Evidence Retrieval

From clean Wikipedia labels to agents that decide where to search.

Every paper fixes a different mismatch between benchmark fact-checking and real fact-checking: realism, evidence evaluation, cost, multi-hop decomposition, and source coverage.

2017

LIAR

political truthfulness labels

2018

FEVER

large Wikipedia verification

2020/21

HOVER / PolitiHop

multi-hop reasoning

2023

AVeriTeC

real claims, web evidence

2024/25

Ev2R / FIRE / AFEV

evaluation, cost, decomposition

2026

WKGFC

KG + web agents

The field stops treating fact-checking as a single classification problem and starts treating it as a controlled evidence-gathering process.

1 / 15

Dataset Landscape

Wikipedia-Based Datasets

These datasets typically rely on information modified or extracted from Wikipedia.

Dataset	Size	Labels	Task / Speciality
FEVER	~185,400 claims	Supported, Refuted, Not Enough Info	Standard single-hop fact verification against Wikipedia; claims are human-written.
HOVER	~26,171 claims; 18,171 train, 4,000 dev, 4,000 test	Supported, Not-Supported	Multi-hop reasoning across 2-4 Wikipedia articles.
VitaminC	~400,000 pairs	Supports, Refutes, Neutral	Robust verification with contrastive evidence pairs and subtle factual changes.
FEVEROUS	N/A	N/A	Verification over unstructured text and structured Wikipedia tables.
WikiFactCheck	N/A	N/A	Claims checked specifically against Wikipedia as a source.

Web-Sourced and Real-World Datasets

These use claims from news organizations, social media, or political speech, often with open-web retrieval.

Dataset	Size	Labels	Task / Speciality
AVeriTeC	4,568 claims; 3,068 train, 500 dev, 1,000 test	Supported, Refuted, Conflicting Evidence / Cherry-picking, Not Enough Evidence	Real-world claims verified through QA decomposition and web evidence; avoids temporal leakage.
LIAR	12,836 statements	6 fine-grained truthfulness levels	Short PolitiFact statements for fake-news detection.
LIAR-PLUS	~12,791 total	6 fine-grained labels	LIAR extension with automatically extracted justifications and evidence.
LIAR-New	N/A	Binarized Supported / Refuted	Updated post-2021 PolitiFact claims to reduce LLM pretraining leakage.
RAWFC	2,012 total	3 classification categories	Fake-news detection with weakly structured evidence.
PolitiHop	933 total	3 classification categories	Multi-hop fact-checking focused on political claims.
Factcheck-Bench	631 claims	Supported, Partially Supported, Not Supported, Refuted	Fine-grained fact-checker evaluation at claim, sentence, and document levels.
FacTool-QA	233 instances	True, False	FacTool subset for claims requiring world knowledge.
FELM-WK	184 instances	True, False	LLM factuality benchmark for world knowledge, including some ungrounded claims.
BingCheck	142 sampled instances	Binary, binarized from original 4 labels	Checks factuality of LLM responses using web-search tools.
MultiFC	N/A	N/A	Large multi-domain dataset from 26 fact-checking organizations.
Covid-Fact	N/A	N/A	COVID-19 fact extraction and verification sourced from Reddit.

Specialized and Domain-Specific Datasets

These focus on specific formats, such as charts and graphs, or domains, such as science and health.

Dataset	Size	Labels	Task / Speciality
PubHealth	N/A	Binarized Supported / Refuted	Public-health claims paired with expert-curated evidence paragraphs.
SciFact	N/A	N/A	Scientific claims verified against research abstracts.
ChartCheck	N/A	N/A	Explainable fact-checking over real-world chart images.
FactKG	N/A	N/A	Verification over knowledge graphs, assuming evidence exists inside the KG.
ClaimPKG	N/A	N/A	Graph-based reasoning with pseudo-subgraph generation.
HealthVer	N/A	N/A	Evidence-based fact-checking for health-related claims.
XFact	N/A	N/A	Multilingual fact-checking benchmark.

2 / 15

Results · AVeriTeC

AVeriTeC Results

Retrieval, veracity, and justification scores

Model	Q only	Q + A	Veracity @ (.2/.25/.3)			Justifications @ (.2/.25/.3)
No search	0.19	0.11	0.03	0.02	0.01	0.02	0.01	0.01
Gold evidence	1.00	1.00	0.49	0.49	0.49	0.28	0.28	0.28
AVeriTeC-BLOOM-7b	0.26	0.21	0.23	0.15	0.00	0.11	0.07	0.05
gpt-3.5-turbo	0.29	0.16	0.17	0.10	0.06	0.06	0.04	0.02

Veracity F1 by label

Model	S	R	C	N	Macro
No evidence	.30	.22	.00	.16	.17
Gold evidence	.48	.74	.15	.59	.49
AVeriTeC	.41	.69	.10	.16	.23
gpt-3.5-turbo	.62	.71	.02	.20	.39

S = supported, R = refuted, C = conflicting evidence / cherry-picking, N = not enough evidence.

3 / 15

Results · FIRE

FIRE Results

Performance comparisons across datasets

Framework	LLM	FacTool-QA						FELM-WK						BingCheck
		Label = True			Label = False			Label = True			Label = False			Label = True			Label = False
		Prec	Recall	F1	Prec	Recall	F1	Prec	Recall	F1	Prec	Recall	F1	Prec	Recall	F1	Prec	Recall	F1
Random	-	0.81	0.47	0.59	0.28	0.64	0.39	0.75	0.49	0.59	0.30	0.57	0.39	0.77	0.67	0.72	0.40	0.52	0.45
Always True	-	0.76	1.0	0.86	0	0	0	0.72	1.0	0.84	0	0	0	0.70	1.0	0.83	0	0	0
Always False	-	0	0	0	0.24	1.0	0.39	0	0	0	0.28	1.0	0.44	0	0	0	0.30	1.0	0.46
FacTool	GPT-4o	0.88	0.81	0.84	0.52	0.66	0.58	0.69	0.53	0.60	0.57	0.73	0.64	0.86	0.57	0.68	0.43	0.79	0.56
FacTool	GPT-4o-mini	0.92	0.68	0.78	0.45	0.82	0.58	0.67	0.37	0.48	0.51	0.78	0.62	0.92	0.55	0.69	0.45	0.88	0.60
Factcheck-GPT	GPT-4o	0.90	0.79	0.84	0.52	0.71	0.60	0.67	0.68	0.67	0.61	0.61	0.61	0.85	0.70	0.77	0.50	0.71	0.59
Factcheck-GPT	GPT-4o-mini	0.85	0.80	0.82	0.47	0.56	0.51	0.61	0.50	0.55	0.51	0.62	0.56	0.88	0.78	0.83	0.60	0.76	0.67
Safe	GPT-4o	0.92	0.88	0.90	0.66	0.77	0.71	0.70	0.80	0.75	0.72	0.60	0.65	0.84	0.90	0.87	0.71	0.60	0.65
Safe	GPT-4o-mini	0.92	0.82	0.87	0.58	0.79	0.67	0.61	0.76	0.68	0.61	0.44	0.51	0.86	0.81	0.84	0.60	0.69	0.64
FIRE	GPT-4o	0.92	0.88	0.90	0.65	0.71	0.68	0.70	0.86	0.77	0.77	0.54	0.63	0.86	0.88	0.87	0.70	0.67	0.68
FIRE	GPT-4o-mini	0.87	0.88	0.87	0.60	0.59	0.59	0.63	0.82	0.71	0.67	0.44	0.53	0.87	0.91	0.88	0.74	0.67	0.70
FIRE (No Reason)	GPT-4o	0.88	0.86	0.87	0.60	0.64	0.62	0.70	0.85	0.77	0.77	0.58	0.66	0.85	0.89	0.87	0.70	0.62	0.66
FIRE (No Reason)	GPT-4o-mini	0.87	0.84	0.86	0.55	0.61	0.58	0.65	0.84	0.73	0.71	0.47	0.57	0.84	0.87	0.85	0.66	0.6	0.62
FIRE (No Search)	GPT-4o	0.86	0.87	0.88	0.61	0.54	0.57	0.69	0.86	0.77	0.77	0.55	0.65	0.86	0.91	0.88	0.79	0.64	0.71
FIRE (No Search)	GPT-4o-mini	0.84	0.84	0.84	0.49	0.48	0.49	0.61	0.86	0.72	0.7	0.36	0.48	0.83	0.9	0.87	0.71	0.57	0.63

LLM/search cost and time

Framework	LLM	LLM cost	Search cost	Time
FacTool	GPT-4o	24.76	3.67	2.92
FacTool	GPT-4o-mini	1.49	3.67	2.34
Factcheck-GPT	GPT-4o	21.41	-	4.25
Factcheck-GPT	GPT-4o-mini	1.28	-	4.09
Safe	GPT-4o	6.34	2.93	4.62
Safe	GPT-4o-mini	0.43	2.93	4.25
FIRE	GPT-4o	3.35	0.60	1.31
FIRE	GPT-4o-mini	0.14	0.20	1.25
FIRE (No Reason)	GPT-4o	1.65	0.68	0.57
FIRE (No Reason)	GPT-4o-mini	0.07	0.59	0.54
FIRE (No Search)	GPT-4o	1.70	-	1.03
FIRE (No Search)	GPT-4o-mini	0.11	-	1.34

Cost is in USD and time is in hours for evaluating 559 atomic claims in FacTool-QA, FELM-WK, and BingCheck.

4 / 15

Results · Ev2R

Ev2R Results

Detailed scores for Ev2R components

Scorer	VitaminC		FEVER		AVeriTeC		Avg
Scorer	ρ	r	ρ	r	ρ	r	\|ρ\|	\|r\|
Reference-based (Precision)
GPT4o	.203	.200	.140	.129	.278	.256	.207	.195
Gemini-Pro	.222	.216	.139	.118	.237	.177	.199	.170
Gemini-Flash	.202	.198	.106	.079	−.290	−.205	.199	.161
Llama 3.1	−.037	−.036	.110	.109	.216	.210	.121	.118
Reference-based (Recall)
GPT4o	.229	.228	.146	.114	.314	.284	.230	.209
Gemini-Pro	.253	.246	.176	.143	.285	.278	.238	.222
Gemini-Flash	.232	.227	.032	.000	−.190	−.039	.151	.089
Llama 3.1	.020	.025	−.011	−.012	.128	.121	.053	.053
Reference-based (F1)
GPT4o	.215	.213	.143	.121	.295	.269	.218	.202
Gemini-Pro	.237	.230	.155	.129	.259	.216	.217	.193
Gemini-Flash	.216	.212	.049	.000	−.230	−.066	.172	.115
Llama 3.1	.087	.164	−.024	−.027	.161	.154	.074	.073
Proxy-reference
Proxy-ref	.456	.479	.290	.480	.487	.459	.411	.473

Reference-based scores use different LLM backbones; proxy-reference uses a DeBERTa-based component.

Correlation with AVeriTeC shared-task human ratings

Metric	Ref-less baselines				Proxy-ref	Ref-based baselines				Ev2R
Metric	GPT-4o	Gem-Pro	Gem-Flash	Llama	DeBERTa	RougeL	BLEU	Meteor	H-Met	GPT-4o	Gem-Pro	Gem-Flash	Llama
COV (ρ)	.237	.287	.275	.297	.338	.150	.236	.229	.005	.321	.341	.221	.328
COV (r)	.261	.296	.287	.286	.348	.169	.184	.240	−.024	.323	.326	.299	.336
REL (ρ)	.292	.203	.287	.227	.298	.086	.107	.062	.008	.297	.278	.403	.224
REL (r)	.360	.263	.333	.250	.374	.099	.079	.076	.003	.332	.315	.404	.277

5 / 15

Results · Ev2R

Ev2R Adversarial Tests

Dataset / Test	Reference-less baselines					Ev2R			Proxy-ref
Dataset / Test	METEOR	ROUGE	BLEU	BLEURT	GPT4o	Prec	Rec	Proxy-comp.	DeBERTa
Semantics-altering tests
Completeness	−30.56	−53.19	−77.87	−143.70	−91.76	−21.2	−59.5	−39.35	−82.68
Random shuffle	−32.32	−89.94	−94.20	−1.02	0.55	−6.8	−6.0	−32.22	−92.32
Average	−31.44	−71.57	−86.04	−72.36	−45.61	−14.0	−32.75	−35.79	−87.50
Semantics-preserving tests
Invariance contraction	−0.53	−0.97	−1.35	−1.65	0	−0.8	−0.7	−0.85	0.02
Invariance num2text	−8.34	−10.24	−14.10	−22.13	−3.85	0.1	0	−4.02	−35.64
Invariance text2num	−0.04	−0.90	−1.15	−0.81	−4.40	0	−0.3	−1.59	0.01
Invariance synonyms	−8.22	−36.64	−51.83	−51.00	−2.75	−11.6	−11.3	−15.35	−85.71
Redundancy sent	−26.16	−46.52	−56.15	−80.45	3.30	0	−0.1	−5.04	−67.95
Redundancy words	−5.73	−10.68	−32.69	−31.90	−2.75	−1.0	0.2	−2.89	−52.55
Fluency	−11.52	−23.33	−29.26	−41.69	−2.20	−4.3	−3.9	−2.6	−57.12
Noise	−31.82	−31.81	−31.81	−21.20	−1.65	−28.7	−2.2	2.89	−62.42
Argument structure	−0.01	0	0	0	1.65	0	0	−0.74	−28.93
Average	−10.33	−18.43	−23.94	−27.87	−15.04	−5.16	−2.9	−3.78	−39.79

Values are score differences, in percent, between initial evidence and manipulated evidence. The evaluated Ev2R reference-based scorer uses GPT4o.

6 / 15

Results · Multihop

Multihop Results

First 200 examples of the dev set

System	Supp F1	Ref F1	NEI F1	Conf F1	Acc	AVeriTeC 0.25
AllAtOnce	.591	.813	0	0	.705	.340
LLM+LLM	.644	.821	0	0	.720	.385
Seq+Seq	.638	.816	0	0	.715	.370
4 class	.486	.593	.148	.069	.415	.245
No late verdict	.643	.811	0	0	.705	.450
No long doc	.577	.819	0	0	.705	.465
Multi-doc	.673	.837	0	0	.735	.460
No metadata	.575	.810	0	0	.700	.410
No paraphrase	.701	.839	0	0	.745	.225
Repeat not para	.624	.813	0	0	.710	.340
Algorithm 1	.716	.841	0	0	.750	.495

Final results on full datasets

Data	Submission	Supp F1	Ref F1	NEI F1	Conf F1	Acc	AVeriTeC 0.25
Dev	Algorithm 1	.698	.853	0	0	.754	.486
Dev	Inflated to 10	.698	.853	0	0	.754	.510
Test	Algorithm 1	—	—	—	—	—	.445
Test	Inflated to 10	—	—	—	—	—	.477

7 / 15

Results · Multihop Multi-Evidence

Balanced Accuracy Across Fact-Checking Datasets

Best results are bold; second-best results are underlined. Grey-highlighted columns in the original table denote models that use self-retrieved rather than gold evidence.

Method	Wikipedia sourced		Web sourced		With gold evidence				Overall Avg. (%)
Method	FEVER	HOVER	LIAR-New	AVeriTeC	SummEval	AFact-CNN	PubHealth	Avg. (%)	Overall Avg. (%)
Large-scale LLMs
Gemini-2.5-flash	54.8	62.2	59.1	61.0	69.9	64.6	70.8	68.4	63.2
GPT-4	51.4	62.5	58.6	60.2	69.7	60.7	73.2	67.9	62.3
GPT-4o	55.3	63.8	60.1	61.5	76.3	66.8	67.0	70.0	64.4
Claude 3.5-Sonnet	57.1	64.4	61.3	63.0	77.3	68.8	73.8	73.3	66.5
DeepSeek-V3 67B	53.5	61.7	58.8	60.5	68.3	63.2	72.9	68.1	62.7
Small-scale LLMs
Llama3 8B [7]	48.2	55.1	50.3	54.7	51.7	51.3	70.7	57.9	54.6
Qwen2.5 7B [46]	50.1	57.6	52.9	56.3	58.5	45.3	59.1	54.3	54.3
Llama3.3 70B [7]	57.4	61.2	60.5	62.1	57.6	53.5	76.9	62.7	61.3
Qwen2.5 72B [46]	58.1	62.0	59.7	61.4	53.4	49.9	76.7	60.0	60.2
Specialized methods
HerO [49]	67.5	58.1	70.2	71.4	65.8	62.4	70.3	66.2	66.5
GraphRAG (GPT-4o) [8]	-	-	-	-	68.2	60.4	74.2	67.6	-
GraphCheck [5]	-	-	-	-	66.3	61.9	70.9	66.4	-
FIRE [45]	90.6	67.0	72.8	72.8	58.2	56.9	64.1	59.7	68.9
WKGFC(Ours)	91.9	72.8	81.3	73.2	69.8	58.4	72.8	67.0	74.3

8 / 15

Released: arXiv May 22, 2023 · NeurIPS Datasets & Benchmarks 2023

Paper 1 · AVeriTeC

Problem Solved by Averitec

4,568 claims50 fact-checking orgsEnglish varieties: US, IN, GB, ZAκ = 0.619 verdict agreement

Before

Benchmarks often used artificial claims, closed Wikipedia evidence, weak intermediate reasoning, and sometimes evidence published after the claim.

Problem

Journalistic claims need context, temporal discipline, conflicting evidence handling, and evidence that can explain a verdict.

Method

Normalize the claim, then annotate question-answer evidence with textual justifications. Verification becomes a research path.

Limitation

English-only. The metric can check evidence usefulness, but cannot fully guarantee source-answer consistency.

AVeriTeC label	Meaning
Supported	Evidence backs the normalized claim.
Refuted	Evidence contradicts it.
Conflicting Evidence / Cherry-picking	Both support and refutation appear, or the claim selectively uses evidence.
Not Enough Evidence	Available evidence is insufficient.

claim→normalized claim→Q+A evidence→verdict

Talk track

AVeriTeC changes the unit of work. The system is no longer rewarded for only choosing a label. It has to recover the questions a fact-checker would ask.

9 / 15

Released: arXiv Nov 8, 2024 · TACL Apr 15, 2026

Paper 2 · Ev2R

The evaluation problem: on the open web, there is rarely one gold path.

Before

Systems were judged by final verdict accuracy or evidence overlap with reference text. That works better in closed-source settings than on the web.

Problem

Two systems can retrieve different evidence and both be right. Exact text matching can punish valid evidence.

Method

Ev2R combines reference-based, proxy-reference, and reference-less evidence scorers.

Datasets

Evaluated on AVeriTeC, FEVER, and VitaminC.

The contribution is not a new fact-checker. It is a better ruler for measuring evidence retrieval.

Scorer type	What it checks
Reference-based	Does retrieved evidence match the atomic facts in gold evidence?
Proxy-reference	Does evidence support the predicted verdict label?
Reference-less	Is evidence relevant to the input claim without needing a gold passage?

Reported result

Evaluation on AVeriTeC, FEVER, and VitaminC showed high correlation with human judgments on criteria such as verdict agreement.

Talk track

Ev2R matters because open-web fact-checking creates multiple acceptable trails. The metric moves the field away from brittle text overlap and toward evidence sufficiency.

10 / 15

Published: NAACL Findings, April 2025

Paper 3 · FIRE

The cost problem: fixed retrieval wastes searches and LLM calls.

Before

Atomic claims were checked with a fixed retrieval budget, then verified. The same pipeline ran even when the model already had enough information.

Problem

At scale, fact-checking becomes expensive. Fixed retrieval also does not imitate how humans search, pause, and search again only when needed.

Method

FIRE integrates retrieval and verification in one loop. The agent decides whether to answer or issue another search query based on confidence.

Datasets

FacTool-QA, FELM-WK, Factcheck-Bench, and BingCheck.

7.6×

lower LLM cost, average

16.5×

lower search cost, average

judge confidence→search if needed→verify→stop or loop

Limitation

The confidence decision is compressed into the agent loop for cost reasons. Splitting confidence, retrieval, and verification could make behavior more interpretable, but may cost more.

Talk track

FIRE is the efficiency chapter. It asks a practical question: why retrieve ten things for every claim if the system can decide when retrieval is actually necessary?

11 / 15

Released: arXiv June 9, 2025

Paper 4 · AFEV, Fact in Fragments

The complexity problem: multi-hop claims break when decomposed all at once.

Before

Complex claims were often decomposed statically or checked through coarse evidence retrieval. This can cause noisy evidence and accumulated reasoning errors.

Method

AFEV extracts one atomic fact, retrieves and reranks evidence, verifies it, then uses the result to decide the next atomic fact.

HOVER handling

On HOVER, the model follows the multi-hop chain by verifying fragments sequentially across articles instead of treating the whole claim as one retrieval query.

Settings

Five retrieval candidates, two reranked evidence pieces, and one dynamic demonstration per atomic fact.

fact 1→evidence→rationale→fact 2→aggregate

Dataset	Split / size used	Reported result
LIAR-PLUS	10,240 / 1,284 / 1,267	LA 83.73, Macro-F1 83.12
HOVER	18,171 / 4,000 / 4,000	LA 78.87, Macro-F1 78.76
PolitiHop	592 / 141 / 200	LA 74.14, Macro-F1 57.69
LIAR + RAWFC	LIAR 10,269 train; RAWFC 1,612 train	Reported SOTA over LLM baselines such as RAFTS on both.

Talk track

AFEV is the “do not verify the whole thing at once” paper. It turns a complex claim into a chain of smaller verified commitments.

12 / 15

Released: arXiv Feb 27, 2026 · SIGIR 2026 listed

Paper 5 · WKGFC

The source problem: the web is broad but noisy, knowledge graphs are precise but incomplete.

Before

RAG systems relied heavily on semantic similarity over text. KG systems assumed that the right evidence existed inside the graph.

Problem

Real claims operate under open-world evidence conditions. A fact may be missing from a KG, while web evidence may be relevant but unstructured.

Method

WKGFC frames fact-checking as a partially observable decision process. The agent expands a KG, searches the web when needed, aligns passages into triplets, then decides.

Prompt optimization

It uses self-reflection and TextGrad-style prompt optimization, keeping the base LLM frozen.

KG retrieval→expand + prune→web fill→aligned KG→verdict

Evaluation category	Datasets
Wikipedia-based	FEVER, HOVER
Web-sourced	LIAR-New, AVeriTeC
Provided evidence / summaries	SummEval-series, AggreFact-CNN, PubHealth

5%+

absolute balanced accuracy gain reported over strong baselines

Failure modes

Insufficient KG coverage, maximum-step exhaustion, and over-confident early verdicts.

Talk track

WKGFC brings the story into agentic retrieval. It does not choose between structured and unstructured evidence. It uses a graph as the anchor and the web as the expansion layer.

13 / 15

Evaluation Metrics

Evaluation asks what the claim’s truth status is, why, and how costly the evidence path was.

Veracity

The final truth label for the claim: supported, refuted, conflicting evidence / cherry-picking, or not enough evidence. In AVeriTeC, the veracity score is only awarded when retrieval clears the evidence cutoff.

Justification

The free-text explanation for the verdict. It shows how the evidence leads to the label, and AVeriTeC reports justification quality separately from veracity.

METEOR

A text-similarity score for comparing model output with a reference answer. It gives partial credit for close wording or meaning when valid evidence is phrased differently.

Veracity classification

Label Accuracy and Macro-F1 measure verdict prediction quality.
Balanced Accuracy supports fair comparison across uneven datasets.
AVeriTeC Score awards verdict credit only when sufficient evidence is retrieved.

Evidence retrieval

Hungarian METEOR optimally matches generated QA evidence to reference QA pairs.
Ev2R combines atomic-fact precision/recall with proxy verdict confidence.
Precision, Recall, F1 measure sentence-level evidence retrieval.

Efficiency and agents

LLM token cost and search cost capture deployment expense.
Time measures practical scalability.
Composite reward balances correctness, coverage, citation quality, and efficiency.

Human-centered checks

Coverage, relevance, coherence, repetition, verdict agreement validate metric behavior.
Free-marginal κ and Fleiss’ κ measure annotation reliability.
Error analysis tracks insufficient KG, max-step failures, and over-confidence.

In plain language: veracity answers “what is the truth status?” Justification answers “why is that the verdict?” METEOR gives partial credit when the model’s evidence or explanation is close to the gold reference.

14 / 15

Closing synthesis

The field moved from verdict prediction to evidence management.

Paper	Main question	Technical move	Limit left open
AVeriTeC	Does the benchmark resemble real fact-checking?	Real claims, QA evidence, four labels.	English scope, imperfect evidence-answer consistency checks.
Ev2R	How do we judge evidence on the open web?	Evidence scoring beyond exact gold overlap.	Metric quality still depends on language and scorer reliability.
FIRE	How do we make LLM fact-checking cheaper?	Iterative retrieve-or-answer loop.	Confidence and stopping behavior remain hard to inspect.
AFEV	How do we verify complex claims?	Iterative atomic fact extraction, reranking, dynamic demonstrations.	More steps increase compute and can still inherit evidence quality issues.
WKGFC	How do we combine structured and open evidence?	KG-first agent with web augmentation and prompt-level policy optimization.	KG incompleteness, max-step failures, over-confident stopping.

The conclusion is simple: modern AFC is less about one smarter classifier and more about when to search, where to search, how to decompose the claim, and when the evidence is enough.

Sources used: AVeriTeC, Schlichtkrull et al., 2023, arXiv:2305.13117 / NeurIPS Datasets and Benchmarks 2023.

Ev2R, Akhtar et al., arXiv:2411.05375, 2024; TACL article, 2026.

FIRE, Xie et al., Findings of NAACL 2025, DOI 10.18653/v1/2025.findings-naacl.158.

Fact in Fragments / AFEV, Zheng et al., arXiv:2506.07446, 2025.

WKGFC, Multi-Sourced, Multi-Agent Evidence Retrieval for Fact-Checking, arXiv:2603.00267, 2026.

Dataset details also reflect the provided notes for LIAR, HOVER, PolitiHop, RAWFC, PubHealth, Factcheck-Bench, FacTool-QA, FELM-WK, BingCheck, SummEval, AggreFact-CNN, XFact, ChartCheck, DanFEVER, FakeCovid, and SciFact.

15 / 15