PhysDom: Enabling LLM’s to Reason about physical systems

Current generative models for language are unable to reason about physical systems. For example,

imagine querying a Large Language Model (LLM) about the expected behaviour of balls on a

billiards table. Even if the system is fully specified (positions and colours of balls on the table,

their physical parameters, etc.), current models cannot answer questions such as “Can the pink

ball be pocketed by the white ball without hitting the red ball?". To elicit the correct answer along

with reasoning from an LLM for this deceptively simple prompt involves tedious engineering of

the context. We propose to develop a general abstraction where a domain expert (an expert on the

physics of billiards) can encode rules of a physical system in a way that enables subsequent natural

language reasoning and question-and-answering via LLMs. PhysDom will provide an abstraction

for specification across domains/systems and a procedure for auto-generation of context.