Definition
Plain language
A test of whether a model can retrieve key-value pairs hidden deep inside a long context.
As stated in the literature
Multi-Query Associative Recall, a synthetic benchmark designed to measure how well architectures can retrieve key-value bindings from long contexts.
Why it matters: It cleanly separates architectures that truly handle long-range key-value retrieval from ones that only look like they do.
For example, the model reads a list of (key, value) pairs and is later asked to recall the value matching a particular key.
Heard on the show
“There's a benchmark in sequence modeling called MQAR — Multi-Query Associative Recall.”Episode 033 — Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval