📝 SummarXiv

Abstract

Many of language models' impressive capabilities originate from their in-context learning: based on instructions or examples, they can infer and perform new tasks without weight updates. In this work, we investigate \emph{when} representations for new tasks are formed in language models, and \emph{how} these representations change over the course of context. We focus on ''transferrable'' task representations -- vector representations that can restore task context in another instance of the model, even without the full prompt. We show that these representations evolve in non-monotonic and sporadic ways, and are distinct from a more inert representation of high-level task categories that persists throughout the context. Specifically, models often condense multiple evidence into these transferrable task representations, which align well with the performance improvement based on more examples in the context. However, this accrual process exhibits strong locality along the sequence dimension, coming online only at certain tokens -- despite task identity being reliably decodable throughout the context. Moreover, these local but transferrable task representations tend to capture minimal ''task scopes'', such as a semantically-independent subtask, and models rely on more temporally-distributed representations to support longer and composite tasks. This two-fold locality (temporal and semantic) underscores a kind of just-in-time computational process underlying language models' ability to adapt to new evidence and learn new tasks on the fly.