paper presents a three-tier systematic review of 30 foundational and state-of-the-art publications, summarizing the existing knowledge on the error’s nature of hallucinations in LLMs and their antecedents and management. The classification of hallucinations is analyzed as intrinsic, extrinsic, compounding, and multimodal errors; focusing on their phenomenon in tasks of questions answering, summarization, and other multimodal cases. Current mitigation strategies are discussed at the data level (e.g., dataset cleaning and ad hoc knowledge injection), at the model level (e.g., fine-tuning and regularization), and at the inference level (e.g., retrieval-augmented generation and prompt engineering). Benchmarks such as HALUEVAL and TruthfulQA are compared to determine the usefulness and drawbacks for measuring the incidence of hallucinations. Some of the recurring issues such as the scarcity of domain-related corpora, adaptability in the LM inspired by scarce resources, and the ethical use of purposeful hallucinations are observed together with the prospects of a combination of approaches and large-scale evaluation frameworks. It is proposed that this review be used as a guide for future research in the domain of LLMs and as a reference that highlights gaps and possible directions for continuing work towards the development of more dependable, domain-adaptive, and ethically consistent LLMs.