EMNLP 2026 Main

SAGE: Semantic Attribute Graphs
for Multi-Entity Visual Retrieval

Turning visually dense document images into entity–attribute graphs, so fine-grained evidence survives retrieval.

Yongjoo Kim1*, Mincheol Kwon1*, Seonga Choi1, Minseung Lee1,
Kyeong-Jin Oh2, Hyunyoung Lee2, Yunsu Choi2, Jungbeom Lee1†
1Korea University    2KT Corporation
*Equal contribution  ·  †Corresponding author
Semantic dilution and the SAGE entity-graph retrieval framework
Semantic dilution and our entity-graph retrieval framework. (a) Existing methods encode dense product images with coarse image or patch embeddings, where multiple attributes are mixed into a diluted representation. (b) SAGE decomposes the image into semantic attributes, constructs an entity-centered graph, and retrieves query-relevant subgraphs, preserving fine-grained evidence for multi-attribute comparison.
Abstract

What the paper is about

Dense document images often contain many fine-grained visual and textual entities whose relevance depends on a user query. Standard vision–language retrievers encode cropped regions with a single vector, which can mix distinct entity signals and obscure the evidence needed for fine-grained retrieval. We call this failure mode Semantic Dilution and quantitatively show that it degrades entity-level retrieval as a function of entity density.

To mitigate it, we propose SAGE, a training-free framework that parses semantic entities from dense document images, represents them as hierarchical graph nodes with multi-vector embeddings, and retrieves query-relevant evidence through iterative entity-level subgraph matching. We also introduce DEAR, a dataset of 1,055 query–image pairs sourced from product detail pages, where each query requires retrieving and comparing multiple fine-grained entities from visually dense inputs across four question types of increasing complexity.

Experiments show that SAGE substantially reduces semantic dilution and outperforms patch-level and OCR-based retrieval baselines on DEAR, achieving a Recall@3 of 0.849 and a generation score of 2.746 on multi-entity visual comparison queries.

Contribution 1

Semantic Dilution analysis

We formally define semantic dilution as a density-dependent failure mode in dense visual documents and show that increasing entity density degrades entity-level retrieval.

Contribution 2

DEAR benchmark

1,055 query–image pairs from product detail pages, where each query requires retrieving multiple specific entities across dense images and comparing the retrieved evidence.

Contribution 3

SAGE framework

A training-free, entity-centric graph retrieval framework using semantically coherent entities as retrieval units, with iterative subgraph-based evidence retrieval.

The problem

Semantic Dilution

When a whole dense page is compressed into one embedding, individual attributes stop being distinguishable — and the denser the page, the worse it gets.

We measure the semantic dilution score δ(I): the average cosine similarity between a full-image embedding and the embeddings of its isolated attribute crops, δ(I) = (1/n) Σk cos(Enc(I), Enc(rk)). A lower δ(I) means the full-image embedding retains less attribute-specific evidence.

Semantic dilution score decreases as the number of entities per image increases
Dilution grows with entity density. As the number of attributes in a single image increases (16 → 25 → 38), δ(I) drops from 0.6085 to 0.4497, indicating that full-image embeddings retain less attribute-specific evidence under denser visual layouts.

Semantic Dilution. The compression-induced degradation of fine-grained semantics that occurs when layout-agnostic crops or patches mix unrelated attributes into a single dense embedding. As entity density increases, target evidence is more likely to be mixed with irrelevant semantic signals, reducing the embedding’s discriminability for fine-grained retrieval and comparative reasoning.

This degradation carries over to retrieval

We rank each ground-truth image against all 502 collected images using its question. Across four different encoders, higher region density consistently lowers similarity to the ground-truth image and worsens its retrieval rank.

Encoder Sparse (6.8 regions) Dense (40.3 regions)
Sim. ↑Rank ↓Sim. ↑Rank ↓
CLIP0.6395.80.618 (−0.021)20.0 (+14.2)
Jina-CLIP-v20.6776.00.626 (−0.051)17.5 (+11.5)
SigLIP20.55213.00.547 (−0.005)24.5 (+11.5)
ColQwen36.793.528.39 (−8.40)8.0 (+4.5)

Question–image similarity and retrieval rank of the ground-truth image among all 502 collected images. Lower rank is better. CLIP, Jina-CLIP-v2 and SigLIP2 similarities are rescaled to 0–1; ColQwen scores retain their original scale.

The method

SAGE: Semantic Attribute Graph for Entity retrieval

Five modules, all built on frozen pretrained models — the entire pipeline is training-free.

Overview of the SAGE pipeline: attribute parsing, semantic captioning, hierarchical graph construction, retrieval planner, subgraph augmentation
Overview of SAGE. The pipeline (i) parses dense document images into attribute regions, (ii) generates semantic captions for image regions, (iii) constructs a hierarchical entity–attribute graph, (iv) retrieves query-relevant subgraphs through iterative planning, and (v) generates answers from structured multimodal evidence.
Module 1

Attribute Parser

An off-the-shelf OCR layout parser (PaddleOCR) decomposes each image into fine-grained regions, yielding bounding-box–text pairs plus a visual crop per region for non-textual signals such as icons and diagrams.

Module 2

Attribute Captioner

Image regions carry no OCR text, so a frozen VLM generates a semantic caption describing what each visual crop represents — producing an enriched attribute record (box, crop, text, caption).

Module 3

Hierarchical Graph Constructor

A frozen LLM groups coherent records into attribute-level leaf nodes under entity-level root nodes. Each leaf is embedded by jointly encoding its composite image with its composed text and captions.

Module 4

Retrieval Planner

Two-stage coarse-to-fine search: top-kr root nodes first, then root-conditioned attribute retrieval, reranked by a weighted root/attribute score — refined iteratively against a query checklist.

Module 5

Subgraph Augmentor

The retrieved subgraph is serialized into an ordered multimodal context — entities with their child attributes and visual crops — and handed to a frozen VLM for grounded answer generation.

Key idea. Instead of embedding whole pages or layout-agnostic patches, SAGE embeds only semantically coherent attribute units, and anchors them to the entity they belong to. Root-level anchoring reduces interference from irrelevant regions, while checklist-driven refinement ensures every item the query asks for is actually covered.

Iterative checklist refinement

A frozen LLM parses the query Q into a required-item checklist C(Q) and verifies whether each item is satisfied by the current retrieval set. If any item remains uncovered, the planner reformulates the query with the missing items and retrieves again — until the checklist is covered or a maximum iteration count is reached.

The benchmark

DEAR: Dense Entity-Attribute Retrieval

A diagnostic benchmark for semantic dilution. Every query requires retrieving multiple specific entities from visually dense images and comparing their attributes — within and across images.

1,055
QA pairs
502
Dense images
530
Multi-image queries
4
Question types (A–D)
8
Product categories

Construction pipeline

DEAR construction pipeline: data preparation, attribute tagging, QA generation
Dataset construction pipeline of DEAR. Product detail page images are collected from Best Buy and Auto Catalog Archive, segmented into localized regions with an OpenCV whitespace-band parser, annotated with VLM-assisted attribute tags, and verified by human annotators. Questions are then generated through VLM-guided generation and human verification, covering Type A–D configurations of increasing entity and attribute complexity.

Four question types

Examples of the four DEAR QA types
Examples of DEAR QA types. Each query requires retrieving evidence regions from dense product images and reasoning over the corresponding attributes. Type A–D cover single-entity single-attribute, single-entity multi-attribute, multi-entity single-attribute, and multi-entity multi-attribute questions.
Type A · 212

Single-Entity, Single-Attribute

Direct attribute extraction from one entity, e.g. “What is the earbuds’ charging method?”

Type B · 313

Single-Entity, Multi-Attribute

Several attributes of one entity must be jointly retrieved and weighed against each other.

Type C · 173

Multi-Entity, Single-Attribute

The same attribute is compared across entities, e.g. which product has longer playtime.

Type D · 357

Multi-Entity, Multi-Attribute

The hardest setting: multiple attributes of multiple entities, compared and traded off.

Statistics

Configuration# of Record
Images502
QA pairs1,055
Single-image query525
Multi-image query530
QA types
Type A (Single-Entity, Single-Attribute)212
Type B (Single-Entity, Multi-Attribute)313
Type C (Multi-Entity, Single-Attribute)173
Type D (Multi-Entity, Multi-Attribute)357
Question complexity
Extraction-oriented534
Reasoning-oriented521
Attribute density
Low (0–10)5
Medium (11–100)392
High (>100)108

Statistics of the DEAR benchmark.

Results

SAGE outperforms patch-level and OCR-based retrieval

All methods use the same frozen Qwen2.5-VL-72B generation backbone. Retrieval is Recall@3; generation is scored 0–5 by an LLM judge (GPT-5.4).

MethodType AType BType CType DOverall
Embedding-based
ColPali0.364 / 1.3130.220 / 1.3330.313 / 1.8240.246 / 1.7240.272 / 1.566
ColQwen0.500 / 2.0000.358 / 1.5780.513 / 2.6200.490 / 2.3660.463 / 2.139
OCR-based
PaddleOCR + BM250.649 / 2.1350.633 / 2.2750.617 / 4.6200.559 / 2.9640.606 / 2.097
PaddleOCR + DPR0.325 / 0.7000.206 / 0.8060.234 / 1.2910.160 / 1.7540.217 / 1.182
Caption-based
VLM Captioning + BM250.125 / 2.0000.095 / 2.6190.088 / 2.4120.138 / 2.7240.115 / 2.494
VLM Captioning + DPR0.063 / 1.6250.191 / 1.8570.206 / 2.4120.172 / 2.1720.163 / 2.030
Graph-based
SAGE (Ours)0.948 / 2.4190.927 / 2.7870.766 / 3.3210.718 / 2.8000.827 / 2.804
SAGE (Ours)0.948 / 2.5610.927 / 2.9970.807 / 3.0190.788 / 2.5120.849 / 2.746

Retrieval and generation performance on DEAR, reported as retrieval / generation. Generation is evaluated by LLM-Judge on a 0–5 scale. denotes adaptive top-k enabled.

Takeaway. SAGE improves retrieval across every question type, reaching an overall Recall@3 of 0.827 — and 0.849 with adaptive top-k, alongside the best generation score of 2.746. The gains are largest on Type C and Type D, exactly the queries that need evidence from multiple entities.

Generalization beyond DEAR

The same pipeline and modules, applied unchanged to two other visually dense document collections, using Recall@5 on questions requiring fewer than five supporting documents.

DatasetMethodRecall@5Gen. Score
UniDoc-Bench
(Finance)
ColPali0.68063.0615
ColQwen0.77453.0718
SAGE (Ours)0.80673.1949
SAGE (Ours)0.81143.2205
M3DocVQAColPali0.56442.7414
ColQwen0.55402.8966
SAGE (Ours)0.60983.6034
SAGE (Ours)0.65983.8621

Entity-level evidence transfers to document collections beyond product detail pages.

Component-wise ablations

NA, SA, TR and IR denote node aggregation, subgraph augmentation, two-stage retrieval, and iterative refinement.

IDNASATRIR Type AType BType CType DOverall
(a)0.429 / 2.2440.390 / 2.3140.165 / 2.2070.097 / 2.0680.338 / 2.150
(b)0.597 / 2.2210.596 / 2.4010.554 / 2.9240.500 / 2.4220.553 / 2.501
(c)0.597 / 2.3110.596 / 2.2270.554 / 2.5980.500 / 2.1690.553 / 2.297
(d)0.948 / 2.4190.927 / 2.7870.766 / 3.3210.718 / 2.8000.827 / 2.804
(e)0.948 / 2.5610.927 / 2.9970.807 / 3.0190.788 / 2.5120.849 / 2.746

Each entry is reported as retrieval / generation. The proposed modules are complementary, jointly mitigating semantic dilution and improving multi-entity comparison.

Citation

BibTeX

If you find this work useful, please consider citing:

@inproceedings{kim2026sage,
  title     = {SAGE: Semantic Attribute Graphs for Multi-Entity Visual Retrieval},
  author    = {Kim, Yongjoo and Kwon, Mincheol and Choi, Seonga and Lee, Minseung
               and Oh, Kyeong-Jin and Lee, Hyunyoung and Choi, Yunsu and Lee, Jungbeom},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing (EMNLP)},
  year      = {2026}
}