HELIOS: Harmonizing Early Fusion, Late Fusion, and LLM Reasoning for Multi-Granular Table-Text Retrieval

Sungho Park; Joohyung Yun; Jongwuk Lee; Wook-Shin Han

cs.DBFeb 25, 2026

HELIOS: Harmonizing Early Fusion, Late Fusion, and LLM Reasoning for Multi-Granular Table-Text Retrieval

Authors: Sungho Park, Joohyung Yun, Jongwuk Lee, Wook-Shin Han

Organizations: POSTECH, Republic of Korea · Sungkyunkwan University, Republic of Korea

Abstract

Table-text retrieval aims to retrieve relevant tables and text to support open-domain question answering. Existing studies use either early or late fusion, but face limitations. Early fusion pre-aligns a table row with its associated passages, forming "stars," which often include irrelevant contexts and miss query-dependent relationships. Late fusion retrieves individual nodes, dynamically aligning them, but it risks missing relevant contexts. Both approaches also struggle with advanced reasoning tasks, such as column-wise aggregation and multi-hop reasoning. To address these issues, we propose HELIOS, which combines the strengths of both approaches. First, the edge-based bipartite subgraph retrieval identifies finer-grained edges between table segments and passages, effectively avoiding the inclusion of irrelevant contexts. Then, the query-relevant node expansion identifies the most promising nodes, dynamically retrieving relevant edges to grow the bipartite subgraph, minimizing the risk of missing important contexts. Lastly, the star-based LLM refinement performs logical inference at the star graph level rather than the bipartite subgraph, supporting advanced reasoning tasks. Experimental results show that HELIOS outperforms state-of-the-art models with a significant improvement up to 42.6% and 39.9% in recall and nDCG, respectively, on the OTT-QA benchmark.

Figures & tables

Figure 1: Simplified examples of three cases where existing methods struggle to retrieve the question-related documents correctly. (a) Inadequate granularity of retrieval units leading to inaccurate retrieval results. (b) Entity linking results cannot estimate essential query-aware relationships. (c) Inability of advanced reasoning such as table aggregation and multi-hop reasoning.

Figure 2: Overview of HELIOS : The initial graph $G_{init}$ is early-fused to generate a graph $G_{d}$ . Each node and edge of $G_{d}$ are embedded. (1) The edges of $G_{d}$ are retrieved using the query $q$ , then integrated into a candidate bipartite subgraph $G_{c}$ . (2) The most query-relevant nodes in $G_{c}$ are identified as seed nodes. New nodes from $G_{init}$ are expanded from the seed nodes, forming an expanded graph $G_{l}$ . (3) LLM performs aggregation over restored tables to identify new relevant table rows, and then eliminates irrelevant passages.

Figure 3: The overall procedure of query-relevant node expansion. The beam width

b

is set as 2 in this example. The purple-colored nodes indicate the selected seed nodes.

Model	AR@2	AR@5	AR@10	AR@20	AR@50	nDCG@50	HITS@4k
Iterative Retriever	–	–	–	–	–	–	27.2
Fusion Retriever	–	–	–	–	–	–	52.4
OTTeR †	31.4	49.7	62.0	71.8	82.0	25.9	70.1
DoTTeR †	31.5	51.0	61.5	71.9	80.8	26.7	70.3
CORE †	35.3	50.7	63.1	74.5	83.1	25.4	77.2
COS †	44.4	61.6	70.8	79.5	87.8	33.6	81.8

Table 1: Retrieval accuracy on OTT-QA ’s dev set. Results marked with

{\dagger}

indicate reproduced values.

Model	AR@2	AR@5	AR@10	AR@20	AR@50	EM	F1
COS †	50.7	59.7	67.1	72.4	79.5	54.4	63.7
HELIOS	70.5	77.8	81.0	82.6	86.2	59.6	69.1

Table 2: Retrieval and end-to-end QA accuracy on the MMQA dev set.

Algorithm	Dev		Test
Algorithm	EM	F1	EM	F1
OTTeR	37.1	42.8	37.3	43.1
DoTTeR	37.8	43.9	35.9	42.0
CORE	49.0	55.7	47.3	54.1
COS	56.9	63.2	54.9	61.5
HELIOS	59.3	65.8	57.0	64.3

Table 3: End-to-end QA accuracy on OTT-QA .

Figure 4: End-to-end QA accuracy comparison across different readers for dev set of OTT-QA

Algorithm	EM	F1
HOLMES †	30.3	42.0
HELIOS	57.1	66.4

Table 4: QA Accuracy Comparison with HOLMES in a distractor setting adapted from OTT-QA .

Algorithm	AR@2	AR@5	AR@10	AR@20	AR@50	nDCG@50	EM	F1
HELIOS	63.3	76.7	85.0	90.4	94.2	47.0	59.3	65.8
w/ Finetuned SLR	63.2	76.8	85.1	90.3	94.8	47.6	59.4	65.9
w/o QNE	62.5	74.7	82.7	88.4	92.7	45.1	56.9	63.2
w/o SLR	60.0	75.2	84.7	90.1	94.6	46.5	59.0	65.7

Table 5: Retrieval accuracy of OTT-QA ’s dev set for HELIOS ’s design factors.

Retrieval Unit	AR@2	AR@5	AR@10	AR@20	AR@50	nDCG@50
Node	29.3	47.4	58.8	68.5	79.5	23.8
Star Graph	37.9	57.4	66.9	76.4	84.5	28.5
Edge	49.1	63.1	70.6	77.6	85.1	34.2

Table 6: Retrieval accuracy across different units

Algorithm	Execution Time (s)	nDCG@50
DoTTeR	0.08	26.7
CORE	4.13	25.4
COS	3.75	33.6
COS w/ ColBERT & bge	5.46	36.5
HELIOS	5.14	47.0
HELIOS w/ Finetuned SLR	4.76	47.6

Table 7: Algorithm execution time comparison

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: High-level schematic workflow of HELIOS .

Figure 6: The overall process of star-based LLM refinement for queries classified as aggregation queries. Table segment nodes of the same color (black, orange) indicate segments that belong to the same original table.

Figure 7: Change in AR@50 with varying beam width

Figure 8: End-to-end QA accuracy comparison across different top-

k

values on the OTT-QA dev set.

Component	Processing Time
Entity recognition	3h 14m 32s
Entity linking	2h 20m 48s
Edge indexing	22h 29m 54s
Passage indexing	7h 49m 01s
Table segment indexing	2h 29m 09s

Appendix

Table 8: Graph construction time by component.

Component	Training Time
Edge retriever	2h 49m 37s
Edge reranker	1h 25m 32s
Node reranker	1h 39m 20s
Expanded query retriever (passages)	36m 32s
Expanded query retriever (tables)	4h 14m 45s

Appendix

Table 9: Fine-tuning time for retriever submodules.

Figure 9: Qualitative analysis on four question-answer pairs. (a) A case where passage verification is successful. (b) A first case where passage verification has failed. (c) A second case where passage verification has failed. (d) A case where table aggregation is successful.

Information Retrieval

Information Retrieval studies how to find, rank, and organize relevant information in large collections of data.

See all 459 papers

FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning
Mar 13, 2026Chaojie Sun, Bin Cao, Tiantian Li +3Table Retrieval LLM Reasoning
Zhejiang University of Technology Hangzhou, China · University of Aberdeen Aberdeen, UK
FT-RAG: A Fine-grained Retrieval-Augmented Generation Framework for Complex Table Reasoning
May 2, 2026Zebin Guo, Weidong Geng, Ruichen MaoMultimodal RAG Table QA
Georgia Institute of Technology Atlanta, Georgia, USA · Zhejiang University Hangzhou, Zhejiang, China · Zhejiang Lab Hangzhou, Zhejiang, China
H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning
Sep 1, 2026Jia Ling, Yangfan Wang, Chen Tang +4Table QA Hierarchical Representation Learning
Harbin Institute of Technology · AI Research Center, Midea Group (Shanghai) Co., Ltd. · Changchun University of Science and Technology +1

Mar 13, 2026J/K move · Enter open · S save

FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning

Chaojie Sun, Bin Cao, Tiantian Li, Chenyu Hou+2

Zhejiang University of Technology Hangzhou, China · University of Aberdeen Aberdeen, UK

Table Retrieval LLM Reasoning

Open paper AskPDF arXiv

With the rapid advancement of large language models (LLMs), growing efforts have been made on LLM-based table retrieval. However, existing studies typically focus on single-table query, and implement it by similarity matching after encoding the entire table. These methods usually result in low accuracy due to their coarse-grained encoding which incorporates much query-irrelated data, and are also inefficient when dealing with large tables, failing to fully utilize the reasoning capabilities of LLM. Further, multi-table query is under-explored in retrieval tasks. To this end, we propose a hierarchical multi-table query method based on LLM: Fine-Grained Multi-Table Retrieval FGTR, a new retrieval paradigm that employs a human-like reasoning strategy. Through hierarchical reasoning, FGTR first identifies relevant schema elements and then retrieves the corresponding cell contents, ultimately constructing a concise and accurate sub-table that aligns with the given query. To comprehensively evaluate the performance of FGTR, we construct two new benchmark datasets based on Spider and BIRD . Experimental results show that FGTR outperforms previous state-of-the-art methods, improving the F_2 metric by 18% on Spider and 21% on BIRD, demonstrating its effectiveness in enhancing fine-grained retrieval and its potential to improve end-to-end performance on table-based downstream tasks.

Abstract

Information Retrieval

Explore similar work

FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning

FT-RAG: A Fine-grained Retrieval-Augmented Generation Framework for Complex Table Reasoning

H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning