GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration

Italo Luis da Silva; Hanqi Yan; Yujing Wang; Jiangnan Ye; Lin Gui; Yulan He

cs.IRAug 23, 2026

GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration

Authors: Italo Luis da Silva, Hanqi Yan, Yujing Wang, Jiangnan Ye, Lin Gui, Yulan He

Organizations: Department of Informatics King’s College London London, England, UK

Abstract

Scientific papers may relate by problem, method, result, or contribution, but document-level retrievers collapse these into a single similarity score without saying why they are related. Citation- and similarity-based retrieval alone also confines search to the neighbourhood of what is already known, whereas generative retrieval generates document identifiers directly, enabling the exploratory retrieval that scientific discovery depends on. We connect papers in a graph whose edges are typed by these four facets, derived from facet items and citation signals, and distil it into a generative retriever whose identifiers are the papers' own facet text. Two graph properties do not survive naive distillation. First, because every training pair is an edge, naive enumeration indexes just 84% of the corpus. Coverage-aware distillation makes every paper learnable through a reverse-neighbour fallback, a minimum-coverage threshold, and edge-importance weighting. Second, constrained decoding guarantees that every generated identifier is a valid paper, but not that the graph connects it to the query. Graph-weighted reciprocal rank fusion scales each candidate's rank term by its query-candidate edge weight, dropping unsupported ones. On LitWeave, our constructed corpus of 11,359 NLP papers, Graft recovers 91% of its graph teacher's Recall@20 with no nearest-neighbour index or encoder at inference, and outperforms the graph teacher on query papers outside the corpus. It reproduces the graph's own facet labels at 0.922 precision, so every returned paper arrives labelled with the facet that surfaced it rather than an opaque score.

Figures & tables

Figure 1: Left: existing retrieval reduces relatedness to a single score and cannot say why two papers are related. Centre: LitWeave instead types every edge by facet, so a pair may align on one facet or on several. Right: Graft distils that graph into a generative retriever and fuses the per-facet lists with graph-RRF, so the final ranking keeps the typing. The diagram is schematic; measured results are in § 5 .

Figure 2: Graft on the CEval query. Top: different facets connect the query to different papers. E.g., a paper on YESciEval is related through method alone, a connection a document-level dense retriever can miss. Training: the typed graph teacher supplies the supervision for coverage-aware distillation, which raises target coverage from 84% to the whole corpus. Inference: the same graph supplies the sparse edge weights that graph-RRF uses, producing ranked papers whose facet provenance is marked by colour, as in the top panel, and which may credit more than one facet. Decoding runs separately per facet; the method run is shown. The diagram is schematic.

Method	R@20	R@50
Dense-graph	0.357	0.560
MiniLM (title+abstract)	0.295	0.440
Dense-only (facet)	0.274	0.440
BM25 (title+facets)	0.274	0.433
SPECTER2 (title+abstract)	0.220	0.341
S2 Recommendations API	0.124	0.129

Table 1: Main results. The full generative method approximates the dense-graph baseline at R@20 without a nearest-neighbour index or query-time encoder. The lower block is Graft ’s supervision comparison.

Method	R@5	R@10	R@20	R@50
Dense-graph	0.218	0.360	0.501	0.657
BM25	0.164	0.230	0.309	0.434
MiniLM	0.152	0.225	0.293	0.408
SPECTER2	0.121	0.170	0.225	0.314
Graft	0.367	0.453	0.548	0.709

Table 2: Retrieval for 500 query papers outside the corpus. Graft leads at every cutoff. Baseline configurations are those of Table 1 .

Fusion	R@20	R@50
Best single facet (problems)	0.115	0.115
Uniform RRF (no graph)	0.284	0.447
MINDER combination	0.310	0.464
Graph-RRF (all four facets)	0.322	0.447

Table 3: Comparison between fusion methods.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Facet	Mean	${\geq}\,4$
Problems	4.68	47/50 (94%)
Methods	4.83	49/50 (98%)
Results	4.77	47/50 (94%)
Contributions	4.85	50/50 (100%)
All	4.78	193/200 (96.5%)

Appendix

Table 4: Pooled groundedness ratings by facet, with core items averaged across their three raters: mean score and fraction of items rated

{\geq}\,4

Year	Train	Dev	Test
2019	1,479
2020	1,727
2021	1,839
2022	1,830
2023		1,543
2024			1,582

Appendix

Table 5: Corpus by publication year and split.

$n$	Venue	$n$	Venue
2,891	ACL	137	TACL
2,842	EMNLP	135	AACL-IJCNLP
1,273	NAACL	111	CoNLL
438	EACL	102	WMT
398	LREC	91	INLG
335	Findings of ACL	78	SIGDIAL

Appendix

Table 6: The twenty most frequent venues in the corpus, after venue-string normalisation.

$\alpha$	$\beta$	R@20	R@50
0.00	1.00	0.323	0.524
0.25	0.75	0.339	0.552
0.50	0.50	0.356	0.568
0.75	0.25	0.363	0.569
1.00	0.00	0.273	0.437

Appendix

Table 7: Edge-weight sweep over the facet/citation combination, dense-graph retrieval on the development split.

\alpha

weights the facet score,

\beta=1-\alpha

the citation score.

$K$	Coverage
10	52.9%
20	84.0%
50	98.5%

Appendix

Table 8: Top-

K

pruning level against target coverage. Our corpus uses

K=20

(bold).

Walk depth	R@20	R@50
1-hop	0.318	0.465
2-hop (ours)	0.357	0.560
3-hop	0.357	0.565

Appendix

Table 9: Walk depth for dense-graph retrieval on the test queries. Two hops capture almost all of the available recall; a third adds cost without benefit.

Facet	Stored facet item	Identifier (first 12 tokens)
Problems	Lack of systematic, comparable evaluation of post-hoc explainability techniques: existing studies focus on narrow setups (single method, task, or architecture) making it hard to choose suitable explainers for a given model and task.	lack of systematic, comparable evaluation of post-hoc explainability
Methods	Empirically test these diagnostic measures across three text-classification datasets (e-SNLI, TSE, IMDB) and three model architectures (CNN, LSTM, Transformer), including multiple random seeds and randomly initialized models to assess consistency.	empirically test these diagnostic measures across three text-classification datasets
Results	Rationale and dataset consistency measures show low-to-moderate correlations overall; no single explainer is universally best for rationale consistency, indicating that faithfulness, human-alignment, confidence indication, and consistency are separable properties.	rationale and dataset consistency measures show low-to-moderate correlations
Contributions	A comprehensive, practical suite of diagnostic properties for explainability methods together with explicit, automated measures (MAP, AUC-TP, SD+LR MAE, Spearman correlations) that are applicable across tasks and architectures.	a comprehensive, practical suite of diagnostic properties for explainability methods

Appendix

Table 10: The four identifiers contributed by “A Diagnostic Study of Explainability Techniques for Text Classification” , one per facet, beside the stored facet items they derive from. Identifiers are lowercased, stripped of leading markers, and cut to twelve tokens.

Facet	Nodes	Leaves
Problems	396,133	41,997
Methods	481,796	47,490
Results	469,724	46,827
Contributions	447,281	47,974
Total	1,794,934	184,288

Appendix

Table 11: Per-facet prefix-trie size. Internal nodes exclude the root; each leaf is one distinct DocID.

Facet	Query $\rightarrow$ retrieved paper	Rank	Shared, in different words
problems	Higher-order comparisons of sentence encoder representations $\rightarrow$ Designing and interpreting probes with control tasks	92	Whether an interpretability method measures the representation or the instrument: via representational geometry, and via probe expressivity.
methods	Findings from the Bambara–French machine translation competition $\rightarrow$ AmericasNLI	84	Building a parallel resource for a genuinely low-resource language and adapting a multilingual pretrained model to it.
results	Semantic representation for dialogue modeling $\rightarrow$ Entity, relation and event extraction with contextualized span representations	74	An explicit graph over a contextual encoder recovers long-range structure the encoder alone misses: across utterances, and across sentences.
contributions	A challenge dataset for conversational stance detection $\rightarrow$ Build it break it fix it for dialogue safety	80	A multi-turn dialogue dataset built to capture what single-turn collection misses, for stance and for safety.

Appendix

Table 12: One judged non-obvious retrieval per facet. “Rank” is the retrieved paper’s rank in the facet-similarity ranking.

DocID scheme	R@20
Numeric codebook	0.121
C2T cluster keywords	0.260
Facet-item bullet	0.296

Appendix

Table 13: Identifier scheme (uniform-sampling baseline, graph-RRF R@20).

$K_{\text{min}}$	R@20
1	0.299
2	0.304
3	0.322
4	0.294

Appendix

Table 14: Coverage-floor sweep: minimum training views per DocID.

Fusion	R@20	$\Delta$
Graph-RRF (all four facets)	0.322
$-$ problems	0.303	-0.020
$-$ methods	0.321	-0.002
$-$ results	0.317	-0.005
$-$ contributions	0.316	-0.006

Appendix

Table 15: Graph-RRF with one facet removed (

K_{\text{min}}{=}3

model), and

\Delta

its change in R@20 from the full fusion.

Query facet	Top retrieved related paper
Problems	Flexible Text Generation for Counterfactual Fairness Probing
Methods	YESciEval: Robust LLM-as-a-Judge for Scientific Question Answering
Results	Optimal and Efficient Text Counterfactuals Using Graph Neural Networks
Contributions	DISCO: Distilling Counterfactuals with Large Language Models

Appendix

Table 16: Facet-attributed retrieval for the query paper “CEval: A Benchmark for Evaluating Counterfactual Text Generation” . Each facet surfaces a different neighbour, retaining why it was recommended.

Variant	R@20	R@50
Uniform sampling (base)	0.296	0.430
CE reweighting by edge score	0.303	0.436
Graph-weighted sampling (ours)	0.307	0.442
Exhaustive Cartesian pairing	0.242	0.328
Hard negatives, top-50	0.211	0.277
Hard negatives, top-100	0.201	0.275

Appendix

Table 17: Supervision and objective alternatives. Top: ways of using the edge score. Bottom: alternative training-pair construction and negative sampling.

Model	R@20	R@50
LoRA $r{=}16$	0.169	0.232
LoRA $r{=}32$	0.205	0.282
LoRA $r{=}64$	0.217	0.293
LoRA $r{=}128$	0.230	0.314
LoRA $r{=}256$	0.237	0.337
Full fine-tuning	0.322	0.447

Appendix

Table 18: LoRA rank sweep vs. full fine-tuning, graph-RRF recall.

Base model	R@20	R@50
Llama-3.2 1B	0.322	0.447
Qwen3 0.6B	0.248	0.332
SmolLM2 360M	0.089	0.130
SmolLM2 135M	0.084	0.112

Appendix

Table 19: Base-model size (graph-RRF recall).

Package	Version	Licence
accelerate	1.12.0	Apache-2.0
beartype	0.22.9	MIT
bitsandbytes	0.49.2	MIT
fastapi	0.135.1	MIT
leidenalg	0.12.0	GPL-3.0
matplotlib	3.10.7	PSF

Appendix

Table 20: Python packages used during development.

Information Retrieval

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

See all 459 papers

Beyond Single-Score Ranking: Facet-Aware Reranking for Controllable Diversity in Paper Recommendation
Mar 11, 2026Duan Ming TaoRecommender Systems Information Retrieval
EECS International Graduate Program, National Yang Ming Chiao Tung University
Graph2Idea:Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts
Jun 8, 2026Xu Li, Hanzhe Tu, Xun HanGraph Retrieval-Augmented Generation Scientific Hypothesis Generation
Southwest Petroleum University, Chengdu, Sichuan, China · Sichuan Police College, Luzhou, Sichuan, China
GRASP: Graph-Reasoning Aided Survey Planning for High-Fidelity Related Work Generation
Jul 4, 2026Haoming Li, Jessica OuyangCitation Network Analysis Graph-Based Planning
Department of Computer Science University of Texas at Dallas Richardson, TX 75080

Mar 11, 2026J/K move · Enter open · S save

Beyond Single-Score Ranking: Facet-Aware Reranking for Controllable Diversity in Paper Recommendation

Duan Ming Tao

EECS International Graduate Program, National Yang Ming Chiao Tung University

Recommender Systems Information Retrieval Learning to Rank Cross-Encoder Reranking

Open paper AskPDF arXiv

Current paper recommendation systems output a single similarity score that mixes different notions of relatedness, so users cannot specify why papers should be similar. We present SciFACE (Scientific Faceted Cross-Encoder), a reranking framework that models two independent facets: Background (what problem is studied) and Method (how it is solved). SciFACE trains two separate cross-encoders on 5,891 real seed-candidate paper pairs labeled by GPT-4o-mini with facet-specific criteria and validated against human judgments. On CSFCube, SciFACE reaches 70.63 NDCG@20 on Background (5.9 points above SPECTER) and 49.06 NDCG@20 on Method (31.1 points above SPECTER), competitive with state-of-the-art results. Compared with FaBLE without citation pre-training, SciFACE improves Method NDCG@20 by 4.1 points while using 5,891 labeled pairs versus 40K synthetic augmentations. These results show that high-quality grounded facet labels can be more data-efficient than large-scale synthetic augmentation for learning fine-grained scientific similarity.

Abstract

Information Retrieval

Explore similar work

Beyond Single-Score Ranking: Facet-Aware Reranking for Controllable Diversity in Paper Recommendation

Graph2Idea:Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts

GRASP: Graph-Reasoning Aided Survey Planning for High-Fidelity Related Work Generation