Just-in-time and distributed task representations in language models (arxiv.org)

The authors analyze how and when language models form internal representations for new tasks during in-context learning, focusing on "transferrable" vector representations that can recreate task context if moved to another instance of the model. They find task representations evolve non-monotonically and sporadically: although task identity is decodable across the entire prompt, compact transferrable vectors often "come online" only at specific tokens, where the model condenses multiple pieces of evidence. These localized representations tend to capture minimal, semantically-independent "task scopes" (for example, a subtask), while longer or composite tasks are supported by more temporally distributed representations. Performance gains from additional examples align with the accrual of these transferrable vectors, revealing a two-fold locality—temporal (sequence-positioned) and semantic (scope-limited)—that the authors characterize as a just-in-time computation strategy. This work matters for interpretability and practical prompt design because it shows in-context learning is not simply a smooth, cumulative encoding of examples but relies on discrete, moveable task sketches plus distributed context signals. Technical implications include new diagnostics for probing representation transferability, potential targeted interventions or edits at tokens where task vectors appear, and guidance for few-shot prompting or chain-of-thought strategies that aim to trigger or consolidate these just-in-time representations. The findings point to modular, time-sensitive mechanisms underlying how large models adapt to new instructions without weight changes.