How do Language Models Bind Entities in Context?

arXiv:2310.17191v2 Announce Type: replace
Abstract: To correctly use in-context information, language models (LMs) must bind entities to their attributes. For example, given a context describing a "green square" and a "blue circle", LMs must bind the shapes to their respective colors. We analyze LM representations and identify the binding ID mechanism: a general mechanism for solving the binding problem, which we observe in every sufficiently large model from the Pythia and LLaMA families. Using causal interventions, we show that LMs' …

abstract analyze arxiv circle colors context cs.ai cs.cl cs.lg example green identify information language language models lms square type