Seminario Danilo Croce
Il LabSIL è lieto di annunciare il primo incontro del ciclo di seminari «La Babele dei Golem», dal titolo Hallucinations in LLMs.
Il seminario, a cura del Prof. Danilo Croce (Università degli studi di Roma, Tor Vergata), si svolgerà presso il Dipartimento di Filosofia della Sapienza, Università di Roma (Villa Mirafiori, Aula seminari, I piano).
---
Hallucinations in LLMs
This talk provides an exploration of the nature, causes, and mitigation of hallucinations in state-of-the-art language models. Starting with an introduction to the foundations of LLMs, it emphasizes the transformative role of technologies like transformers and pre-training in language understanding and generation. The main focus is on understanding the “hallucination” phenomenon, where models produce factually incorrect or contextually inappropriate outputs. We classify and dissect different types of hallucinations, exploring both factual inaccuracies and faithfulness issues in responses. We then address the causes of these hallucinations, spanning data limitations, training misalignments, and inference challenges. The talk also summarizes current strategies for detecting hallucinations, alongside methods for mitigating them through improved data handling or refined training techniques. Finally, we discuss open challenges and questions surrounding the reliability, knowledge boundaries, and ethical responsibilities of deploying LLMs. This presentation aims to foster an understanding of LLM limitations and engage participants in deeper reflection on the implications of these powerful AI systems in shaping human knowledge and ethics.
---
Link per partecipare:
https://uniroma1.zoom.us/j/88081003648?pwd=fK3obDsDhM2w4iuEGnmBXUITRZ7wLg.1