Daily AI Paper

Today’s pick is Telescopic Language Models, a new way to make one language model serve many compute budgets without training separate versions. The problem is simple but important: in real deployments, some requests need a fast, cheap answer, while others can afford more layers and more reasoning. Normally, that means building a whole family of models or compression schemes. This paper instead trains a single Transformer to work at every depth, from shallow to full size. The key idea is prefix supervision: at each step, the model learns from a randomly chosen truncated prefix of its layers, plus a full-capacity pass, so every prefix becomes a valid language model. That gives one nested model spanning the whole quality-speed continuum. The result is practical and elegant: no architectural changes, no extra inference machinery, and much better performance across budgets than fixed-exit baselines. For anyone building LLM systems, this is a compelling step toward truly elastic models.