Teaching Language Models how to coordinate - Livestream

Large Language Models, despite their astounding reasoning abilities, are not faithful problem solvers. While their abilities are strongly correlated with scale, even humongous models like GPT-3.5 or GPT-4 can become inconsistent reasoners. Recent advances in verbose prompting techniques like chain-of-thought try to elicit step-by-step decomposition so that the model can solve a sequence of simpler problems to finally reach the goal. Augmenting external tools like web search or calculators has also been proposed to offload deterministic tasks. However, foundational language models learn neither problem decomposition nor tool usage. In this talk, the speaker will present potent solutions towards offloading reasoning subtasks in the case of mathematical problem solving: how to teach an auxiliary (and potentially frugal) language model to coordinate with black-box solvers, symbolic or language model-based, to successfully answer mathematical problems. This talk will focus on successfully teaching language models to perform reasoning from non-human feedback and how rewards beyond just the correctness of the final answer are essential for better learning.

Speaker: Tanmoy Chakraborty, IIT Delhi

Register to attend here