LLMs and Alignment. Research Talks: Francesco Croce

Join us for the second installment of the Research Talks series, focusing on LLMs and Alignment! 🎓✨

When: March 26, 17:00 - 19:00
Where: Maarintie 8

In this session, Professor Francesco Croce will delve into his groundbreaking research on Multimodal Chain-Of-Thought Reasoning Generalization.

Talk Overview:

Integrating reasoning into large language models (LLMs) and vision-language models has driven remarkable advancements in their capabilities. However, the nuances of reasoning model generalization remain elusive.

This talk will introduce a robust evaluation framework designed to scrutinize the effectiveness of chain-of-thought (CoT) strategies in a straightforward planning task. By employing varied input forms (both visual and textual) and CoT reasoning methods, we can meticulously assess model performance across in-distribution (ID) and out-of-distribution (OOD) scenarios.

Our findings indicate that while CoT reasoning enhances ID generalization across all input types, OOD generalization is severely constrained. Fascinatingly, we observe that reasoning patterns that amalgamate various text formats yield the most promising OOD performance. 🧠💡

Don't miss out on this enlightening talk! Click here for more information: Event Details