q135/random_k2/random_1.md

eprints@whiterose.ac.uk
https://eprints.whiterose.ac.uk

Universities of Leeds, Sheffield and York

Deposited via The University of Sheffield.

White Rose Research Online URL for this paper:
https://eprints.whiterose.ac.uk/id/eprint/228448/

Version: Published Version

Proceedings Paper:
Deng, X., Wang, X. and Stevenson, R. (2025) The next phase of scientific fact-checking:
advanced evidence retrieval from complex structured academic papers. In: Zamani, H.,
(ed.) ICTIR '25: Proceedings of the 2025 International ACM SIGIR Conference on
Innovative Concepts and Theories in Information Retrieval (ICTIR). ICTIR '25: International
ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval, 18
Jul 2025, Padua, Italy. ACM. ISBN: 9798400718618.

https://doi.org/10.1145/3731120.3744614

Reuse
This article is distributed under the terms of the Creative Commons Attribution (CC BY) licence. This licence
allows you to distribute, remix, tweak, and build upon the work, even commercially, as long as you credit the
authors for the original work. More information and the full terms of the licence here:
https://creativecommons.org/licenses/

Takedown
If you consider content in White Rose Research Online to be in breach of UK law, please notify us by
emailing eprints@whiterose.ac.uk including the URL of the record and the reason for the withdrawal request.

The Next Phase of Scientific Fact-Checking: Advanced Evidence
Retrieval from Complex Structured Academic Papers

Xingyu Deng
xdeng37@sheffield.ac.uk
University of Sheffield
Sheffield, UK

Xi Wang
xi.wang@sheffield.ac.uk
University of Sheffield
Sheffield, UK

Mark Stevenson
mark.stevenson@sheffield.ac.uk
University of Sheffield
Sheffield, UK

Abstract
Scientific fact-checking aims to determine the veracity of scien-
tific claims by retrieving and analysing evidence from research
literature. The problem is inherently more complex than general
fact-checking since it must accommodate the evolving nature of
scientific knowledge, the structural complexity of academic litera-
ture and the challenges posed by long-form, multimodal scientific
expression. However, existing approaches focus on simplified ver-
sions of the problem based on small-scale datasets consisting of
abstracts rather than full papers, thereby avoiding the distinct chal-
lenges associated with processing complete documents. This paper
examines the limitations of current scientific fact-checking systems
and reveals the many potential features and resources that could be
exploited to advance their performance. It identifies key research
challenges within evidence retrieval, including (1) evidence-driven
retrieval that addresses semantic limitations and topic imbalance
(2) time-aware evidence retrieval with citation tracking to mit-
igate outdated information, (3) structured document parsing to
leverage long-range context, (4) handling complex scientific expres-
sions, including tables, figures, and domain-specific terminology
and (5) assessing the credibility of scientific literature. Preliminary
experiments were conducted to substantiate these challenges and
identify potential solutions. This perspective paper aims to advance
scientific fact-checking with a specialised IR system tailored for
real-world applications.

CCS Concepts
· Information systems → Specialized information retrieval.

Keywords
Evidence retrieval, Scientific fact-checking

ACM Reference Format:
Xingyu Deng, Xi Wang, and Mark Stevenson. 2025. The Next Phase of Scien-
tific Fact-Checking: Advanced Evidence Retrieval from Complex Structured
Academic Papers. In Proceedings of the 2025 International ACM SIGIR Con-
ference on Innovative Concepts and Theories in Information Retrieval (ICTIR)
(ICTIR ’25), July 18, 2025, Padua, Italy. ACM, New York, NY, USA, 13 pages.
https://doi.org/10.1145/3731120.3744614

This work is licensed under a Creative Commons Attribution 4.0 International License.
ICTIR ’25, Padua, Italy
© 2025 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-1861-8/2025/07
https://doi.org/10.1145/3731120.3744614

1 Introduction
Fact-checking aims to assess the veracity of factual claims based
on credible evidence [37, 116] and serves as a crucial safeguard
for mitigating misinformation. Scientific fact-checking is a spe-
cialised variant of this task, grounded in scientific knowledge, with
the objective of combating misinformation that affects the pub-
lic, helping researchers in knowledge discovery and assisting in-
dividuals in understanding scientific advancements [97]. This is
particularly important given the rapid emergence of new scien-
tific findings, where both professionals and the public must assess
the credibility of information. A prominent case occurred during
the COVID-19 pandemic, in which politically motivated misinfor-
mationÐranging from inflated infection statistics to unsupported
treatmentsÐcirculated extensively, eroding public trust and endan-
gering health communication [58]. However, existing approaches
to scientific fact-checking remain limited, primarily relying on the
retrieval of evidence from relatively simple and small-scale sources
[16, 49, 62, 72, 75, 100, 101, 104]. For example, SciFact-Open [101],
the largest available dataset for scientific fact-checking, contains
500,000 documents ś substantially smaller than PubMed, which con-
tains over 37 million biomedical publications. In addition, SciFact-
Open consists only of abstracts, rather than full-text papers, thereby
excluding critical structural and citation information, ignoring long-
range context and scientific expression conveyed through tables
and figures. These design simplifications may hinder the applicabil-
ity of current approaches in real-world settings, where scientific
evidence is embedded in long and structurally complex documents
with multimodal content.

Fact-checking is a knowledge-intensive task, where the verifica-
tion process relies on sourcing evidence from a reliable upstream
Information Retrieval (IR) system. Emerging findings indicate the
value of effective retrieval in improving fact-checking systems.
For example, introducing even a small amount of noise into evi-
dence can significantly degrade fact-checking performance [76].
Recent Retrieval-Augmented-Generation (RAG) techniques have
been widely used for fact-checking [28, 47, 63, 70, 80, 84, 94], where
retrieval models are fine-tuned to identify high-quality evidence
for claim verification. These observations underscore the critical
role of robust evidence retrieval, as an ideal IR system for fact-
checking should rank all relevant evidence at the top while filtering
out non-evidential noise. Ensuring retrieval robustness is crucial to
maintaining sufficient yet relevant evidence, which is essential for
improving scientific fact-checking accuracy.

A major trend in fact-checking research is to consider realistic set-
tings that employ rich, diverse and timely evidence sources, as seen
in FEVER (using Wikipedia) [90] and AVeriTeC (using web-wide
resources) [78]. In document level evidence retrieval, current

436

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

general fact-checking systems over-rely on commercial search APIs,
which do not consider the specific requirements of fact-checking
[77, 91, 116]. Such reliance on commercial search APIs ś with lim-
ited adaptability ś has left document retrieval methodologies under-
explored in fact-checking, especially for domain-specific corpora
such as scientific fact-checking. Current scientific fact-checking sys-
tems primarily employ off-the-shelf IR methods [97], such as lexical
matching and semantic relevance ranking, which do not scale effec-
tively for large-scale scientific corpora. In addition, the distribution
of relevant evidence across scientific topics is highly imbalanced,
which degrades both retrieval effectiveness and efficiency, espe-
cially for claims with scarce supporting literature. SciFact-Open
[101], which extends the original SciFact dataset [100] for large-
scale evaluation, illustrates this issue: verification performance on
SciFact-Open drops by at least 15 F1 points for all well-performed
fact-checking systems developed in SciFact [101]. While increasing
corpus size enhances evidence diversity, it also amplifies retrieval
noise, reducing efficiency in both retrieval and verification. Beyond
that, high semantic relevance does not guarantee high evidential
relevance, and irrelevant yet semantically similar documents can
introduce noise into downstream verification [117]. These chal-
lenges underscore the necessity of developing tailored document
retrieval systems specifically designed for scientific fact-checking,
as effective retrieval is a prerequisite for accurate claim verification.
Beyond document-level evidence retrieval, within-document
evidence retrieval is also essential for processing complex sci-
entific literature. Scientific papers, unlike general fact-checking
documents, are long, structured, domain-specific and involve addi-
tional metadata. As scientific fact-checking evolves from abstract-
based to full-paper retrieval, retrieval models must account for
metadata (e.g., publish date, citations) and complex structured data
format (e.g., charts, tables, figures). This necessitates the adaptation
of verification models such as SciBERT [11] for domain-specific
terminologies [100], Longformer [12] for long-range dependen-
cies [102] and TAPAS [38] for tabular data verification [3]. Sci-
entific expressions in academic papers are highly structured and
contextually interdependent, where textual content, tabular data,
and figures mutually reinforce the conveyed information. How-
ever, existing scientific fact-checking systems primarily operate
at the abstract level, adopting methodologies similar to general
fact-checking [49, 50, 68, 72, 75, 97, 100, 102, 111, 119], albeit incor-
porating domain-specific models such as BioSentVec [21] and SciB-
ERT [11]. The development of public full-paper datasets aligns with
the requirement of real-world scientific fact-checking systems for
effective verification. This highlights the urgent need for retrieval
and verification methodologies that can leverage entire scientific
documents. Accordingly, within-document evidence retrieval and
its integration into verification pipelines should be explored to fully
unlock the potential of scientific literature for fact-checking.

This perspective paper presents a comprehensive examination of
the challenges associated with evidence retrieval in scientific fact-
checking, highlighting challenges that are not typically faced
within general fact-checking, leading to the need for specialised
retrieval and verification strategies. We advocate for proactive re-
search efforts to develop scalable methodologies while addressing
the limitations of current datasets. We structure our discussion into
two parts following a typical fact-checking pipeline: Sections 2ś3

explore document-level evidence retrieval while Sections 4ś8 ex-
plore within-document evidence retrieval in scientific publication
for scientific fact-checking. Each of the following sections identi-
fies a research challenge followed by a tentative and illustrative
research direction (RD).

2 Beyond Semantics
Evidence retrieval for fact checking is closely related to traditional
document retrieval techniques, which typically focus on retrieving
documents that are semantically similar to a query or contain match-
ing keywords. While this approach is effective in many scenarios,
it often fails to address the ultimate objective of fact-checking ś
successful claim verification. Evidence retrieval that relies solely on
semantic similarity may prioritise irrelevant or low-context infor-
mation, introducing noise into the subsequent verification process.
Recent IR studies [67, 92] show that semantic relevance alone may
not ensure utility in knowledge-intensive NLP tasks under the RAG
framework, suggesting the importance of utility-aware retrieval
strategies. To improve the verification utility of evidence retrieval
systems, techniques such as fine-tuning, joint optimisation, and
learning from verification feedback have been developed. These
approaches leverage relevance labels derived from annotated gold
evidence [40, 68, 117, 119, 120]. Although graded relevance has
been extensively explored in general IR, current evidence retrieval
systems for fact-checking often oversimplify relevance as binary,
failing to differentiate between fully non-evidential and partially
relevant evidence. This coarse-grained labelling scheme fails to
differentiate between completely non-evidential documents and
partially relevant (plausible) evidence. Negative examples and ran-
domly retrieved examples are equally treated as 0, despite exhibiting
varying degrees of evidential support. We argue that evidence re-
trieval should distinguish between non-evidential information and
plausible evidence, enhancing the model’s ability to identify previ-
ously unobserved but potentially useful evidence within large-scale
corpora.

Developing an IR system that can effectively differentiate be-
tween evidential and non-evidential information requires access
to fine-grained relevance labels during training. However, manu-
ally constructing negative samples is both complex and resource-
intensive due to the vast number of unlabelled documents and
sentences that lack explicit pairing with given claims. Furthermore,
assessing the degree of evidential support for a claim within unla-
belled documents is inherently challenging. To validate the impact
of fine-grained evidential relevance, beyond semantic relevance, we
carry out preliminary experiments which explore the use of down-
stream verification feedback to capture different levels of evidential
values.

Experiment Overview. The experiment investigates whether
combining verification feedback with semantic relevance improves
the performance of document evidence retrieval. Figure 1 presents
the pipeline, where probabilities from downstream verification
serve as feedback. Specifically, we integrate two components: 1) the
semantic relevance score, computed using an off-the-shelf reranker
model, and (2) the verification success feedback score, derived from
a fine-tuned verifier model, indicating the degree to which a docu-
ment is evidential for a given claim.

437

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

Table 1: Retrieval result on SciFact-Open and Check-COVID

SciFact-Open
R@50 R@20 R@10
BM25
45.22
54.78
66.09
71.09
79.13
88.91
monoT5-3B
71.30
81.95
91.96
+Verification
Check-COVID R@50 R@20 R@10
75.02
87.91
89.49
95.84
91.48
96.13

BM25
monoT5-3B
+Verification

81.96
93.16
94.55

R@5
38.04
57.17
62.61
R@5
67.59
82.06
84.04

R@3
30.87
48.26
52.83
R@3
61.35
74.93
77.80

R@1
20.22
31.09
32.17
R@1
46.18
58.28
61.84

Table 2: Evidence positions in the retrieved list. We select an ex-
ample from the SciFact-Open dataset. Claim: Female carriers of the
Apolipoprotein E4 (APOE4) allele have a reduced risk for Alzheimer’s
disease. Gold Evidence: [E1,E2,E3,E4,E5]

reranker model E1
BM25
monot5-3B
+Verification

838th
1st
1st

E2
141th
8th
3rd

E3
7th
16th
5th

E4
163th
2nd
2nd

E5
67th
302nd
27th

Datasets. We conduct our evaluation using datasets including:
(1) SciFact [100]: A corpus of 5,183 abstracts from scientific arti-
cles, with 809/300/300 samples for train, validation and test sets.
The test set is not publicly accessible. (2) SciFact-Open [101]: An
extended version of SciFact with 500K abstracts, re-annotating evi-
dence documents for 279 claims from the original SciFact test set. (3)
Check-COVID [104]: A COVID-19-specific fact-checking dataset,
containing 347 abstracts from CORD-19 journal articles and 1,504
expert annotated news-related claims.

MultiVerS is trained on the SciFact train set to create a verifier
model that provides verification feedback. Since the SciFact test
set is inaccessible, we evaluate document evidence retrieval on
SciFact-Open and full Check-COVID.

Results. The integration of verification feedback consistently en-
hanced document evidence retrieval, as +Verification outperformed
monoT5-3B across nearly all cut-off thresholds in both SciFact-Open
and Check-COVID (Table 1). The improvements are particularly ev-
ident at lower cut-offs, where retrieving the most relevant evidence
is crucial. In SciFact-Open, Recall@5 and Recall@3 increased from
57.17% to 62.61% and from 48.26% to 52.83%, respectively. Similarly,
in Check-COVID, these metrics improved from 82.06% to 84.04%
and from 74.93% to 77.80%. These improvements are particularly
meaningful given that the average number of gold evidence doc-
uments per claim is 1 in Check-COVID and approximately 2 in
SciFact-Open. To illustrate this improvement, we conducted a case
study examining how different retrieval methods ranked specific
evidence documents (Table 2). Compared to monoT5-3B, +Verifi-
cation successfully elevated the ranks of E2, E3 and E5, retrieving
three additional pieces of evidence within the top 5 results. Notably,
E5, ranked 302nd by monoT5-3B, was effectively rescued by adding
verification feedback in +Verification, demonstrating the value of
integrating verification-informed retrieval signals.

These findings highlight a promising research direction: shift-
ing from semantic-only retrieval towards evidence-aware retrieval,
where retrieval models explicitly account for evidential value. Our
results suggest that leveraging well-performing verification models
can help refine retrieval systems by distinguishing between purely

Claim

Retr ieved
documents

0.86

0.92

0.31

0.07

0.04

0.02

0.01

r elevance scor es

Document
Rer anker

monoT5-3B

Top-k
documents

Ver ification

M ultiVer S

Suppor t

0.7

Refute

0.1

ver ification
feedback

NEI

0.2

0.8

Figure 1: Pipeline of experiment

Approaches. monoT5-3B [64] has demonstrated strong perfor-
mance as a reranker for SciFact, as evidenced by its widespread
use as a strong baseline in studies [55, 86] on the BEIR benchmark
[89]. It also demonstrated state-of-the-art performance in evidence
retrieval for verification [68, 97, 102]. The model assigns a predicted
score, 𝑠𝑟
𝑐,𝑑 , representing the semantic relevance for a document 𝑑
to a claim 𝑐 as defined in Equation 1.

(1)

𝑐,𝑑 ∈ (0, 1)

𝑓 (𝑐, 𝑑) → 𝑠𝑟

𝑐,𝑑, 𝑠𝑟
MultiVerS is the best-performing verifier model on SciFact [97,
102]. We reproduced this model using the official implementation1
while adjusting the negative sampling parameter from 20 to 5 to
avoid over-fitted verification feedback. The model predicts the veri-
fication outcome as per the calculated probabilities for a document
𝑑 either supporting 𝑝𝑟
𝑐,𝑑 or providing insufficient in-
formation 𝑝𝑛

𝑐,𝑑 , refuting 𝑝𝑠
𝑐,𝑑 , relative to a claim 𝑐, as follows:
𝑐,𝑑, 𝑝𝑠

𝑉 (𝑐, 𝑑) = 𝑎𝑟𝑔𝑚𝑎𝑥 (𝑝𝑟

𝑐,𝑑, 𝑝𝑛

𝑐,𝑑 )

(2)

+Verification (Ideal Reranker Model) combines semantic relevance
and verification feedback to refine evidential retrieval. The final
retrieval score is calculated by summing the semantic score (𝑠𝑟
𝑐,𝑑 )
and the verification probabilities (𝑝𝑟
𝑐,𝑑 ), followed by a nor-
malisation step to ensure the score remains within (0,1), formulated
as:

𝑐,𝑑 and 𝑝𝑛

𝑠𝑟 +𝑣
𝑐,𝑑

= 1/2 ∗ (𝑠𝑟

𝑐,𝑑 + 𝑝𝑟
This formulation ensures that documents contributing to support
or refute labels receive higher retrieval scores, enhancing the evi-
dential quality of retrieved documents.

𝑐,𝑑 ) ∈ (0, 1)

𝑐,𝑑 + 𝑝𝑠

(3)

To evaluate retrieval effectiveness, we use Recall@k (R@k), a
common evaluation approach that measures the proportion of rele-
vant evidence successfully retrieved within the top 𝑘 results.

1https://github.com/dwadden/multivers

438

r

e

r

a

n

k

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

semantic relevance and plausible evidential relevance among unan-
notated documents. Furthermore, to continually improve evidence-
aware retrieval, we propose the development of tailored IR systems
capable of identifying evidential information, thereby enhancing
evidence retrieval for scientific fact-checking.

RD.1. Benchmark tailored IR system for fact-checking

The preliminary study presented in this work outlined a framework
to enhance evidence retrieval beyond only semantic relevance. To
overcome the limitations of existing IR systems in scientific fact-
checking scenarios, it is imperative to develop specialised IR sys-
tems capable of handling the specific challenges of verification tasks.
However, training an IR system on a single fact-checking dataset
risks poor generalizability and potential overfitting, particularly due
to data imbalance, a common issue in the relatively small datasets
characteristic of scientific fact-checking [97, 116]. Furthermore,
poor verification performance deteriorates retrieval accuracy, cre-
ating a vicious feedback loop that further degrades overall system
effectiveness. A multi-pronged strategy could mitigate these chal-
lenges by pooling verification signals from various high-performing
verifier models, leveraging large-scale datasets such as FEVER [90]
to improve training robustness, and providing a shared retrieval
checkpoint enable subsequent studies to fine-tune the model for
specific scenarios or datasets while reducing training cost. Recent
work [51, 74] has explored unified retrieval models for knowledge-
intensive NLP tasks, focusing on retrieval quality and downstream
task utility [73, 115], including question answering (QA) and fact-
checking. Similarly, we propose a verification-driven IR system for
evidence retrieval, which explicitly incorporates evidential infor-
matics. This approach follows a two-step training paradigm: general
pre-training on large, diverse datasets followed by domain-specific
fine-tuning. This approach balances scalability and domain speci-
ficity, ensuring IR models are both robust across different contexts
and highly effective in targeted fact-checking applications. Addi-
tionally, a corresponding benchmark should employ a diverse set
of evaluation metrics beyond for the fact-checking task to ensure
comprehensive assessment of performance within fact-checking
[6]. These metrics could include verification accuracy, reflecting the
downstream utility of retrieved evidence; decision latency, measur-
ing the computational efficiency of retrieval models; and robustness
to real-world conditions such as noisy data and incomplete evidence,
to improve system resilience.

Integrating verification feedback into evidence retrieval improves
relevance assessment beyond binary labels, enhancing retrieval per-
formance. Future research should focus on developing a benchmark
IR system tailored for fact-checking, incorporating fine-grained
relevance labels and verification-driven retrieval models. A scalable
pre-training and fine-tuning approach has the potential to improve
retrieval robustness and generalizability thereby producing more
accurate and efficient fact-checking systems.

3 Imbalanced resources of scientific topics
In existing general fact-checking datasets, such as FEVER which is
based on Wikipedia, the distribution of gold evidence per claim is
relatively even and sufficient. However, a significant imbalance of
evidence is observed in scientific fact-checking. While the SciFact

Table 3: Sufficient-evidence claim and none-evidence claim exam-
ples in SciFact-Open. ‘Evidence’ is the number of evidence in SciFact-
Open corpus and ‘Entities’ is the number of entities by searching
bold-keyword in PubMed.

Claim
Obesity is determined in part by
genetic factors.
LRBA controls CTLA - 4 expression.

Evidence Entities

24

0

499k

0.27k

corpus (~5K documents) maintains a relatively balanced number of
evidence documents per claim, this balance was disrupted when the
dataset was expanded to create SciFact-Open (~500K documents).
In this larger corpus, the majority of claims have none or only
one piece of supporting evidence while others have over. Claims
related to less-researched topics are generally associated with fewer
scientific publications, as illustrated by the examples in Table 3.

However, most fact-checking systems do not explicitly account
for evidence imbalance. A common approach is to use a fixed re-
trieval cut-off (i.e., selecting a predefined number of top-ranked
documents for verification). One of the most inefficient approaches
is setting the cut-off equal to the maximum number of evidence
per claim in the dataset, ensuring that all possible evidence is re-
trieved. This heuristic has not previously caused major issues since
general fact-checking datasets contain a relatively balanced num-
ber of supporting documents per claim. However, the imbalance
in SciFact-Open suggests that the simple approach may not be
suitable for open-domain scientific fact-checking with two major
drawbacks:

(1) Inefficiency. Although the maximum number of gold evi-
dence documents in the SciFact-Open dataset is 24, less than one
third of claims have more than two. Using the maximum number as
a cut-off would be inefficient due to the large number of documents
that would have to be processed by the verifier.

(2) Inaccuracy. Introducing irrelevant evidence into downstream
verification degrades fact-checking performance, whether the noise
is semantically related or completely random [76] (as discussed in
Section 2).

To address these challenges in the current and future studies of
scientific fact-checking, we proposed a research direction based
on flexible cut-off strategy for retrieving evidence based on claim
characteristics.

RD.2. Flexible cut-off for Retrieved Evidence

Ranked List Truncation (RLT) refers to the task of selecting an opti-
mal prefix of a ranked list of retrieved documents, with the goal of
balancing retrieval effectiveness and efficiency. Prior work explores
both heuristic and learned approaches, using either relevance labels
or features derived from score distributions to determine the cut-off
point [10, 52, 59, 61, 103, 109]. A related line of work is stopping
methods in technology-assisted review (TAR) [14, 15, 85], which
aim to retrieve as much relevant information as possible while min-
imising the effort spent on examining irrelevant documents. Both
approaches aim to optimise an expected metric over candidate cut
positions, typically using metrics such as 𝐹 1@𝑘 or 𝑟𝑒𝑐𝑎𝑙𝑙@𝑘. The
datasets used in these prior studies on RLT and stopping methods,
such as CLEF and TREC [25ś27, 33, 43ś45], exhibit imbalance but

439

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

typically contain enough relevant items per query to support recall-
based supervision and evaluation. This dependence on sufficient
relevance labels becomes problematic in fact-checking scenarios,
where gold evidence documents are typically rare, making both
recall-based stopping and supervised RLT approaches unsuited to
this problem.

To explore whether relevance score distributions indicate evi-
dence sufficiency, we compare well-studied and less-studied claims
from SciFact-Open. Following Wadden et. al. [101], claims with
four or more gold evidence documents are considered to be well-
studied and those with none to be less-studied. Using monoT5-3B,
we compute several statistics over the ranked document scores,
including first-document relevance, average and total scores, score
decay, and the initial-to-final score ratio. Table 4 shows that well-
studied claims tend to have higher top-ranked scores and sharper
decay patterns, suggesting that relevance distributions may serve
as indicators of evidence sufficiency.

Table 4: Statistical analysis of average relevance scores for
well-studied and less-studied claims. ‘I/F’ ratio is Initial-to-
Final ratio. ‘Exp k’ denotes the exponential decay factor k.

Metric
Less-
Well-

1st Doc Mean
0.502
0.741

0.941
0.995

Sum I/F ratio Exp k
1.544
3.514
25.097
0.651
1.963
37.052

Based on these findings, one possible direction is to leverage
existing techniques to estimate whether a claim is less-studied or
well-studied. A prediction module could utilise statistical features
of relevance distribution such as those presented in Table 4. In addi-
tion, metadata such as retrieved entity counts in PubMed (Table 3)
can serve as auxiliary signals to refine the prediction. Claims pre-
dicted as less-studied ś e.g., with low total relevance or steep score
decay ś may be assigned smaller cut-offs to reduce verification cost,
while RLT and stopping techniques could be applied to well-studied
claims where concentrated high scores suggest richer evidence.

While this naive strategy relies on heuristic features, it does not
explicitly optimise verification performance. To address this, future
approaches could explore learning a cut-off policy using feedback
from the verification stage. Specifically, truncation points may be
selected based on reward signals, such as whether the claim is cor-
rectly verified or the confidence of the verifier. This would bypass
the need for relevance-labelled supervision, which is often infeasi-
ble in scientific fact-checking due to sparse annotations. Inspired
by prior RLT and stopping method work, such a learned policy
could optimise both efficiency and factual accuracy by aligning
truncation decisions with downstream verification performance.

4 Time and Citation
General fact-checking evidence corpus such as Wikipedia and fact-
checking websites, often lack sentence-level evidence timestamps,
making it difficult to determine the original publish time of sentence
evidence in verification and hindering the development of time-
aware retrieval methods. Timeliness is important in fact-checking,
but in science, evolving evidence makes outdated studies particu-
larly problematic. For instance, early COVID-19 treatment studies

Table 5: Results of health QA task considering the different
thresholds of the published time of literature [99]

Year
≥2020
≥2018
≥2015
≥2010
≥2005
≥2000
≥1990
≥1980

Precision Recall

59.7
59.6
61.1
63.4
68.1
66.1
65.6
64.2

60.3
58.0
56.0
55.6
56.5
56.8
55.4
54.7

F1 score
58.7
57.9
53.9
52.8
52.0
51.8
51.3
50.0

were later refuted, and outdated evidence may lead to harmful de-
cisions. This section discusses whether scientific publications are
more suitable for time-aware fact-checking and explores possible
ways to leverage their inherent temporal characteristics.

4.1 Timeliness of evidence
Unlike general fact-checking, where historical and static facts re-
main unchanged, scientific knowledge continuously evolves. This
fundamental difference necessitates time-aware retrieval in scien-
tific fact-checking to ensure that retrieved evidence remain valid
and reflective of the latest scientific consensus. A time-sensitive
retrieval mechanism should prioritise recent publications to en-
sure that fact-checking systems incorporate the most up-to-date
methodologies and factual updates. This is especially crucial in
fields like healthcare, where relying on outdated information could
lead to misleading conclusions or incorrect decisions. For example,
during a rapidly evolving pandemic, a medical treatment initially
considered effective might later be deemed unreliable. This sec-
tion discusses the challenges and opportunities of integrating the
evidence timestamp into scientific fact-checking.

Scientific fact-checking aims to find evidential information in
the literature to verify a claim. Intuitively, considering outdated
literature negatively affects verification. Research in healthcare
question answering (QA) has demonstrated that time-aware re-
trieval improves system performance [99], as shown in Table 5. The
F1 score of the healthcare QA system improves as the publication
year of evidence documents becomes more recent. By extension,
scientific fact-checking on a large-scale corpus may also suffer from
incorporating outdated evidence. These findings highlight the need
for time-aware filtering in scientific fact-checking systems to en-
hance reliability. Fact-checking datasets could explicitly incorporate
timestamps as metadata to facilitate research into temporal rele-
vance in retrieval and verification. A general fact-checking study
[9] collected the ‘timestamp of last update’ of claims and evidential
documents, allowing later research [8] to explore the impact of
temporal data on verification. This study found that a time-aware
system achieved a 15% improvement in macro F1 score, underscor-
ing the importance of temporal information. However, this work
was limited by: (1) focusing on verification only, where evidence
documents had already been retrieved in a separate initial step,
without ensuring that retrieval prioritised high-quality and tem-
porally relevant evidence, and (2) the use of fragmented evidence ś

440

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

the general fact-checking datasets usually only contain small snip-
pets of web documents, omitting many important time expressions
present in full texts.

Given these findings, it is essential to develop a fact-checking
system that explicitly incorporates temporal awareness across both
retrieval and verification stages to meet real-world applications.

RD.3. Time-Aware Retrieval and Verification

Traditional fact-checking approaches often handle conflicting evi-
dence for a single claim by assigning neutral veracity labels, such
as “mixture," “unproven," or “not enough information" [37, 97].
However, conflicting evidence often arises due to outdated stud-
ies included in the retrieval process, which introduces noise and
adversely affects prediction accuracy [99]. This issue has been ob-
served in healthcare QA systems, where outdated evidence degrades
performance [99]. To address this issue, we propose a time-aware
approach that incorporates temporal information into both retrieval
and verification stages:

1.Retrieval Stage: Outdated evidence should be filtered or de-
prioritised during retrieval to ensure that the retrieved evidence
set is temporally aligned with the latest scientific findings.

2.Verification Stage: After filtering by time-aware retrieval, all
retrieved evidence should be considered but with differentiated
weighting based on temporal relevance. Recent evidence should be
prioritised through higher weights, while older evidence should
serve as supplementary or contextual information rather than pri-
mary evidence.

By processing outdated evidence differently across retrieval and
verification components, this direction explores how temporal in-
formation can reduce noise and improve the reliability of scientific
fact-checking systems.

4.2 Indirect evidence through citation
General fact-checking faces a number of challenges when attempt-
ing to determine the timeliness of evidence: (1) Lack of publication
timestamp [8]. Many fact-checking sources, such as Wikipedia and
fact-checking websites, do not provide precise publication dates
for individual sentences or paragraphs. Instead, they record only
the last edited timestamp, which does not accurately reflect when a
fact was first published. (2) Tracking the origin of evidence. Evidence
is often copied or paraphrased across multiple sources, making it
difficult to determine the original publication date of a statement.
(3) Search engine bias in retrieval. Pre-established fact-checking
datasets retrieve evidence using top-k search engine results, where
ranking mechanisms may prioritise recent documents due to time-
aware ranking biases. This can misrepresent the actual chronology
of claims and lead to fragmented evidence, making time extraction
unreliable.

While these challenges also affect scientific fact-checking, they
can be naturally mitigated by leveraging full-text papers rather than
abstract-only sources. The main advantages of using full-text papers
include: (1) Explicit publication metadata: Each piece of literature
has a clearly defined publication date, ensuring accurate temporal
tracking. (2) Citation tracing for indirect evidence: Mandatory cita-
tion rules in academic publications facilitate source tracing, even
when statements are referenced indirectly. (3) Structured nature

of academic papers: provides an indication of the origin of state-
ments. For example, evidence in the background section typically
references prior studies, whereas those in the abstract or results
sections represent findings from the current publication. Given
these inherent advantages, incorporating timestamps as metadata
offers a promising research direction for full-paper-based scien-
tific fact-checking. Exploring the temporal dynamics of claims and
evidence should further enhance retrieval accuracy. Additionally, ci-
tation tracking to trace the original source of paraphrased evidence,
ensuring the first-published timestamp is accurately recorded.

Citation-based tracking provides a promising approach to esti-
mating evidence timestamps. However, the widespread presence of
multiple citations in scientific literature makes it difficult to iden-
tify which references should be tracked, increasing computational
costs and reducing efficiency. Moreover, indirect citations and para-
phrased references, particularly in introductory sections, further
obscure the retrieval of the first-published source. To address these
issues, we propose the following research direction to develop a
more effective approach for citation tracking and timestamp attri-
bution.

RD.4. Citation-Based Evidence Tracking

Intuitively, self-contained evidence refers to information directly
presented in the body of the current paper, while cited evidence is
derived from external sources referenced by the paper. To explore
this distinction, we analysed 22 accessible full papers out of 24
gold evidence for the claim in Table 2. We prompted GPT-4o to
search for supporting/refuting evidence and determine whether
each piece of evidence was paraphrased/summarised from a citation
or is self-contained.

Figure 2: Evidence sources in scientific literature.

The result shown in Figure 2 reveals, as expected, that cited
evidence is predominantly located in the introduction, while self-
contained evidence is more common in the results and conclusion
sections. While we observed a few inaccurate outcomes (e.g., one
‘From Cited’ evidence appearing in ‘Abstract’ is misjudged), the
overall distribution remains discernible and interpretable. In ad-
dition, we also observed that sentences in scientific literature fre-
quently contain multiple citations, making it costly to manually
extract the original timestamp of cited evidence. To address this
challenge, we propose a two-step approach:

1. Identify track-worthy citations. Not all citations are equally
important for fact verification. Track-worthy citations should in-
clude: conclusive evidence directly influences claim verification and

441

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

plausible evidence that may impact claim assessment [108]. Since
checking every cited document is expensive, an initial filtering
step is required. A potential solution is to rank citations based on
their relevance and function. Citation recommendation [30, 34] is a
similar task to identify relevant publications for a given statement
using retrieval models. Beyond relevance, the function of citation
can be referred to as a signal to adjust priority. Citations are classi-
fied into eight categories: background, motivation, uses, extends,
similarities, differences, compare/contrast, and future work [42].
While single evidence has multiple citations, the function of cita-
tion can help identify the citations that align with the evidence.
For example, for predicted evidence ‘Experiments show model A
outperforms previous SOTA model B [citation 1,2,3]’, cited papers
for ‘model A/B’ in the ‘introduction’ function are less check-worthy
than the citation in the ‘compare/contrast’ function. By prioritising
high-impact citations, retrieval costs can be significantly reduced.
2. Track the original timestamp of evidence. Once track-
worthy citations are identified, the next step is to trace the original
timestamp of cited evidence via direct and indirect citation tracking.
Direct citation tracking can be applied if the publication date of the
cited paper is straightforward to retrieve. However, some evidence
is paraphrased or indirectly cited, requiring a deep tracking mech-
anism to trace the citation path. Citation graph analysis [17, 96]
can help map citation paths using directed graphs and applying
search algorithms to identify the earliest relevant source. A recent
study [118] found that reference errors ś references do not include
information to support statement ś frequently appear ranging from
11% to 41% across domains. Addressing these errors introduces
a sub-task for scientific fact-checking: verifying whether a cited
reference truly supports the claim. To ensure feasibility, an early
explorative study can assume that scientific literature generally
adheres to citation conventions, preventing infinite citation loops
in verification.

In summary, to explore time-aware fact-checking for scientific
literature, we propose two potential research directions: (1) Integrat-
ing temporal information into retrieval and verification to handle
outdated evidence and (2) Developing citation-based tracking meth-
ods to identify the original source and timestamp of evidence. These
directions provide a basis for studying the impact of time-aware
mechanisms in scientific fact-checking systems.

5 Structured Long-Context Evidence Retrieval
Existing scientific fact-checking datasets construct evidence cor-
pora using fragmented sentences, paragraphs, or abstracts [16, 49,
62, 72, 75, 100, 104]. However, scientific literature is typically pre-
sented in structured, visually rich formats, often as PDF documents,
where different sections serve distinct functions: abstracts, results,
and conclusions summarise key findings, while background and in-
troduction sections provide prior research context. With the diverse
and unique functionalities of scientific literature components, this
section explores challenges and potential research directions for
advancing scientific fact-checking at the full-paper level. Existing
fact-checking pipelines typically follow a document retrieval and
then sentence selection paradigm [37, 97, 116]. For general fact-
checking, evidence retrieval often uses top-ranked sentences from
top-ranked documents, treating them as self-contained evidence

Title
Authors

Publish Venue
Publish Time

Citations:
[ 25, 44 , 60 , 67, 78 , 82 , 92

Figure 3: An example for parsing the scientific literature

units. However, this approach neglects long-range context, as using
the extracted sentences ignores surrounding information to sup-
port verification. In contrast, scientific documents exhibit higher
document-level consistency of verdict, commonly one paper hav-
ing a sole standpoint to a given question, making document-level
processing necessary for scientific fact-checking [50, 102, 119].

LLMs have recently demonstrated growing capability to process
long contexts. However, it remains challenging to fact-check an
entire full-text document in a single pass [105]. LLMs are prone to
hallucinating content that is not grounded in the provided docu-
ments [2, 20, 23, 88] and are often susceptible to distraction from
irrelevant context [82]. Their reasoning capabilities also degrade
as text length increases [46]. A potential solution is context distil-
lation, as used in retrieval-augmented generation (RAG), to filter
high-quality context and mitigate hallucination, which improves
the performance of downstream QA tasks [106]. However, unlike
the QA task, fact-checking requires explicit retrieval of explicit. Fil-
tering context may remove crucial supporting evidence, leading to
incomplete verification. Moreover, much of the historical scientific
literature exists in PDF format. Although recent multimodal LLMs
are capable of consuming PDFs and conducting reasoning tasks,
their capabilities are still limited to surface-level understanding.
For example, they frequently fail to capture cross-page content
and complex layout structures [60, 87, 95]. By prompting GPT-4o
to locate evidence in PDF literature with the results presented in
Section 4.2, we observed that it mislocated evidence in incorrect
sections or non-existent sections. These limitations indicate that ex-
isting LLMs are not capable of supporting end-to-end fact-checking
for long-context scientific literature ś not only for raw PDFs but
also for plain-text documents.

In addition to their length, scientific papers follow a structured
format that introduces additional challenges for verification. Back-
ground and introduction sections often cite prior work, which
may conflict with conclusions drawn later, while discussion sec-
tions highlight limitations that can cast doubt on earlier findings.
These internally inconsistent signals may introduce misleading
information, demanding reasoning that accounts for both the fac-
tual assertions and the functional roles of different sections. Pars-
ing scientific documents into structured units enables a modular
pipeline where retrieval and verification can be independently opti-
mised. This modular structure facilitates interpretability, robustness,
and denoising of conflicting or irrelevant content. Layout-aware
tools offer a practical foundation for such structure-aware process-
ing [19, 32, 41, 56, 81], as illustrated in Figure 3.

442

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

To address these challenges, we advocate for a retrieval frame-
work that explicitly considers the document structure commonly
found within scientific literature. Such a system should identify
targeted evidence, capture long-range context across sections, and
suppress irrelevant or conflicting content. We next outline a direc-
tion toward adaptive, section-aware retrieval strategies designed
to meet these requirements.

RD.5. Adaptive Section-Aware Evidence Retrieval

Scientific literature follows a structured format where different sec-
tions serve distinct functions, presenting challenges for traditional
evidence retrieval and verification in fact-checking systems. Exist-
ing retrieval methods often operate at the sentence or paragraph
level, neglecting the long-range context and the structured nature
of scientific documents. Additionally, large language models (LLMs)
struggle to accurately associate claims with the appropriate sec-
tions, leading to potential misinterpretations and inconsistencies.
A promising research direction is adaptive section-aware ev-
idence retrieval, which dynamically adjusts retrieval strategies
based on document structure and claim types. This approach con-
sists of two key components:

Claim-evidence matching. Claims should first be matched to
the most relevant sections. For instance, experiment-driven claims,
such as “X method improves accuracy compared to Y,ž should pri-
marily retrieve evidence from the ‘Results’ and ‘Conclusion’ sec-
tions, as these contain empirical findings. In contrast, background
or theoretical claims, such as “X method is widely used in Y applica-
tions,ž should focus on the Introduction and Background sections,
which provide foundational knowledge. Prioritising section-aware
evidence retrieval helps filter irrelevant context and reduces re-
trieval noise.

Contextual expansion. After retrieving primary evidence, the
system should augment it with relevant contextual information
from other sections to improve interpretability. For example, method-
ological details from the ‘Analysis’ section can provide additional
support for experimental claims, while historical context from the
Background section can clarify theoretical claims. Access to the full
document allows for retrieving finer-grained evidence or comple-
mentary details that fragmented approaches may overlook.

In summary, an effective adaptive section-aware retrieval sys-
tem must overcome challenges in accurately parsing document
structures, prioritising relevant sections based on claim types, and
efficiently handling conflicting evidence. By integrating structured
document parsing, hierarchical retrieval strategies, and context-
aware reasoning, future systems can leverage richer evidence from
full-paper scientific literature while reducing LLM hallucinations.
Advancing these techniques will enhance the reliability and inter-
pretability of scientific fact-checking systems.

6 Multimodal content in Science
Scientific literature often conveys key evidence using non-textual
elements such as tables, charts, and figures, which are commonly
used to present experimental results, statistical analyses, and the-
oretical models, as shown in Figure 3. For full-text scientific fact-
checking, especially across various fields of science and technol-
ogy, it is crucial to move beyond text and accurately interpret

these structured elements to ensure comprehensive verification.
Fact-checking and misinformation detection on individual modal-
ities ś such as figures [1, 65, 66, 69, 93, 112], charts [4, 7], and
tables [3, 13, 22, 29, 36, 57, 79, 83, 113] ś has been studied inde-
pendently [5]. To improve tabular reasoning, transformer-based
approaches such as TAPAS [38] and Table-BERT [114] have been de-
veloped. However, these techniques perform poorly in SCITAB [57],
a dataset for scientific fact-checking on tables, with results barely
above random. One possible cause is the lack of contextual ground-
ing for tables, which are rarely self-contained. Similarly, figures and
charts in scientific papers often require surrounding textual expla-
nations for correct interpretation. Recent datasets like AVerImaTeC
[18] address image-text verification using web-sourced data, but sci-
entific domains present greater challenges: figures are densely struc-
tured, often span multiple sections, and require domain-specific
understanding capability.

We argue that scientific fact-checking should shift toward full-
paper analysis, where structured elements are interpreted along-
side their textual context. Unlike standalone multimodal models,
document-level processing enables cross-referencing between fig-
ures/tables/charts and their descriptions, facilitating more faithful
and complete verification.

RD.6. Multi-modal Evidence Alignment

Scientific fact-checking requires integrating evidence across multi-
ple modalities, including text, tables, and figures, to ensure consis-
tency and completeness. Recent multimodal information retrieval
datasets in the scientific domain [71, 110] provide aligned pairs
of textual and structured content, offering a foundation for cross-
modal reasoning. However, in real scientific documents, structured
elements, such as figures and tables, are not always located close
to their descriptive text, making alignment a non-trivial challenge.
A promising direction is to explicitly align structured elements
with their corresponding textual explanations within the same
document. In scientific articles, tables and figures are typically ex-
plained through captions or surrounding sentences. Layout-aware
parsing techniques can help identify these elements and link them
to relevant text spans. Once aligned, their contents can be jointly
encoded, enabling claim verification that draws on both structured
data and contextual text. This alignment facilitates more coherent
retrieval and reasoning across modalities, improving the reliability
of multimodal fact-checking.

This unified framework contrasts with traditional multimodal
systems that process each modality in isolation. Effective alignment
demands progress in scientific document understanding [24, 60],
visual structure parsing [32], and domain-specific retrieval [71].
Integrating these efforts will support full-document, multimodal
verification, where structured evidence is faithfully grounded in its
textual context.

7 Credibility of scientific literature
Existing evidence retrieval models often favour high-ranking doc-
uments based on semantic relevance, often overlooking scientific
rigour [98]. As a result, low-quality documents may be retrieved
as evidence, undermining the credibility of scientific fact-checking.
The reliability of a claim verification process is inherently tied to

443

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

the quality of the supporting literature, making evidence credibility
a crucial factor in scientific fact-checking.

In the domain of scientific literature, credibility assessment is
influenced by multiple factors, including peer-review status, which
ensures methodological scrutiny, citation impact which indicates
how influential a study is within its field, and experimental rigour,
reflecting the robustness of a study’s methodology. However, the
proliferation of non-peer-reviewed manuscripts and publications
from venues with varying editorial standards, particularly in open-
access repositories, poses a growing challenge. Such sources may
lack the rigorous methodological scrutiny necessary to ensure re-
liable scientific conclusions. Therefore, incorporating additional
quality indicators is essential for enhancing the robustness of sci-
entific fact-checking.

RD.7. Extending Indicators of Evidence Quality

Evaluating scientific literature quality extends beyond content
reliability, and should consider factors including venue reputa-
tion, methodological rigour, and experimental transparency. High-
impact journals and prestigious conferences generally enforce strin-
gent peer-review standards, contributing to the credibility of pub-
lished research. Similarly, the expertise and prior contributions of
an author, particularly in reputable venues, can provide further
insight into the credibility of a study. In addition to traditional
metadata, emerging indicators such as replication status, data avail-
ability, and adherence to reporting guidelines can further reflect
methodological soundness. Although metadata-based credibility
assessment is a useful heuristic, it is not foolproof. For example,
selective reporting and statistical manipulation still exist, as some
widely cited studies have later been retracted due to methodolog-
ical flaws [31]. While the integration of such metadata remains a
reasonable approach, as these indicators generally correlate with lit-
erature quality, their limitations must be acknowledged, given that
even widely cited studies can occasionally be subject to retraction
due to undetected methodological flaws.

8 Scientific terminology complexity
Scientific fact-checking systems often encounter challenges when
aligning claims with supporting evidence due to mismatches in
terminology granularity [101, 107]. The prevalence of hierarchical
and synonymous scientific terms introduces significant challenges
in fact verification. Many concepts exist at multiple levels of speci-
ficity, where broader categories encompass more specific subtypes,
leading to ambiguity in claim-evidence alignment. This issue is
further exacerbated by high token-level similarity among related
terms, making it difficult for models to differentiate between general
and specific concepts. As a result, models often misinterpret evi-
dence relevance, increasing the likelihood of incorrect verification
outcomes.

This issue arises when a claim uses a broad term, while the sup-
porting evidence provides a more specific instance, or vice versa.
As in the following example, such mismatches can lead to incor-
rect veracity assignments, as existing models struggle to recognise
hierarchical relationships between concepts.

Claim: Cancer risk is lower in individuals with a history of

alcohol consumption.

Supports: Alcohol consumption was associated with a decreased

risk of thyroid cancer.

This issue is common within scientific fact checking and has been
reported to occur within 44% of annotated examples in SciFact-Open
[101]. Hence, we argue that capturing the hierarchical relationship
could be a research direction to improve verification performance
in the scientific domain, by solving the mismatch problem.

RD.8. Hierarchical Concept Modelling

To address this issue, ontology-based reasoning can be integrated
into fact-checking pipelines. Structured ontologies such as MeSH
[54] and UMLS [53] define hierarchical relationships that help sys-
tems infer term specificity. Recognising that ‘lung cancer’ is a sub-
type of ‘cancer’ enables better claim-evidence alignment, mitigat-
ing errors caused by lexical similarity. Beyond that, the knowledge
graph can enrich ontological reasoning by encoding both hier-
archical and associative relationships among scientific concepts
[39, 48]. However, its potential for resolving terminology granular-
ity mismatches in scientific fact-checking remains unexplored. Ad-
ditionally, representation learning techniques such as contrastive
learning can embed these hierarchical relationships into vector
space representations, reducing reliance on token-level similarity.
Domain-specific models like SciBERT [11] and PubMedBERT [35]
can further enhance contextual understanding by incorporating
structured knowledge into retrieval and verification processes. By
leveraging ontological reasoning, knowledge graphs, and struc-
tured embeddings, scientific fact-checking systems can better align
claims with relevant evidence, reducing verification errors caused
by terminology granularity mismatches. In addition to enhancing
precision, this also has potential to improve interpretability by
making model decisions more transparent.

9 Conclusion
This paper explores the evolution of scientific fact-checking method-
ologies from abstract-level approaches to full-paper frameworks
on large-scale corpora. As the volume and diversity of scientific
knowledge continue to grow, the challenges of verifying claims
across heterogeneous sources become increasingly complex. By
addressing the complexities inherent in scientific literature, includ-
ing its evolving nature, structured format, and the necessity for
precise evidence retrieval, we underscore the importance of de-
veloping specialised retrieval systems capable of managing large,
multimodal, and time-sensitive evidence retrieval. Furthermore,
this work proposes several research directions aimed at improving
scientific fact-checking efficiency and reliability. They include the
integration of time-aware evidence retrieval to ensure the use of the
most relevant and up-to-date findings, adaptive document process-
ing to enable context-sensitive retrieval strategies, and multi-modal
evidence alignment to integrate text, tables and figures to enhance
verification accuracy. Overcoming these challenges has potential to
improve the accuracy of fact-checking processes and also facilitate
the scalability and applicability of these systems in diverse scientific
fact-checking scenarios. By bridging the gap between scientific fact-
checking and effective evidence retrieval, these advancements will
contribute to more robust, interpretable, and trustworthy scientific
fact-checking methodologies for real-world applications.

444

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

References

[1] Sahar Abdelnabi, Rakibul Hasan, and Mario Fritz. 2022. Open-domain, content-
based, multi-modal fact-checking of out-of-context images via online resources.
In Proceedings of the IEEE/CVF conference on computer vision and pattern recog-
nition. 14940ś14949.

[2] Vaibhav Adlakha, Parishad BehnamGhader, Xing Han Lu, Nicholas Meade,
and Siva Reddy. 2024. Evaluating Correctness and Faithfulness of Instruction-
Following Models for Question Answering. Transactions of the Association for
Computational Linguistics 12 (2024), 681ś699. doi:10.1162/tacl_a_00667

[3] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. 2022. PubHealthTab: A
Public Health Table-based Dataset for Evidence-based Fact Checking. In Findings
of the Association for Computational Linguistics: NAACL 2022, Marine Carpuat,
Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz (Eds.). Association
for Computational Linguistics, Seattle, United States, 1ś16. doi:10.18653/v1/
2022.findings-naacl.1

[4] Mubashara Akhtar, Oana Cocarascu, and Elena Simperl. 2023. Reading and
Reasoning over Chart Images for Evidence-based Automated Fact-Checking.
In Findings of the Association for Computational Linguistics: EACL 2023, An-
dreas Vlachos and Isabelle Augenstein (Eds.). Association for Computational
Linguistics, Dubrovnik, Croatia, 399ś414. doi:10.18653/v1/2023.findings-eacl.30
[5] Mubashara Akhtar, Michael Schlichtkrull, Zhijiang Guo, Oana Cocarascu, Elena
Simperl, and Andreas Vlachos. 2023. Multimodal Automated Fact-Checking:
A Survey. In Findings of the Association for Computational Linguistics: EMNLP
2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Com-
putational Linguistics, Singapore, 5430ś5448. doi:10.18653/v1/2023.findings-
emnlp.361

[6] Mubashara Akhtar, Michael Schlichtkrull, and Andreas Vlachos. 2024. Ev2R:
Evaluating Evidence Retrieval in Automated Fact-Checking. arXiv preprint
arXiv:2411.05375 (2024).

[7] Mubashara Akhtar, Nikesh Subedi, Vivek Gupta, Sahar Tahmasebi, Oana Co-
carascu, and Elena Simperl. 2024. ChartCheck: Explainable Fact-Checking
over Real-World Chart Images. In Findings of the Association for Computational
Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.).
Association for Computational Linguistics, Bangkok, Thailand, 13921ś13937.
doi:10.18653/v1/2024.findings-acl.828

[8] Liesbeth Allein, Marlon Saelens, Ruben Cartuyvels, and Marie-Francine Moens.
2023. Implicit Temporal Reasoning for Evidence-Based Fact-Checking. In Find-
ings of the Association for Computational Linguistics: EACL 2023, Andreas Vla-
chos and Isabelle Augenstein (Eds.). Association for Computational Linguistics,
Dubrovnik, Croatia, 176ś189. doi:10.18653/v1/2023.findings-eacl.13

[9] Isabelle Augenstein, Christina Lioma, Dongsheng Wang, Lucas Chaves Lima,
Casper Hansen, Christian Hansen, and Jakob Grue Simonsen. 2019. MultiFC: A
Real-World Multi-Domain Dataset for Evidence-Based Fact Checking of Claims.
In Proceedings of the 2019 Conference on Empirical Methods in Natural Language
Processing and the 9th International Joint Conference on Natural Language Pro-
cessing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan
(Eds.). Association for Computational Linguistics, Hong Kong, China, 4685ś4697.
doi:10.18653/v1/D19-1475

[10] Dara Bahri, Yi Tay, Che Zheng, Donald Metzler, and Andrew Tomkins. 2020.
Choppy: Cut transformer for ranked list truncation. In Proceedings of the 43rd
International ACM SIGIR Conference on Research and Development in Information
Retrieval. 1513ś1516.

[11] Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A Pretrained Language
Model for Scientific Text. In Proceedings of the 2019 Conference on Empirical
Methods in Natural Language Processing and the 9th International Joint Confer-
ence on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang,
Vincent Ng, and Xiaojun Wan (Eds.). Association for Computational Linguistics,
Hong Kong, China, 3615ś3620. doi:10.18653/v1/D19-1371

[12] Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The

Long-Document Transformer. arXiv:2004.05150 (2020).

[13] Chandra Sekhar Bhagavatula, Thanapon Noraset, and Doug Downey. 2013.
Methods for exploring and mining tables on wikipedia. In Proceedings of the
ACM SIGKDD workshop on interactive data exploration and analytics. 18ś26.

[14] Reem Bin-Hezam and Mark Stevenson. 2023. Combining Counting Processes
and Classification Improves a Stopping Rule for Technology Assisted Review.
In Findings of the Association for Computational Linguistics: EMNLP 2023, Houda
Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational
Linguistics, Singapore, 2603ś2609. doi:10.18653/v1/2023.findings-emnlp.171

[15] Reem Bin-Hezam and Mark Stevenson. 2024. RLStop: A Reinforcement Learning
Stopping Method for TAR. In Proceedings of the 47th International ACM SIGIR
Conference on Research and Development in Information Retrieval. 2604ś2608.

[16] Jannis Bulian, Jordan Boyd-Graber, Markus Leippold, Massimiliano Ciaramita,
and Thomas Diggelmann. 2020. Climate-fever: A dataset for verification of
real-world climate claims. In NeurIPS 2020 Workshop on Tackling Climate Change
with Machine Learning.

[17] Peter Buneman, Dennis Dosso, Matteo Lissandrini, and Gianmaria Silvello. 2021.
Data citation and the citation graph. Quantitative Science Studies 2, 4 (2021),

1399ś1422.

[18] Rui Cao, Zifeng Ding, Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos.
2025. AVerImaTeC: A Dataset for Automatic Verification of Image-Text Claims
with Evidence from the Web. arXiv:2505.17978 [cs.CL] https://arxiv.org/abs/
2505.17978

[19] Catherine Chen, Zejiang Shen, Dan Klein, Gabriel Stanovsky, Doug Downey,
and Kyle Lo. 2023. Are Layout-Infused Language Models Robust to Layout
Distribution Shifts? A Case Study with Scientific Documents. In Findings of the
Association for Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-
Graber, and Naoaki Okazaki (Eds.). Association for Computational Linguistics,
Toronto, Canada, 13345ś13360. doi:10.18653/v1/2023.findings-acl.844

[20] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2024. Benchmarking large
language models in retrieval-augmented generation. In Proceedings of the AAAI
Conference on Artificial Intelligence, Vol. 38. 17754ś17762.

[21] Qingyu Chen, Yifan Peng, and Zhiyong Lu. 2019. BioSentVec: creating sen-
tence embeddings for biomedical texts. In 2019 IEEE International Conference on
Healthcare Informatics (ICHI). IEEE, 1ś5.

[22] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang,
Shiyang Li, Xiyou Zhou, and William Yang Wang. 2020. TabFact : A Large-scale
Dataset for Table-based Fact Verification. In International Conference on Learning
Representations (ICLR). Addis Ababa, Ethiopia.

[23] Sabrina Chiesurin, Dimitris Dimakopoulos, Marco Antonio Sobrevilla Cabezudo,
Arash Eshghi, Ioannis Papaioannou, Verena Rieser, and Ioannis Konstas. 2023.
The Dangers of trusting Stochastic Parrots: Faithfulness and Trust in Open-
domain Conversational Question Answering. In Findings of the Association for
Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and
Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto,
Canada, 947ś959. doi:10.18653/v1/2023.findings-acl.60

[24] Jaemin Cho, Debanjan Mahata, Ozan Irsoy, Yujie He, and Mohit Bansal. 2024.
M3docrag: Multi-modal retrieval is what you need for multi-page multi-
document understanding. arXiv preprint arXiv:2411.04952 (2024).

[25] Gordon V Cormack and Maura R Grossman. 2014. Evaluation of machine-
learning protocols for technology-assisted review in electronic discovery. In
Proceedings of the 37th international ACM SIGIR conference on Research & devel-
opment in information retrieval. 153ś162.

[26] Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021.
Overview of the TREC 2020 deep learning track. arXiv:2102.07662 [cs.IR]
https://arxiv.org/abs/2102.07662

[27] Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M
Voorhees. 2020. Overview of the TREC 2019 deep learning track. arXiv preprint
arXiv:2003.07820 (2020).

[28] Alphaeus Dmonte, Roland Oruche, Marcos Zampieri, Prasad Calyam, and Is-
abelle Augenstein. 2024. Claim Verification in the Age of Large Language
Models: A Survey. arXiv preprint arXiv:2408.14317 (2024).

[29] Haoyu Dong and Zhiruo Wang. 2024. Large language models for tabular data:
Progresses and future directions. In Proceedings of the 47th International ACM
SIGIR Conference on Research and Development in Information Retrieval. 2997ś
3000.

[30] Michael Färber and Adam Jatowt. 2020. Citation recommendation: approaches

and datasets. International Journal on Digital Libraries 21, 4 (2020), 375ś405.

[31] Aaron HA Fletcher and Mark Stevenson. 2025. Predicting retracted research: a
dataset and machine learning approaches. Research Integrity and Peer Review
10, 1 (2025), 1ś10.

[32] Raymond Fok, Hita Kambhamettu, Luca Soldaini, Jonathan Bragg, Kyle Lo, Marti
Hearst, Andrew Head, and Daniel S Weld. 2023. Scim: Intelligent skimming
support for scientific papers. In Proceedings of the 28th International Conference
on Intelligent User Interfaces. 476ś490.

[33] Maura R Grossman, Gordon V Cormack, and Adam Roegiest. 2016. TREC 2016

Total Recall Track Overview.. In TREC.

[34] Nianlong Gu, Yingqiang Gao, and Richard HR Hahnloser. 2022. Local citation
recommendation with hierarchical-attention text encoder and scibert-based
reranking. In European conference on information retrieval. Springer, 274ś288.
[35] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong
Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2020. Domain-Specific
Language Model Pretraining for Biomedical Natural Language Processing.
arXiv:arXiv:2007.15779

[36] Zihui Gu, Ju Fan, Nan Tang, Preslav Nakov, Xiaoman Zhao, and Xiaoyong Du.
2022. PASTA: Table-Operations Aware Fact Verification via Sentence-Table
Cloze Pre-training. In Proceedings of the 2022 Conference on Empirical Methods
in Natural Language Processing, Yoav Goldberg, Zornitsa Kozareva, and Yue
Zhang (Eds.). Association for Computational Linguistics, Abu Dhabi, United
Arab Emirates, 4971ś4983. doi:10.18653/v1/2022.emnlp-main.331

[37] Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. A survey on
automated fact-checking. Transactions of the Association for Computational
Linguistics 10 (2022), 178ś206.

[38] Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Müller, Francesco Piccinno,
and Julian Eisenschlos. 2020. TaPas: Weakly Supervised Table Parsing via
Pre-training. In Proceedings of the 58th Annual Meeting of the Association for

445

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

Computational Linguistics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel
Tetreault (Eds.). Association for Computational Linguistics, Online, 4320ś4333.
doi:10.18653/v1/2020.acl-main.398

[39] Jason Hoelscher-Obermaier, Edward Stevinson, Valentin Stauber, Ivaylo Zhelev,
Viktor Botev, Ronin Wu, and Jeremy Minton. 2022. Leveraging knowledge
graphs to update scientific word embeddings using latent semantic imputation.
In Proceedings of the first Workshop on Information Extraction from Scientific
Publications, Tirthankar Ghosal, Sergi Blanco-Cuaresma, Alberto Accomazzi,
Robert M. Patton, Felix Grezes, and Thomas Allen (Eds.). Association for Com-
putational Linguistics, Online, 43ś53. doi:10.18653/v1/2022.wiesp-1.6

[40] Xuming Hu, Zhaochen Hong, Zhijiang Guo, Lijie Wen, and Philip Yu. 2023. Read
it twice: Towards faithfully interpretable fact verification by revisiting evidence.
In Proceedings of the 46th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 2319ś2323.

[41] Po-Wei Huang, Abhinav Ramesh Kashyap, Yanxia Qin, Yajing Yang, and Min-Yen
Kan. 2022. Lightweight Contextual Logical Structure Recovery. In Proceedings
of the Third Workshop on Scholarly Document Processing, Arman Cohan, Guy
Feigenblat, Dayne Freitag, Tirthankar Ghosal, Drahomira Herrmannova, Petr
Knoth, Kyle Lo, Philipp Mayr, Michal Shmueli-Scheuer, Anita de Waard, and
Lucy Lu Wang (Eds.). Association for Computational Linguistics, Gyeongju,
Republic of Korea, 37ś48. https://aclanthology.org/2022.sdp-1.5/

[42] Tomoki Ikoma and Shigeki Matsubara. 2023. On the Use of Language Models for
Function Identification of Citations in Scholarly Papers. In Proceedings of the Sec-
ond Workshop on Information Extraction from Scientific Publications, Tirthankar
Ghosal, Felix Grezes, Thomas Allen, Kelly Lockhart, Alberto Accomazzi, and
Sergi Blanco-Cuaresma (Eds.). Association for Computational Linguistics, Bali,
Indonesia, 130ś135. doi:10.18653/v1/2023.wiesp-1.15

[43] E. Kanoulas, Dan Li, Leif Azzopardi, and René Spijker. 2018. CLEF 2017 Tech-
nologically Assisted Reviews in Empirical Medicine Overview. In Conference
and Labs of the Evaluation Forum. https://api.semanticscholar.org/CorpusID:
246440739

[44] Evangelos Kanoulas, Dan Li, Leif Azzopardi, and Rene Spijker. 2018. CLEF
2018 technologically assisted reviews in empirical medicine overview. In CEUR
workshop proceedings, Vol. 2125.

[45] Evangelos Kanoulas, Dan Li, Leif Azzopardi, and Rene Spijker. 2019. CLEF 2019
technology assisted reviews in empirical medicine overview. In CEUR workshop
proceedings, Vol. 2380. 250.

[46] Marzena Karpinska, Katherine Thai, Kyle Lo, Tanya Goyal, and Mohit Iyyer.
2024. One Thousand and One Pairs: A “novelž challenge for long-context
language models. In Proceedings of the 2024 Conference on Empirical Methods in
Natural Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung
Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA,
17048ś17085. doi:10.18653/v1/2024.emnlp-main.948

[47] Mohammed Abdul Khaliq, Paul Yu-Chun Chang, Mingyang Ma, Bernhard
Pflugfelder, and Filip Miletić. 2024. RAGAR, Your Falsehood Radar: RAG-
Augmented Reasoning for Political Fact-Checking using Multimodal Large
Language Models. In Proceedings of the Seventh Fact Extraction and VERifica-
tion Workshop (FEVER), Michael Schlichtkrull, Yulong Chen, Chenxi White-
house, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos
Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas
Vlachos (Eds.). Association for Computational Linguistics, Miami, Florida, USA,
280ś296. doi:10.18653/v1/2024.fever-1.29

[48] Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James Thorne, and Edward
Choi. 2023. FactKG: Fact Verification via Reasoning on Knowledge Graphs.
In Proceedings of the 61st Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), Anna Rogers, Jordan Boyd-Graber, and
Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto,
Canada, 16190ś16206. doi:10.18653/v1/2023.acl-long.895

[49] Neema Kotonya and Francesca Toni. 2020. Explainable Automated Fact-
Checking for Public Health Claims. In Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP), Bonnie Webber,
Trevor Cohn, Yulan He, and Yang Liu (Eds.). Association for Computational
Linguistics, Online, 7740ś7754. doi:10.18653/v1/2020.emnlp-main.623

[50] Xiangci Li, Gully A Burns, and Nanyun Peng. 2021. A Paragraph-level Multi-task

Learning Model for Scientific Fact-Verification.. In SDU@ AAAI.

[51] Xiaoxi Li, Zhicheng Dou, Yujia Zhou, and Fangchao Liu. 2024. Corpuslm:
Towards a unified language model on corpus for knowledge-intensive tasks.
In Proceedings of the 47th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 26ś37.

[52] Yen-Chieh Lien, Daniel Cohen, and W Bruce Croft. 2019. An assumption-free
approach to the dynamic truncation of ranked lists. In Proceedings of the 2019
ACM SIGIR International Conference on Theory of Information Retrieval. 79ś82.
[53] Donald AB Lindberg, Betsy L Humphreys, and Alexa T McCray. 1993. The
unified medical language system. Yearbook of medical informatics 2, 01 (1993),
41ś51.

[54] Carolyn E Lipscomb. 2000. Medical subject headings (MeSH). Bulletin of the

Medical Library Association 88, 3 (2000), 265.

[55] Qi Liu, Bo Wang, Nan Wang, and Jiaxin Mao. 2025. Leveraging Passage Em-
beddings for Efficient Listwise Reranking with Large Language Models. In Pro-
ceedings of the ACM on Web Conference 2025 (Sydney NSW, Australia) (WWW
’25). Association for Computing Machinery, New York, NY, USA, 4274ś4283.
doi:10.1145/3696410.3714554

[56] Kyle Lo, Joseph Chee Chang, Andrew Head, Jonathan Bragg, Amy X. Zhang, Cas-
sidy Trier, Chloe Anastasiades, Tal August, Russell Authur, Danielle Bragg, Erin
Bransom, Isabel Cachola, Stefan Candra, Yoganand Chandrasekhar, Yen-Sung
Chen, Evie Yu-Yen Cheng, Yvonne Chou, Doug Downey, Rob Evans, Raymond
Fok, Fangzhou Hu, Regan Huff, Dongyeop Kang, Tae Soo Kim, Rodney Kinney,
Aniket Kittur, Hyeonsu B. Kang, Egor Klevak, Bailey Kuehl, Michael J. Lan-
gan, Matt Latzke, Jaron Lochner, Kelsey MacMillan, Eric Marsh, Tyler Murray,
Aakanksha Naik, Ngoc-Uyen Nguyen, Srishti Palani, Soya Park, Caroline Paulic,
Napol Rachatasumrit, Smita Rao, Paul Sayre, Zejiang Shen, Pao Siangliulue,
Luca Soldaini, Huy Tran, Madeleine van Zuylen, Lucy Lu Wang, Christopher
Wilhelm, Caroline Wu, Jiangjiang Yang, Angele Zamarron, Marti A. Hearst, and
Daniel S. Weld. 2024. The Semantic Reader Project. Commun. ACM 67, 10 (Sept.
2024), 50ś61. doi:10.1145/3659096

[57] Xinyuan Lu, Liangming Pan, Qian Liu, Preslav Nakov, and Min-Yen Kan. 2023.
SCITAB: A Challenging Benchmark for Compositional Reasoning and Claim
Verification on Scientific Tables. In Proceedings of the 2023 Conference on Em-
pirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino,
and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore,
7787ś7813. doi:10.18653/v1/2023.emnlp-main.483

[58] María Luengo and David García-Marín. 2020. The performance of truth: politi-
cians, fact-checking journalism, and the struggle to tackle COVID-19 misinfor-
mation. American Journal of Cultural Sociology 8, 3 (2020), 405.

[59] Yixiao Ma, Qingyao Ai, Yueyue Wu, Yunqiu Shao, Yiqun Liu, Min Zhang, and
Shaoping Ma. 2022. Incorporating retrieval information into the truncation
of ranking lists for better legal search. In Proceedings of the 45th International
ACM SIGIR Conference on Research and Development in Information Retrieval.
438ś448.

[60] Yubo Ma, Yuhang Zang, Liangyu Chen, Meiqi Chen, Yizhu Jiao, Xinze Li,
Xinyuan Lu, Ziyu Liu, Yan Ma, Xiaoyi Dong, Pan Zhang, Liangming Pan, Yu-
Gang Jiang, Jiaqi Wang, Yixin Cao, and Aixin Sun. 2024. MMLongBench-Doc:
Benchmarking Long-context Document Understanding with Visualizations.
arXiv:2407.01523 [cs.CV] https://arxiv.org/abs/2407.01523

[61] Chuan Meng, Negar Arabzadeh, Arian Askari, Mohammad Aliannejadi, and
Maarten de Rijke. 2024. Ranked list truncation for large language model-based
re-ranking. In Proceedings of the 47th International ACM SIGIR Conference on
Research and Development in Information Retrieval. 141ś151.

[62] Isabelle Mohr, Amelie Wührl, and Roman Klinger. 2022. CoVERT: A Corpus
of Fact-checked Biomedical COVID-19 Tweets. In Proceedings of the Thirteenth
Language Resources and Evaluation Conference, Nicoletta Calzolari, Frédéric
Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck,
Sara Goggi, Hitoshi Isahara, Bente Maegaard, Joseph Mariani, Hélène Mazo, Jan
Odijk, and Stelios Piperidis (Eds.). European Language Resources Association,
Marseille, France, 244ś257. https://aclanthology.org/2022.lrec-1.26

[63] Yuki Momii, Tetsuya Takiguchi, and Yasuo Ariki. 2024. RAG-Fusion Based
Information Retrieval for Fact-Checking. In Proceedings of the Seventh Fact Ex-
traction and VERification Workshop (FEVER), Michael Schlichtkrull, Yulong Chen,
Chenxi Whitehouse, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang
Guo, Christos Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne,
and Andreas Vlachos (Eds.). Association for Computational Linguistics, Miami,
Florida, USA, 47ś54. doi:10.18653/v1/2024.fever-1.4

[64] Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. 2020. Docu-
ment Ranking with a Pretrained Sequence-to-Sequence Model. In Findings of the
Association for Computational Linguistics: EMNLP 2020, Trevor Cohn, Yulan He,
and Yang Liu (Eds.). Association for Computational Linguistics, Online, 708ś718.
doi:10.18653/v1/2020.findings-emnlp.63

[65] Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and
Panagiotis Petrantonakis. 2023. Synthetic misinformers: Generating and com-
bating multimodal misinformation. In Proceedings of the 2nd ACM International
Workshop on Multimedia AI against Disinformation. 36ś44.

[66] Stefanos-Iordanis Papadopoulos, Christos Koutlis, Symeon Papadopoulos, and
Panagiotis C Petrantonakis. 2024. VERITE: a Robust benchmark for multimodal
misinformation detection accounting for unimodal bias. International Journal
of Multimedia Information Retrieval 13, 1 (2024), 4.

[67] Andrew Parry, Debasis Ganguly, and Manish Chandra. 2024. In-Context Learn-
ing" or: How I learned to stop worrying and love" Applied Information Retrieval.
In Proceedings of the 47th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 14ś25.

[68] Ronak Pradeep, Xueguang Ma, Rodrigo Nogueira, and Jimmy Lin. 2021. Scientific
Claim Verification with VerT5erini. In Proceedings of the 12th International
Workshop on Health Text Mining and Information Analysis, Eben Holderness,
Antonio Jimeno Yepes, Alberto Lavelli, Anne-Lyse Minard, James Pustejovsky,
and Fabio Rinaldi (Eds.). Association for Computational Linguistics, online,
94ś103. https://aclanthology.org/2021.louhi-1.11/

446

ICTIR ’25, July 18, 2025, Padua, Italy

Xingyu Deng, Xi Wang, Mark Stevenson

[69] Peng Qi, Zehong Yan, Wynne Hsu, and Mong Li Lee. 2024. SNIFFER: Multi-
modal Large Language Model for Explainable Out-of-Context Misinformation
Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition. 13052ś13062.

[70] Vatsal Raina and Mark Gales. 2024. Question-Based Retrieval using Atomic
Units for Enterprise RAG. In Proceedings of the Seventh Fact Extraction and VER-
ification Workshop (FEVER), Michael Schlichtkrull, Yulong Chen, Chenxi White-
house, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos
Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas
Vlachos (Eds.). Association for Computational Linguistics, Miami, Florida, USA,
219ś233. doi:10.18653/v1/2024.fever-1.25

[71] Jonathan Roberts, Kai Han, Neil Houlsby, and Samuel Albanie. 2024. SciFIBench:
Benchmarking Large Multimodal Models for Scientific Figure Interpretation.
arXiv preprint arXiv:2405.08807 (2024).

[72] Arkadiy Saakyan, Tuhin Chakrabarty, and Smaranda Muresan. 2021. COVID-
Fact: Fact Extraction and Verification of Real-World Claims on COVID-19 Pan-
demic. In Proceedings of the 59th Annual Meeting of the Association for Com-
putational Linguistics and the 11th International Joint Conference on Natural
Language Processing (Volume 1: Long Papers), Chengqing Zong, Fei Xia, Wen-
jie Li, and Roberto Navigli (Eds.). Association for Computational Linguistics,
Online, 2116ś2129. doi:10.18653/v1/2021.acl-long.165

[73] Alireza Salemi and Hamed Zamani. 2024. Evaluating retrieval quality in retrieval-
augmented generation. In Proceedings of the 47th International ACM SIGIR Con-
ference on Research and Development in Information Retrieval. 2395ś2400.
[74] Alireza Salemi and Hamed Zamani. 2024. Towards a search engine for machines:
Unified ranking for multiple retrieval-augmented large language models. In
Proceedings of the 47th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 741ś751.

[75] Mourad Sarrouti, Asma Ben Abacha, Yassine Mrabet, and Dina Demner-
Fushman. 2021. Evidence-based Fact-Checking of Health-related Claims. In
Findings of the Association for Computational Linguistics: EMNLP 2021, Marie-
Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.).
Association for Computational Linguistics, Punta Cana, Dominican Republic,
3499ś3512. doi:10.18653/v1/2021.findings-emnlp.297

[76] Artsiom Sauchuk, James Thorne, Alon Halevy, Nicola Tonellotto, and Fabrizio
Silvestri. 2022. On the role of relevance in natural language processing tasks.
In Proceedings of the 45th International ACM SIGIR Conference on Research and
Development in Information Retrieval. 1785ś1789.

[77] Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng,
Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana
Cocarascu, Arpit Mittal, James Thorne, and Andreas Vlachos. 2024. The Auto-
mated Verification of Textual Claims (AVeriTeC) Shared Task. In Proceedings
of the Seventh Fact Extraction and VERification Workshop (FEVER), Michael
Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng, Mubashara
Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana Cocarascu,
Arpit Mittal, James Thorne, and Andreas Vlachos (Eds.). Association for Compu-
tational Linguistics, Miami, Florida, USA, 1ś26. doi:10.18653/v1/2024.fever-1.1
[78] Michael Schlichtkrull, Zhijiang Guo, and Andreas Vlachos. 2024. Averitec: A
dataset for real-world claim verification with evidence from the web. Advances
in Neural Information Processing Systems 36 (2024).

[79] Michael Sejr Schlichtkrull, Vladimir Karpukhin, Barlas Oguz, Mike Lewis, Wen-
tau Yih, and Sebastian Riedel. 2021. Joint Verification and Reranking for Open
Fact Checking Over Tables. In Proceedings of the 59th Annual Meeting of the
Association for Computational Linguistics and the 11th International Joint Confer-
ence on Natural Language Processing (Volume 1: Long Papers), Chengqing Zong,
Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational
Linguistics, Online, 6787ś6799. doi:10.18653/v1/2021.acl-long.529

[80] Özge Sevgili, Irina Nikishina, Seid Muhie Yimam, Martin Semmann, and
Chris Biemann. 2024. UHH at AVeriTeC: RAG for Fact-Checking with Real-
World Claims. In Proceedings of the Seventh Fact Extraction and VERifica-
tion Workshop (FEVER), Michael Schlichtkrull, Yulong Chen, Chenxi White-
house, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos
Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas
Vlachos (Eds.). Association for Computational Linguistics, Miami, Florida, USA,
55ś63. doi:10.18653/v1/2024.fever-1.5

[81] Zejiang Shen, Kyle Lo, Lucy Lu Wang, Bailey Kuehl, Daniel S. Weld, and Doug
Downey. 2022. VILA: Improving Structured Content Extraction from Scien-
tific PDFs Using Visual Layout Groups. Transactions of the Association for
Computational Linguistics 10 (2022), 376ś392. doi:10.1162/tacl_a_00466
[82] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H
Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be
easily distracted by irrelevant context. In International Conference on Machine
Learning. PMLR, 31210ś31227.

[83] Qi Shi, Yu Zhang, Qingyu Yin, and Ting Liu. 2021. Logic-level Evidence Retrieval
and Graph-based Verification Network for Table-based Fact Verification. In
Proceedings of the 2021 Conference on Empirical Methods in Natural Language
Processing, Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott
Wen-tau Yih (Eds.). Association for Computational Linguistics, Online and Punta

Cana, Dominican Republic, 175ś184. doi:10.18653/v1/2021.emnlp-main.16
[84] Ronit Singal, Pransh Patwa, Parth Patwa, Aman Chadha, and Amitava Das.
2024. Evidence-backed Fact Checking using RAG and Few-Shot In-Context
Learning with LLMs. In Proceedings of the Seventh Fact Extraction and VERifi-
cation Workshop (FEVER), Michael Schlichtkrull, Yulong Chen, Chenxi White-
house, Zhenyun Deng, Mubashara Akhtar, Rami Aly, Zhijiang Guo, Christos
Christodoulopoulos, Oana Cocarascu, Arpit Mittal, James Thorne, and Andreas
Vlachos (Eds.). Association for Computational Linguistics, Miami, Florida, USA,
91ś98. doi:10.18653/v1/2024.fever-1.10

[85] Mark Stevenson and Reem Bin-Hezam. 2023. Stopping Methods for Technology-
assisted Reviews Based on Point Processes. ACM Transactions on Information
Systems 42, 3 (2023), 1ś37.

[86] Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin
Chen, Dawei Yin, and Zhaochun Ren. 2023.
Is ChatGPT Good at Search?
Investigating Large Language Models as Re-Ranking Agents. In Proceedings of
the 2023 Conference on Empirical Methods in Natural Language Processing, Houda
Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational
Linguistics, Singapore, 14918ś14937. doi:10.18653/v1/2023.emnlp-main.923
[87] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito,
and Kuniko Saito. 2023. Slidevqa: A dataset for document visual question
answering on multiple images. In Proceedings of the AAAI Conference on Artificial
Intelligence, Vol. 37. 13636ś13645.

[88] Liyan Tang, Philippe Laban, and Greg Durrett. 2024. MiniCheck: Efficient Fact-
Checking of LLMs on Grounding Documents. In Proceedings of the 2024 Confer-
ence on Empirical Methods in Natural Language Processing, Yaser Al-Onaizan,
Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Lin-
guistics, Miami, Florida, USA, 8818ś8847. doi:10.18653/v1/2024.emnlp-main.499
[89] Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna
Gurevych. 2021. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of
Information Retrieval Models. In Thirty-fifth Conference on Neural Information
Processing Systems Datasets and Benchmarks Track (Round 2). https://openreview.
net/forum?id=wCu6T5xFjeJ

[90] James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal.
2018. FEVER: a Large-scale Dataset for Fact Extraction and VERification. In Pro-
ceedings of the 2018 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Volume 1 (Long Pa-
pers), Marilyn Walker, Heng Ji, and Amanda Stent (Eds.). Association for Compu-
tational Linguistics, New Orleans, Louisiana, 809ś819. doi:10.18653/v1/N18-1074
[91] James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopou-
los, and Arpit Mittal. 2018. The Fact Extraction and VERification (FEVER)
Shared Task. In Proceedings of the First Workshop on Fact Extraction and VER-
ification (FEVER), James Thorne, Andreas Vlachos, Oana Cocarascu, Christos
Christodoulopoulos, and Arpit Mittal (Eds.). Association for Computational
Linguistics, Brussels, Belgium, 1ś9. doi:10.18653/v1/W18-5501

[92] Fangzheng Tian, Debasis Ganguly, and Craig Macdonald. 2025. Is Relevance
Propagated from Retriever to Generator in RAG?. In Advances in Information
Retrieval: 47th European Conference on Information Retrieval, ECIR 2025, Lucca,
Italy, April 6ś10, 2025, Proceedings, Part I (Lucca, Italy). Springer-Verlag, Berlin,
Heidelberg, 32ś48. doi:10.1007/978-3-031-88708-6_3

[93] Jonathan Tonglet, Marie-Francine Moens, and Iryna Gurevych. 2024. “Image,
Tell me your story!ž Predicting the original meta-context of visual misinfor-
mation. In Proceedings of the 2024 Conference on Empirical Methods in Natural
Language Processing, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen
(Eds.). Association for Computational Linguistics, Miami, Florida, USA, 7845ś
7864. doi:10.18653/v1/2024.emnlp-main.448

[94] Herbert Ullrich, Tomáš Mlynář, and Jan Drchal. 2024. AIC CTU system at
AVeriTeC: Re-framing automated fact-checking as a simple RAG task. In Proceed-
ings of the Seventh Fact Extraction and VERification Workshop (FEVER), Michael
Schlichtkrull, Yulong Chen, Chenxi Whitehouse, Zhenyun Deng, Mubashara
Akhtar, Rami Aly, Zhijiang Guo, Christos Christodoulopoulos, Oana Cocarascu,
Arpit Mittal, James Thorne, and Andreas Vlachos (Eds.). Association for Compu-
tational Linguistics, Miami, Florida, USA, 137ś150. doi:10.18653/v1/2024.fever-
1.16

[95] Jordy Van Landeghem, Rubèn Tito, Łukasz Borchmann, Michał Pietruszka, Pawel
Joziak, Rafal Powalski, Dawid Jurkiewicz, Mickaël Coustaty, Bertrand Anckaert,
Ernest Valveny, et al. 2023. Document understanding dataset and evaluation
(dude). In Proceedings of the IEEE/CVF International Conference on Computer
Vision. 19528ś19540.

[96] Vijay Viswanathan, Graham Neubig, and Pengfei Liu. 2021. CitationIE: Lever-
aging the Citation Graph for Scientific Information Extraction. In Proceedings
of the 59th Annual Meeting of the Association for Computational Linguistics
and the 11th International Joint Conference on Natural Language Processing
(Volume 1: Long Papers), Chengqing Zong, Fei Xia, Wenjie Li, and Roberto
Navigli (Eds.). Association for Computational Linguistics, Online, 719ś731.
doi:10.18653/v1/2021.acl-long.59

[97] Juraj Vladika and Florian Matthes. 2023. Scientific Fact-Checking: A Survey
of Resources and Approaches. In Findings of the Association for Computational
Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki

447

The Next Phase of Scientific Fact-Checking: Advanced Evidence Retrieval from Complex Structured Academic Papers

ICTIR ’25, July 18, 2025, Padua, Italy

SIGIR Conference on Research and Development in Information Retrieval. 2733ś
2743.

[113] Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. 2023.
Large language models are versatile decomposers: Decomposing evidence and
questions for table-based reasoning. In Proceedings of the 46th International
ACM SIGIR Conference on Research and Development in Information Retrieval.
174ś184.

[114] Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020.
TaBERT: Pretraining for Joint Understanding of Textual and Tabular Data. In Pro-
ceedings of the 58th Annual Meeting of the Association for Computational Linguis-
tics, Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault (Eds.). Associa-
tion for Computational Linguistics, Online, 8413ś8426. doi:10.18653/v1/2020.acl-
main.745

[115] Hamed Zamani and Michael Bendersky. 2024. Stochastic rag: End-to-end
retrieval-augmented generation through expected utility maximization. In Pro-
ceedings of the 47th International ACM SIGIR Conference on Research and Devel-
opment in Information Retrieval. 2641ś2646.

[116] Xia Zeng, Amani S Abumansour, and Arkaitz Zubiaga. 2021. Automated fact-

checking: A survey. Language and Linguistics Compass 15, 10 (2021), e12438.

[117] Hengran Zhang, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Yixing Fan, and
Xueqi Cheng. 2023. From Relevance to Utility: Evidence Retrieval with Feedback
for Fact Verification. In Findings of the Association for Computational Linguistics:
EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association
for Computational Linguistics, Singapore, 6373ś6384. doi:10.18653/v1/2023.
findings-emnlp.422

[118] Tianmai M Zhang and Neil F Abernethy. 2024. Detecting Reference Errors in Sci-
entific Literature with Large Language Models. arXiv preprint arXiv:2411.06101
(2024).

[119] Zhiwei Zhang, Jiyi Li, Fumiyo Fukumoto, and Yanming Ye. 2021. Abstract, Ra-
tionale, Stance: A Joint Model for Scientific Claim Verification. In Proceedings of
the 2021 Conference on Empirical Methods in Natural Language Processing, Marie-
Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.).
Association for Computational Linguistics, Online and Punta Cana, Dominican
Republic, 3580ś3586. doi:10.18653/v1/2021.emnlp-main.290

[120] Liwen Zheng, Chaozhuo Li, Xi Zhang, Yu-Ming Shang, Feiran Huang, and
Haoran Jia. 2024. Evidence Retrieval is almost All You Need for Fact Verifi-
cation. In Findings of the Association for Computational Linguistics ACL 2024,
Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Com-
putational Linguistics, Bangkok, Thailand and virtual meeting, 9274ś9281.
https://aclanthology.org/2024.findings-acl.551

(Eds.). Association for Computational Linguistics, Toronto, Canada, 6215ś6230.
doi:10.18653/v1/2023.findings-acl.387

[98] Juraj Vladika and Florian Matthes. 2024. Comparing Knowledge Sources for
Open-Domain Scientific Claim Verification. In Proceedings of the 18th Conference
of the European Chapter of the Association for Computational Linguistics (Volume
1: Long Papers), Yvette Graham and Matthew Purver (Eds.). Association for
Computational Linguistics, St. Julian’s, Malta, 2103ś2114. https://aclanthology.
org/2024.eacl-long.128/

[99] Juraj Vladika and Florian Matthes. 2024. Improving Health Question Answering
with Reliable and Time-Aware Evidence Retrieval. In Findings of the Association
for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and
Steven Bethard (Eds.). Association for Computational Linguistics, Mexico City,
Mexico, 4752ś4763. doi:10.18653/v1/2024.findings-naacl.295

[100] David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen,
Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or Fiction: Verifying Sci-
entific Claims. In Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing (EMNLP), Bonnie Webber, Trevor Cohn, Yulan
He, and Yang Liu (Eds.). Association for Computational Linguistics, Online,
7534ś7550. doi:10.18653/v1/2020.emnlp-main.609

[101] David Wadden, Kyle Lo, Bailey Kuehl, Arman Cohan, Iz Beltagy, Lucy Lu Wang,
and Hannaneh Hajishirzi. 2022. SciFact-Open: Towards open-domain scientific
claim verification. In Findings of the Association for Computational Linguistics:
EMNLP 2022, Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (Eds.). As-
sociation for Computational Linguistics, Abu Dhabi, United Arab Emirates,
4719ś4734. doi:10.18653/v1/2022.findings-emnlp.347

[102] David Wadden, Kyle Lo, Lucy Lu Wang, Arman Cohan, Iz Beltagy, and Han-
naneh Hajishirzi. 2022. MultiVerS: Improving scientific claim verification with
weak supervision and full-document context. In Findings of the Association
for Computational Linguistics: NAACL 2022, Marine Carpuat, Marie-Catherine
de Marneffe, and Ivan Vladimir Meza Ruiz (Eds.). Association for Computational
Linguistics, Seattle, United States, 61ś76. doi:10.18653/v1/2022.findings-naacl.6
[103] Dong Wang, Jianxin Li, Tianchen Zhu, Haoyi Zhou, Qishan Zhu, Yuxin Wen,
and Hongming Piao. 2022. MtCut: A Multi-Task Framework for Ranked List
Truncation. In Proceedings of the Fifteenth ACM International Conference on Web
Search and Data Mining. 1054ś1062.

[104] Gengyu Wang, Kate Harwood, Lawrence Chillrud, Amith Ananthram, Melanie
Subbiah, and Kathleen McKeown. 2023. Check-COVID: Fact-Checking COVID-
19 News Claims with Scientific Evidence. In Findings of the Association for
Computational Linguistics: ACL 2023, Anna Rogers, Jordan Boyd-Graber, and
Naoaki Okazaki (Eds.). Association for Computational Linguistics, Toronto,
Canada, 14114ś14127. doi:10.18653/v1/2023.findings-acl.888

[105] Yuxia Wang, Revanth Gangi Reddy, Zain Muhammad Mujahid, Arnav Arora,
Aleksandr Rubashevskii, Jiahui Geng, Osama Mohammed Afzal, Liangming
Pan, Nadav Borenstein, Aditya Pillai, Isabelle Augenstein, Iryna Gurevych, and
Preslav Nakov. 2024. Factcheck-Bench: Fine-Grained Evaluation Benchmark
for Automatic Fact-checkers. In Findings of the Association for Computational
Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen
(Eds.). Association for Computational Linguistics, Miami, Florida, USA, 14199ś
14230. doi:10.18653/v1/2024.findings-emnlp.830

[106] Zhiruo Wang, Jun Araki, Zhengbao Jiang, Md Rizwan Parvez, and Graham
Neubig. 2023. Learning to filter context for retrieval-augmented generation.
arXiv preprint arXiv:2311.08377 (2023).

[107] Jevin D West and Carl T Bergstrom. 2021. Misinformation in and about science.

Proceedings of the National Academy of Sciences 118, 15 (2021), e1912444117.

[108] Dustin Wright and Isabelle Augenstein. 2021. CiteWorth: Cite-Worthiness
Detection for Improved Scientific Document Understanding. In Findings of the
Association for Computational Linguistics: ACL-IJCNLP 2021, Chengqing Zong,
Fei Xia, Wenjie Li, and Roberto Navigli (Eds.). Association for Computational
Linguistics, Online, 1796ś1807. doi:10.18653/v1/2021.findings-acl.157
[109] Chen Wu, Ruqing Zhang, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi
Cheng. 2021. Learning to truncate ranked lists for information retrieval. In
Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 4453ś4461.
[110] Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kaijing Ma, Cheng-
hao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, Wenhao Huang, Noura
Al Moubayed, Jie Fu, and Chenghua Lin. 2024. SciMMIR: Benchmarking Sci-
entific Multi-modal Information Retrieval. In Findings of the Association for
Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek
Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand,
12560ś12574. doi:10.18653/v1/2024.findings-acl.746

[111] Amelie Wührl and Roman Klinger. 2022. Entity-based Claim Representation
Improves Fact-Checking of Medical Content in Tweets. In Proceedings of the 9th
Workshop on Argument Mining, Gabriella Lapesa, Jodi Schneider, Yohan Jo, and
Sougata Saha (Eds.). International Conference on Computational Linguistics,
Online and in Gyeongju, Republic of Korea, 187ś198. https://aclanthology.org/
2022.argmining-1.18/

[112] Barry Menglong Yao, Aditya Shah, Lichao Sun, Jin-Hee Cho, and Lifu Huang.
2023. End-to-end multimodal fact-checking and explanation generation: A
challenging dataset and models. In Proceedings of the 46th International ACM

448