Researchers released ZooWork-ShopRanker, a family of open e-commerce rerankers at 0.6B, 4B, and 8B parameters, aligned to shopping preferences via LLM-judge-labeled training pairs. They also released ShopRank-Bench, a ~10,000-pair private-traffic benchmark in dual text formats. The 8B and 4B models significantly outperform the strongest open reranker baseline, with gains extending to MTEB benchmarks.
An arXiv paper introduces MERIT-Rank, a reranking framework that evaluates query-document relevance across multiple LLM reasoning trajectories and consolidates them with a joint reranker. A progressive training method, PRPO, stabilizes trajectories while improving ranking quality, and the 4B model beats most 7B and even 32B rerankers on BRIGHT.
An arXiv paper introduces MERIT-Rank, a reranking framework that evaluates query-document relevance across multiple LLM reasoning trajectories and consolidates them with a joint reranker. A progressive training method, PRPO, stabilizes trajectories while improving ranking quality, and the 4B model beats most 7B and even 32B rerankers on BRIGHT.
A new arXiv paper proposes augmenting query-product pairs with structured LLM-generated attributes plus human-validated relevance, explanations, and centrality judgments for e-commerce search. Evaluated with a dual-encoder retriever and MLP re-ranker on an augmented ESCI subset, synthetic signals approach human-oracle performance, with the largest gains on difficult, low-performing queries.
A new arXiv paper proposes augmenting query-product pairs with structured LLM-generated attributes plus human-validated relevance, explanations, and centrality judgments for e-commerce search. Evaluated with a dual-encoder retriever and MLP re-ranker on an augmented ESCI subset, synthetic signals approach human-oracle performance, with the largest gains on difficult, low-performing queries.
A new arXiv paper presents a scalable system for discovery-augmented search in e-commerce, particularly grocery, that generates implicit user intents to expand candidate recall while maintaining relevance. The approach addresses the limitation of traditional search systems that prioritize precision over recall, improving discoverability of substitute, complementary, and thematically related items.
A new arXiv paper presents a scalable system for discovery-augmented search in e-commerce, particularly grocery, that generates implicit user intents to expand candidate recall while maintaining relevance. The approach addresses the limitation of traditional search systems that prioritize precision over recall, improving discoverability of substitute, complementary, and thematically related items.
An AI lab researcher shared a new open-source reranking method built on Relational Transformers, a tiny architecture under 100M parameters that processes typed database cells for prediction tasks. The method aims to predict search result rankings quickly over large result sets, incorporating consumer preferences and other context.
An AI lab researcher shared a new open-source reranking method built on Relational Transformers, a tiny architecture under 100M parameters that processes typed database cells for prediction tasks. The method aims to predict search result rankings quickly over large result sets, incorporating consumer preferences and other context.
An arxiv paper argues that LLM-based scorers (rerankers, reward models, multi-document QA scorers) are selected on ranking quality, but their scores drive decisions like thresholds and selection. It shows equal ranking quality does not imply equal decisions, and proposes training order-consistent scorers. This matters to practitioners deploying rerankers or reward models where retained sets, not just rankings, affect outcomes.
An arxiv paper argues that LLM-based scorers (rerankers, reward models, multi-document QA scorers) are selected on ranking quality, but their scores drive decisions like thresholds and selection. It shows equal ranking quality does not imply equal decisions, and proposes training order-consistent scorers. This matters to practitioners deploying rerankers or reward models where retained sets, not just rankings, affect outcomes.