Retrieving Reference Tokens in k-NN LM Inference

A language model architecture is designed to predict the next token by using two parallel computational streams that originate from the same query vector. The first stream uses the immediate, local context to generate a probability distribution over the vocabulary. The second stream uses the query vector to search a large external datastore, find the most similar historical contexts, and generate a second probability distribution based on the tokens that followed those contexts. The two distributions are then combined to produce the final prediction. What is the primary functional distinction between the information provided by these two streams?

Visual Representation of k-NN Language Model Inference

Diagnosing an Error in a Hybrid Language Model

A language model architecture enhances its predictions by combining information from its immediate context with knowledge from a large external repository. Arrange the following steps to accurately describe the data flow during its inference process.