The ITSERR project welcomes William Rudman (Brown University), recipient of an ITSERR TNA grant, who will deliver a seminar at the CNR Research Area in Pisa on April 4, 2025, at 10:30 AM.
Seminar Title:
“Outlier Dimension in LLMs and Multimodal-LLMs: Mechanisms for Task Adaptation and Factual Recall”.
Understanding how artificial intelligence models embed tokens in vector space is essential for interpreting their behavior. This seminar will explore the geometric properties of Large Language Model (LLM) and Multimodal-LLM (MLLM) representations through three studies:
- IsoScore and Isotropy in LLMs – Introduction of a new metric to measure how variance is distributed in embedding spaces, revealing the dominant role of “outlier dimensions.”
- Task Adaptation in LLMs – Analysis of how LLMs encode task-specific knowledge, highlighting the key role of outlier dimensions.
- Outlier Dimensions and Factual Recall in MLLMs – Presentation of the VisualCounterfact dataset, developed to investigate how multimodal models store factual associations by altering the visual properties of objects.
Rudman’s research provides new insights into the inner workings of artificial intelligence models, contributing to a deeper understanding of how they process and retrieve information.
Location: CNR Research Area in Pisa, Aula Faedo
Remote participation available