#91 - HATTIE ZHOU - Teaching Algorithmic Reasoning via In-context Learning #NeurIPS

Algorithmic Reasoning

Hattie Zhou, a PhD student at Université de Montréal and Mila, is pioneering the understanding of algorithmic reasoning in large language models (LLMs). Her research, in collaboration with the Google Brain team, focuses on teaching LLMs to perform tasks by formulating algorithms as skills and teaching them how to combine and use these skills effectively. This approach, known as algorithmic prompting, has led to a significant reduction in errors on certain tasks, showcasing its potential for enhancing reasoning capabilities in LLMs 1.

The breakthrough demonstrates algorithmic prompting's viability as an approach for teaching algorithmic reasoning to large language models, and may have implications for other tasks requiring similar reasoning capabilities.

--- Hattie Zhou

This method not only allows LLMs to apply known algorithms to new situations but also opens the possibility for models to discover new algorithms independently 2.

   

Defining Algorithms

Algorithmic reasoning, as defined by Hattie Zhou, involves solving tasks using specific algorithms that are input-independent, ensuring consistent results across various input distributions. This concept is crucial for tasks that can be algorithmically solved, although it also applies to softer algorithms where steps are less rigidly defined 2.

Performing reasoning by running an algorithm is what we refer to as algorithmic reasoning.

--- Hattie Zhou

The emergence of algorithmic reasoning in LLMs is a byproduct of their training, where models learn to identify patterns and apply them to new inputs, thus compressing vast datasets into functional algorithms 3.