Reproduce “Just-in-time and distributed task representations in language models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yuxuan Li, Declan Campbell, Stephanie C. Y. Chan, Andrew Kyle Lampinen

Google DeepMind researchers investigated how large language models represent tasks during in-context learning, revealing that functional task representations are dynamically formed "just-in-time" at specific tokens, contrasting with continuously decodable task identity information. They found these transferrable representations often have a localized scope, supporting only initial subtasks in complex generations, challenging prior assumptions of gradual, monotonic refinement.

Key insights

  • Language models maintain two distinct types of task representations: a broadly decodable, persistent "task identity" and a functionally active, "just-in-time" transferrable representation.
  • Functional task knowledge is not continuously refined across the prompt but rather activates sporadically and primarily at key tokens immediately preceding answer generation, demonstrating strong temporal locality.
  • Transferrable representations often have a limited semantic scope, primarily supporting initial subtasks in complex mixed-generation scenarios and showing decay over successive output units.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.