CELEBRATING THE PHD DEFENSE OF XIYAN FU

We celebrate the successful PhD defense of Xiyan Fu!

Xiyan Fu's thesis

“Understanding and Improving the Compositional Generalization Abilities of LLMs in Reasoning”

provides new insights
into the compositional generalization abilities of (L)LMs, focusing on Natural Language Inference and other advanced reasoning tasks, as a testing ground, making several important contributions to our field:

Xiyan defines novel compositional generalization benchmarks for NLI and reasoning tasks that make use of different structured formats and that have been tailored to diverse, emerging applications, such as conversational AI systems, or graph-based generation tasks.
Through thoroughly designed experimental setups for compositional reasoning tasks, Xiyan investigates new design principles for data presentation in different learning modes, through which we gain new insights into how presenting tasks in an easy-to-hard ordering impacts learning success of LLMs in compositional generalization tasks.
The principle connects to learning aids that are successfully applied with humans, and therefore underline the close, but still underexplored relation between neural models, their learning mechanisms and those of the human mind.

This outstanding thesis offers a range of surprising insights into the generalization abilities of LLMs for compositional reasoning, their current weaknesses and how to efficiently improve on them!

Congratulations, Xiyan, for this fine work. We wish you all the best for the future!