Language is an imperfect vessel for meaning. It captures only a subset of the perceptions, intuitions, and intents that pass between speakers. A new paper on arXiv asks what this means for large language models, and proposes a way to measure the limits of what they can articulate.
Building on the observation that language is inherently lossy, the authors introduce the idea of an 'articulation upper bound' for LLMs. The work is described as a first step toward quantifying how much of the underlying meaning space a model can actually express in words.
Because the abstract is brief, the specific method and results are not yet available from this source. The paper's significance lies in framing a question that is often overlooked: not just what LLMs get right, but what they cannot say at all.