ABSTRACT
This paper proposes MaRI-MT, an extension of MaRI for multi-task retrieval, introducing a task-adaptive re-parameterization module and a routing mechanism for dynamic task-specific expert selection, achieving state-of-the-art performance with reduced parameters.
PAPER · PDF
Loading PDF…
Key findings
MaRI-MT extends MaRI for multi-task retrieval with shared user representation.
Introduces a task-adaptive re-parameterization module based on Low-Rank Adaptation.
Dynamically selects task-specific low-rank experts for fine-grained knowledge sharing.
Achieves state-of-the-art performance across material retrieval, product search, and recommendation tasks.
Limitations & open questions
The paper does not discuss potential limitations or open questions.