Modern LLMs process text by letting every word attend to every earlier word, which is slow for long documents. Faster splits the context into chunks, compresses each into a memory, then combines those memories. Past methods judged a chunk only by its average similarity to the current word, so a chunk with a few highly relevant words looked the same as uniformly weak matches.
MERLIN also stores each chunk's variance, the spread of similarities. The current word then sees both the level and consistency of relevance before weighting each chunk's summary.
On long-context question answering, MERLIN beat prior methods by 8.3% on Gated DeltaNet and 7.0% on Mamba-2. Cheaper attention models can find relevant details in long contexts.