When language models get things wrong, and how to tell: hallucinations, machine-generated text, uncertainty.
Language models are now used to search, summarise and write, but they can be confidently wrong. We study how they behave and where their limits are: spotting invented facts (hallucinations), telling machine-generated text from human writing, testing whether models understand cause and effect, and checking systems that answer questions from documents.
We also look inside models to understand how they reach their outputs, for example at which point a model decides which writing system to use.
Language preference and hallucination in multilingual retrieval-augmented generation
Do language models prefer some languages over others when answering from documents? We test this for chatbots that answer questions from a collection of documents in several languages.
Coordination of IDSIA's participation in the Swiss AI Initiative
Coordinated IDSIA's part in the national effort to build and apply large AI models on the Alps supercomputer at CSCS.
UBS–IDSIA AI Lab collaboration
The NLP group participates in the collaboration across different NLP-related projects.
Detecting misinformation and machine-generated text
Developed and tested ways to detect misleading content and text written by language models. Produced the BUST benchmark.
Misinformation in social media during COVID-19
Studied how to identify misinformation on social media, and public stance towards health policies, during the pandemic.
Oscar William Lithgow-Serrano
Group leader
Joseph Cornelius
Researcher
Vani Kanjirangat
Researcher
Gualtiero Marenco
Researcher