Events

Date: 12 February 2026   Time: 13:00 - 14:00

Location: MB-503

Recent advances in large language models have encouraged the view that natural language, mediated through prompting, could act as a universal interface for general-purpose learning. In this talk, I challenge that assumption by analysing the structural limits of prompt-conditioned learning. I discuss how language is a constrained representation class: some task distinctions cannot be reliably encoded or recovered through prompts alone, leading to irreducible generalisation floors. I then present complementary work on meta-learning for formal reasoning, which may help address these limitations by learning shared structure across task families rather than relying on linguistic descriptions alone. Together, these results suggest that truly general learning requires representations that go beyond natural language, and instead exploit deeper task structure.