Neural retrieval systems are increasingly deployed in multi-domain settings, where each domain may have its own vocabulary and word distribution. A common approach is to fine-tune separate models for each domain, achieving strong in-domain performance while increasing training costs and storage requirements. In contrast, a single model trained jointly across domains avoids model duplication but weakens robustness under domain shift. This creates a trade-off between effectiveness and resource-efficiency. To address this issue, we propose EMAR, a retrieval system that employs a single frozen model shared across all domains and introduces lightweight domain-specific adapters. At inference time, a learned routing mechanism dynamically selects the most suitable adapter for each query. EMAR enables efficient domain-specific retrieval by sharing the computationally expensive backbone across multiple domains and without maintaining multiple full models. Experimental results on three multi-domain benchmarks spanning seventeen domains and three retrieval architectures show that EMAR achieves retrieval effectiveness on par or even better than fully fine-tuned domain-specific models while requiring storing only a small fraction of parameters per domain. Furthermore, EMAR exhibits strong generalization to previously unseen domains, highlighting its effectiveness as an extensible solution for multi-domain retrieval.
Pasha Abrishamchian, Eugene Yang, Andrew Yates, Nazli Goharian, Ophir Frieder
原文摘要
Coverage-evaluation benchmarks quantify whether retrieved evidence covers the different pieces of information needed to satisfy an information need. Existing benchmarks mostly consider factual claims that may already be known to pretrained retrieval components, leaving open whether retrievers find corpus evidence when it contradicts those priors. We study how retrievers behave when their parametric knowledge and the information in a corpus disagree. To do so, we develop a process to augment a nugget-based coverage benchmark with controlled contradictions. For each nugget, we create a negated counterpart and edit its supporting documents to support that negation while preserving the context. Applying this process to CoverageBench, a benchmark measuring information coverage, we create CoverageBench-Negate, spanning six datasets and roughly 63,000 original-edited evidence pairs (one per nugget-document link). We attach rich construction metadata to every pair and additionally provide out-of-family LLM judgments over all pairs and a human audit of a sample. In baseline experiments, BM25 separates original from edited near-twin evidence substantially better than dense retrievers (Qwen3 and Jina), which stay near chance. We release the benchmark on Hugging Face Datasets (nuggets, documents, pairs) together with the construction materials, including the prompts used to create it.