ResearchPod Summary
How can e-commerce platforms unify six distinct relevance tasks—ranging from query-product matching to product-type similarity—into a single model without sacrificing performance due to conflicting signals or varying data volumes? The authors investigate how different model architectures (decoder-only LLMs vs. encoder-based cross-encoders) handle task identity and propose a unified multi-task learning (MTL) framework to improve knowledge transfer, particularly for data-scarce tasks.
The authors construct a 2.27M-example dataset by combining 800K human-labeled query-ad pairs with five derived tasks (e.g., query-query, product-product) validated by human annotators. They compare three task-routing architectures:
They evaluate these architectures across LoRA-adapted LLMs (Gemma-2, Llama-3) and fully fine-tuned cross-encoders (ModernBERT, ELECTRA), ultimately using a majority-vote ensemble to combine the strengths of these diverse architectures.
This work provides a practical blueprint for e-commerce search and recommendation systems to move away from maintaining fragmented, task-specific models. By demonstrating that private transformer layers are essential for decoder-only LLMs to handle multiple tasks effectively, the authors offer actionable guidance for deploying unified, parameter-efficient models that leverage cross-task knowledge.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.