Abstract
I limiti dei Large Language Models
I Large Language Models hanno raggiunto capacità linguistiche estremamente sofisticate, ma la fluidità del linguaggio viene spesso interpretata come evidenza di comprensione, ragionamento o autonomia cognitiva. L’intervento discuterà alcuni limiti strutturali di questi sistemi: assenza di grounding epistemico, dipendenza da correlazioni statistiche, vulnerabilità a perturbazioni minime e difficoltà nella validazione autonoma delle informazioni. Verranno inoltre presentate evidenze recenti sul divario tra performance apparente e processi di valutazione sottostanti, con particolare attenzione alle implicazioni scientifiche e sociali dell’uso crescente di questi modelli.
Presentazioni
- DOCUMENTO
Video
Bio
Walter Quattrociocchi is a Full Professor of Computer Science at Sapienza University of Rome, where he directs the Center for Data Science and Complexity for Society. His research lies at the intersection of data science, network science, and computational social science, with a particular focus on the dynamics of misinformation, polarization, echo chambers, and the attention economy in digital ecosystems.
Abstract
The limitations of Large Language Models
Large Language Models have achieved extremely sophisticated linguistic capabilities, yet their fluency is often interpreted as evidence of understanding, reasoning, or cognitive autonomy. This talk will discuss some of the structural limitations of these systems: the absence of epistemic grounding, dependence on statistical correlations, vulnerability to minimal perturbations, and difficulties in the autonomous validation of information. It will also present recent evidence on the gap between apparent performance and the underlying evaluation processes, with particular attention to the scientific and societal implications of the growing use of these models.
Attribuzione 4.0 Internazionale