Do Current Retrievers Cover All the Evidence? A Controlled Study of Conjunctive Cross-Page Retrieval

2026-07-27 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A controlled study using n-Clue, comprising 1,000 queries over 2,021 documents, investigates the gap in conjunctive cross-page retrieval between finding a relevant document and ensuring it contains all requested evidence across different pages. The research reveals that condition-wise decomposition improves two dense backbones by 6.8-7.3 points, and lexical-visual fusion adds 8.7 points. Conversely, four generic rerankers reduce Gold-NDCG. Scaling a dense family from 0.6B to 8B parameters showed no change (0.0 points) in complete-first success. The strongest hybrid system found a gold document for 81.1% of queries but achieved complete-first success on only 35.8%. Page-aware visual systems performed even lower, surfacing all conditions for only 5.1-5.3% of queries. These findings identify condition coverage, rather than just gold document discovery, as the primary bottleneck in such retrieval tasks.

Key takeaway

For Information Retrieval Engineers developing systems for complex, multi-part queries, you must prioritize ensuring complete evidence coverage across documents, not just overall document relevance. Focus on integrating condition-wise decomposition and lexical-visual fusion techniques, which demonstrably improve the retrieval of all required evidence. Be cautious with generic rerankers, as this study shows they can significantly reduce complete-first success, creating a substantial gap between finding a document and confirming all conditions are met.

Key insights

Current retrieval systems struggle to ensure all conjunctive evidence is covered across document pages, even when relevant documents are found.

Principles

• Condition coverage is the central bottleneck in conjunctive retrieval.
• Decomposition and lexical-visual fusion enhance multi-condition evidence retrieval.
• Generic rerankers can degrade performance for conjunctive tasks.

Method

The study uses n-Clue, a controlled instrument with 1,000 queries over 2,021 documents, to measure "complete-first success" by pairing all-condition golds with subset-only documents.

In practice

• Implement condition-wise decomposition for multi-part queries.
• Integrate lexical-visual fusion to improve evidence coverage.
• Re-evaluate generic rerankers for conjunctive cross-page tasks.

Topics

• Conjunctive Retrieval
• Cross-Page Retrieval
• Information Retrieval
• Document Retrieval
• Lexical-Visual Fusion
• Dense Retrievers
• n-Clue Dataset

Best for: Research Scientist, AI Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Lost in a Single Vector: Improving Long-Document Retrieval with Chunk Evidence Aggregation — Improving long-document retrieval by aggregating independently encoded chunk-level evidence.
• From Articles to Premises: Building PrimeFacts, an Extraction Methodology and Resource for Fact-Checking Evidence — Decontextualizing hyperlink-anchored evidence from fact-checks significantly improves automated retrieval and verification performance.
• Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus — Cross-lingual deep research significantly degrades agent performance, revealing both retrieval and agent-side evidence integration challenges.
• Document Overlap Is Not Evidence Continuity: Measuring Retrieval Jitter in Citation-Based RAG Evaluation — Retrieval jitter causes significant instability in exact evidence spans, even when documents remain consistent, challenging RAG evaluation reproducibility.
• HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering — HiKEY uses document hierarchy and multimodal fusion for efficient, accurate RAG in ODQA.
• When Does Retrieval Beat Direct LLM Diagnosis in Rare Disease? An Empirical Study of Ontology Coverage — Retrieval-augmented LLM performance in rare disease diagnosis depends on ontology coverage, necessitating hybrid approaches.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.