MERLIN
A multilingual multimodal entity linking benchmark covering five languages—Hindi, Indonesian, Japanese, Tamil, and Vietnamese—used to evaluate the cross-lingual generalization capability of entity linking systems.
Timeline (last 90 days)
MERLIN works on both Gated DeltaNet and Mamba-2, validating its architecture-agnostic design claim.
On multiple LongBench subtasks, MERLIN improves by 12.0% on Mamba-2-780M.
MERLIN even surpasses the baseline method that consumes 4 times the memory.
MERLIN's current experiments are concentrated on models at the 1.3B and 780M parameter levels, and performance on larger-scale models remains to be seen.
When aggregating information across memory states, MSLA relies only on the average alignment between the query and each segment's key, ignoring the distribution of alignment within the segment.
MERLIN tracks both the first and second moments of each segment's key, enabling the query to recover the mean and variance of alignment with each segment.
MERLIN is a simple and general routing mechanism that does not depend on a specific architecture and can be embedded in different MSLA variants.
Moment tracking only requires maintaining a few statistics, so its computational overhead is far lower than computing the alignment distribution exactly at each position.
MERLIN was validated on multi-state variants of two architectures, Gated DeltaNet-1.3B and Mamba-2-780M.
On RULER needle-in-a-haystack and long-context question answering tasks, MERLIN improves over the baseline by 8.3% on Gated DeltaNet-1.3B.
4 more timeline events
All Facts (14)
MERLIN works on both Gated DeltaNet and Mamba-2, validating its architecture-agnostic design claim.
50%UnverifiedOn multiple LongBench subtasks, MERLIN improves by 12.0% on Mamba-2-780M.
50%UnverifiedMERLIN even surpasses the baseline method that consumes 4 times the memory.
50%UnverifiedMERLIN's current experiments are concentrated on models at the 1.3B and 780M parameter levels, and performance on larger-scale models remains to be seen.
50%UnverifiedWhen aggregating information across memory states, MSLA relies only on the average alignment between the query and each segment's key, ignoring the distribution of alignment within the segment.
50%UnverifiedMERLIN tracks both the first and second moments of each segment's key, enabling the query to recover the mean and variance of alignment with each segment.
50%UnverifiedMERLIN is a simple and general routing mechanism that does not depend on a specific architecture and can be embedded in different MSLA variants.
50%UnverifiedMoment tracking only requires maintaining a few statistics, so its computational overhead is far lower than computing the alignment distribution exactly at each position.
50%UnverifiedMERLIN was validated on multi-state variants of two architectures, Gated DeltaNet-1.3B and Mamba-2-780M.
50%UnverifiedOn RULER needle-in-a-haystack and long-context question answering tasks, MERLIN improves over the baseline by 8.3% on Gated DeltaNet-1.3B.
50%UnverifiedOn multiple LongBench subtasks, MERLIN improves by 13.5% on Gated DeltaNet-1.3B.
50%UnverifiedThe framework achieves an overall accuracy improvement of 6.9% over the current state of the art on the MERLIN benchmark.
50%UnverifiedThe framework achieves an accuracy improvement of up to 23.3% on rare entity slices.
50%UnverifiedMERLIN is a multilingual multimodal entity linking benchmark covering five languages: Hindi, Indonesian, Japanese, Tamil, and Vietnamese.
50%