Source-linked AI summary
PDMR: Passage-Driven Multi-ID Document Retrieval
Smail Oussaidene, Mohand Boughanem
TL;DR
Single-identifier generative retrieval can compress multifaceted documents into a lossy target, leaving diverse query intents to compete for one access path. PDMR assigns multiple passage-level identifiers and trains with multiple valid targets; its ablations find passage-level supervision improves over document-level supervision, although PDMR does not reach state-of-the-art full-set performance on NQ320K.
Problem
Single-identifier retrieval compresses multifaceted documents into one target, creating a structural bottleneck for facet-specific queries.
Source: §1 · p. 1, §2.2 · p. 2Method
PDMR represents documents with multiple passage-level identifiers and uses multi-target training to distribute probability across valid targets while prioritizing the most relevant passage.
Source: §1 · p. 2, §3.6 · p. 4Results
Ablations find passage-level supervision consistently improves over document-level supervision on NQ and MSM, while the best query-to-passage assignment differs by dataset.
Source: §5.6 · p. 9Takeaways & Limitations
PDMR is complementary to identifier-learning and set-based decoding methods, rather than addressing all challenges in generative retrieval.
Source: §5.7 · p. 10PDMR does not reach state-of-the-art full-set performance on NQ320K, where GenRET [15], NOVO, Few-Shot GR [1], and TSGen [22] report stronger results.
Source: §5.7 · p. 10
Abstract
from arXiv · showhide
Generative Retrieval (GR) models map queries directly to document identifiers, replacing conventional retrieval over external sparse or dense indexes with autoregressive identifier generation. However, most generative retrieval frameworks rely on a single-identifier assumption, mapping each document to a single target sequence. This forces the model to represent all document content with one sequence. Since documents are often multi-faceted, this can lead to lossy representations and reduced robustness to query variation, where multiple query intents must compete for a single generative access path. In this work, we introduce Passage-Driven Multi-ID Retrieval (PDMR), a generative retrieval framework that represents documents through multiple passage-level identifiers. PDMR segments each document and assigns one identifier to each selected passage, which provides multiple semantic entry points for retrieving the same document. This multi-entry representation allows the model to align queries with specific semantic facets, thereby reducing the dependence on a single document-level target. To address the supervision ambiguity of this one-to-many mapping, we formulate training as a multi-target learning problem and explore an objective function designed to distribute probability mass across multiple valid passage-level identifiers. We evaluate PDMR on NQ320K and MS MARCO Document. On NQ320K, PDMR improves over strong generative and non-generative baselines on Recall@1 and MRR@100. On MS MARCO Document, PDMR achieves the best Recall@1 and MRR@10 among the reported methods, while remaining competitive on Recall@10. Controlled ablations further show that passage-level supervision, identifier design, training-query augmentation, and multi-target learning contribute complementary gains.
1 Introduction
Single-identifier generative retrieval can lose document facets and become sensitive to query variation; PDMR instead uses passage-level identifiers as multiple semantic entry points. It frames training as multi-target learning to distribute probability across valid passage identifiers.
1 Introduction
- A single generative target can lose diverse document aspects, making retrieval sensitive to queries about facets absent from the global identifier.Source: §1 · p. 1
- Unlike MINDER [8] and Few-Shot GR [1], which provide multiple views of a fixed retrieval unit, PDMR assigns identifiers to distinct internal passages.Source: §1 · p. 1
- PDMR represents each document with multiple passage identifiers, creating several semantic entry points for retrieval.Source: §1 · p. 1
- A weighted multi-target loss distributes probability across valid passage identifiers while prioritizing the passage most relevant to the query.Source: §1 · p. 2
- PDMR improves over single-identifier baselines on NQ320K and MS MARCO Document and remains competitive with strong retrieval methods.Source: §1 · p. 2
- A two-step LLM pipeline extracts passages and constructs Title-Composed passage identifiers (TC-IDs), with experiments and ablations on two benchmarks.Source: §1 · p. 2
2 Related Work
Related work develops generative retrieval, identifier designs, and multiple descriptive views, while PDMR shifts the retrieval unit to semantically distinct passages. This connects generative retrieval to fine-grained matching through localized access points.
2.1 Foundations of Generative Retrieval
- Generative retrieval maps queries directly to symbolic document or passage identifiers, with methods including natural-language names, learned codes, and n-gram identifiers.Source: §2.1 · p. 2
2.2 Identifier Design Strategies
- Identifier designs span textual and non-textual forms, with learned semantic identifiers refining how document-specific semantics are represented.Source: §2.2 · p. 2
- Textual identifiers expose content but may be long or ambiguous, whereas discrete or structured identifiers can be easier to constrain but offer weaker semantic signals.Source: §2.2 · p. 2
2.3 Multiview versus Multi-Facet Identifiers
- MINDER [8] and Few-Shot GR [1] offer multiple descriptions of a retrieval unit, while PDMR assigns identifiers to semantically distinct passages within documents.Source: §2.3 · p. 2
2.4 Fine-Grained Alignment Beyond Single Vectors
- Dense passage retrieval and late-interaction models preserve fine-grained matching; PDMR applies that general principle to generative identifier design.Source: §2.4 · pp. 2–3
3 Methodology
PDMR replaces one document-level target with multiple passage-level identifiers, then aligns document-level query supervision to those targets and trains across valid identifiers. At inference, it generates passage identifiers and ranks documents using the strongest associated passage score.
3.1 Task Formulation
- PDMR represents each document with multiple passage-level identifiers, each corresponding to a semantic facet, instead of one global identifier.Any selected passage identifier can retrieve its document, providing multiple semantic entry points while retrieval remains document-level.Source: §3.1 · p. 3
3.2 Overview of PDMR
- PDMR addresses missing passage-level relevance labels through query–identifier alignment and handles multiple valid targets with a specialized training objective.Its components are passage extraction, identifier construction, query–identifier alignment, and multi-target training.Source: §3.2 · p. 3
3.3 Passage Extraction
- A two-step LLM procedure first extracts short passages for coverage, then deduplicates and selects informative, distinct passages for precision.Passages are intended to capture a single idea, fact, or event; the initial set may include overlapping or low-importance passages.Source: §3.3 · p. 3
3.4 Passage Identifier Construction
- PDMR assigns each selected passage its own identifier and investigates random numeric, passage-title, and title-composed textual identifier families.Title-composed IDs combine document and passage titles; the document-first ordering gives passages a shared document-title prefix before the local passage title.Source: §3.4 · p. 3
3.5 Query–Identifier Alignment
- Training pairs combine passage-generated synthetic queries with original queries aligned to relevant passages using either all-passage or best-passage assignment.All-passage assignment broadens target coverage but may add noise; BM25-based best-passage assignment is more focused but covers fewer targets.Source: §3.5 · p. 4
3.6 Multi-Target Training Losses
- The weighted multi-target loss trains on a primary passage identifier and additional identifiers from the relevant document, distributing probability mass while prioritizing the primary passage.The single-target baseline treats query–identifier pairs independently and does not count other identifiers of the same document as valid targets.Source: §3.6 · p. 4
3.7 Inference and Document Ranking
- At inference, constrained beam search generates valid passage identifiers in one decoding pass, and each document receives its highest likelihood across associated identifiers.This max-likelihood aggregation turns multiple passage-level predictions into a document-level ranking score.Source: §3.7 · pp. 4–5
4 Experimental Setup
The experiments evaluate PDMR on NQ320K and MS MARCO Document against sparse, dense, and generative retrieval baselines. They also examine passage alignment and pose four research questions about supervision, identifier design, query augmentation, and multi-target training.
4.1 Datasets
- NQ320K contains 109,739 documents and 7,830 development queries, while the MS MARCO Document subset contains 323,569 documents and 5,187 development queries.NQ results include seen and unseen query splits; the MS MARCO subset retains documents associated with at least one training or development query.Source: §4.1 · p. 5
4.2 Baselines
- The comparison includes sparse, dense, and generative retrieval methods, with MINDER as the closest baseline because both it and PDMR relax the single-identifier assumption.MINDER provides multiple views of one retrieval unit, whereas PDMR assigns identifiers to distinct passage-level semantic units.Source: §4.2 · p. 5
4.5 Dataset-level passage alignment
- At the 25% threshold, MSM’s mean weak passage ratio is 38.50%, versus 22.62% for NQ, and 42.39% versus 24.34% of positive pairs have at least half their passages below threshold.Table 2 reports passage-level alignment statistics for positive query-document pairs; the authors suggest all-passage supervision may introduce more noisy targets on MSM than on NQ.Source: Table 2 · p. 6, §4.5 · p. 6
4.3 Evaluation Metrics
- NQ320K is evaluated with R@1, R@10, and MRR@100, while MS MARCO Document uses R@1, R@10, and MRR@10.Source: §4.3 · p. 6
4.4 Implementation Details
- Experiments use T5 Base for three epochs, with constrained beam search and a 0.6 weight on the primary passage in multi-target training.Qwen3-8B segments documents, and each original training query is replicated 20 times per selected passage-level identifier.Source: §4.4 · p. 6
4.6 Research Questions
- The experiments ask whether passage-level supervision improves retrieval, how identifier design affects performance, how query augmentation varies across datasets, and how multi-target training influences results.Source: §4.6 · p. 6
5 Results and Analysis
Across NQ320K and MS MARCO Document, PDMR performs strongly among identifier-based generative retrieval methods, with results favoring passage-level targets and dataset-dependent query-to-passage assignment. The ablations find complementary gains from passage supervision, identifier design, query augmentation, and multi-target learning, while the paper notes remaining limits in autoregressive decoding and fixed identifiers.
5.1 Overall Results
- 37.71 R@1 and 48.66 MRR@10 are best among reported MS MARCO methods; PDMR also improves over DSI-QG by +4.5 R@1 and +4.9 MRR@100 on NQ320K.On MS MARCO, R@10 is 71.49 and remains competitive; on NQ320K, PDMR records 67.6 R@1, 85.8 R@10, and 74.4 MRR@100.Source: §5.1 · pp. 6–7, Table 4 · p. 7, Table 3 · p. 7
5.2 RQ1: Passage-level supervision
- Passage-level supervision improves all reported metrics on NQ and MS MARCO against document-level supervision under single-target training.For example, NQ R@1 rises from 40.54 to 45.62 (+5.08), while MS MARCO R@10 rises from 41.30 to 52.27 (+10.97).Source: §5.2 · p. 8, Table 5 · p. 8
5.3 RQ2: Identifier design
- Textual passage identifiers outperform random numeric IDs, and Doc→Pass title-composed IDs are strongest among single-target variants overall.On NQ, Doc→Pass improves over Pass→Doc by +4.26 R@1, +2.22 R@10, and +3.84 MRR; on MS MARCO, Pass→Doc has slightly higher R@10.Source: §5.3 · p. 8
5.4 RQ3: Training-query augmentation
- Training-query augmentation helps both datasets, but all-passages assignment performs best on NQ whereas best-passage assignment is preferable on MS MARCO.The paper relates this difference to more weakly query-aligned passages in full-length MS MARCO web pages.Source: §5.4 · pp. 8–9
5.5 RQ4: Multi-target training objective
- Multi-target training improves RNID and TC-ID under synthetic-only supervision, with larger gains when combined with training-query augmentation.With TC-ID on MS MARCO, best-passage assignment reaches 37.71 R@1, 71.49 R@10, and 48.66 MRR, versus 30.02, 62.21, and 40.14 for all-passages assignment.Source: §5.5 · p. 9
5.6 Summary of Findings
- The ablations support multiple passage-grounded targets, semantic identifier content, and dataset-dependent query-to-passage assignment as complementary design choices.All-passages assignment can improve coverage for more uniformly aligned documents but may introduce noisy targets for heterogeneous documents.Source: §5.6 · pp. 9–10
5.7 Limitations.
- PDMR does not reach state of the art on NQ320K and retains autoregressive decoding sensitivity to early pruning and identifier ambiguity.It uses fixed identifiers derived from document structure rather than identifiers learned end-to-end for retrieval.Source: §5.7 · p. 10
6 Conclusion
PDMR replaces one document-level identifier with multiple passage-level identifiers and maps them back to the document during retrieval. Experiments show gains over document-level identifiers, with complementary contributions from identifier design, query augmentation, and multi-target learning.
6 Conclusion
- PDMR represents each document with multiple passage-level identifiers, exposing passage-based access points and mapping them back to the document during retrieval.The design addresses the mismatch between multi-faceted documents and single-target generative retrieval.Source: §6 · p. 10
- Across NQ320K and MS MARCO Doc, passage-level identifiers improve retrieval effectiveness over document-level identifiers.Identifier design, training-query augmentation, and multi-target learning contribute complementary gains.Source: §6 · p. 10
- Future work can explore multi-stage training and end-to-end frameworks that jointly learn passage selection, identifier construction, and query-identifier matching.The proposed multi-stage direction first learns reliable passage-level targets, then optimizes document-level retrieval.Source: §6 · p. 10
A Training Computation
On NQ320K, passage-level supervision increases training-instance counts, while multi-target training increases per-step computation and total training time.
A Training Computation
- 45.70M instances result from all-passage assignment with original training queries, compared with 26.07M for best-passage assignment.Single-target PDMR passage-level supervision uses 20.97M instances, up from 5.16M for the single-ID baseline (4.06×).Source: A Training Computation · p. 10
- 41.53 h is the multi-target training time, compared with 10.66 h for single-target PDMR.Both settings use the same number of NQ320K instances; multi-target loss evaluates several valid passage identifiers per query.Source: A Training Computation · p. 10
- Passage-level supervision mainly increases training-instance counts, whereas multi-target training increases per-step computation on NQ320K.The multi-target loss evaluates multiple valid passage identifiers for the same input.Source: A Training Computation · p. 10
B Illustration of Intra-Document Semantic Diversity
The MS MARCO Document example illustrates that different queries for one document can align with different internal passages. PDMR provides separate passage-level access points rather than compressing those facets into one document-level target.
B Illustration of Intra-Document Semantic Diversity
- Different queries for MS MARCO Document D83711 align with different internal passages, illustrating intra-document semantic diversity.Table 10 presents an example of queries for one document aligning with different passage-level semantic units.Source: B Illustration of Intra-Document… · p. 10, Table 10 · p. 11
- D83711 covers the endocrine system, while its queries target different internal passages.The example motivates passage-level identifiers as separate semantic access points.Source: B Illustration of Intra-Document… · p. 10
- Table 10 compares query-to-passage alignments for the same MS MARCO Document document.The caption identifies the table as an example of different queries aligning with different passage-level semantic units.Source: Table 10 · p. 11