Fact-checking is necessary to address the increasing volume of misinformation. Traditional fact-checking relies on manual analysis to verify claims, but it is slow and resource-intensive. This study establishes baseline comparisons for Automated Fact-Checking (AFC) using Large Language Models (LLMs) across multiple labeling schemes (binary, three-class, five-class) and extends traditional claim verification by incorporating analysis, verdict classification, and explanation in a structured setup to provide comprehensive justifications for real-world claims. We evaluate Llama-3 models of varying sizes (3B, 8B, 70B) on 17,856 claims collected from PolitiFact (2007-2024) using evidence retrieved via restricted web searches. We utilize TIGERScore as a reference-free evaluation metric to score the justifications. Our results show that larger LLMs consistently outperform smaller LLMs in classification accuracy and justification quality without fine-tuning. We find that smaller LLMs in a few-shot inference scenario provide comparable task performance to fine-tuned Small Language Models (SLMs) with large context sizes, while larger LLMs consistently surpass them. Evidence integration improves performance across all models, with larger LLMs benefiting most. Distinguishing between nuanced labels remains challenging, emphasizing the need for further exploration of labeling schemes and alignment with evidences. Our findings demonstrate the potential of retrieval-augmented AFC with LLMs.
Misinformation, whether spread inadvertently or with the intention to deceive, is a global challenge
that can be mitigated efectively through fact-checking eforts [
CEUR
ceur-ws.org
retrieve relevant and credible references, and provide evidence-based verdicts on the aggregated results.
As opposed to style-based detection approaches that learn to distinguish claims based on writing
patterns, AFC systems follow a knowledge-based approach that relies on verification knowledge to
make judgements on claims [
Humans often distrust fact-checking work that challenges their beliefs, perceiving them as biased or
manipulated [
This paper investigates the task formulation and assessment for AFC of real-world claims with LLMs to establish baselines in various settings and to evaluate whether truthfulness ratings can be efectively modeled or if alternative approaches to claim annotations and task formulation are necessary. Based on these findings, future approaches can make more informed design choices and improve the reliability and efectiveness of AFC systems. We propose a framework for AFC with LLMs in a few-shot setup without model fine-tuning for claim analysis, claim veracity prediction, and the generation of justifications as natural language explanations. In the scope of this work, we assess the performance of our framework on 17,856 real-world claims from PolitiFact based on three labeling schemes, with or without web evidence, and across models of diferent sizes (i.e. 3B, 8B, 70B). Using reference-free evaluation metrics and conducting extensive experiments, we provide insight into how evidence integration, model size, and labeling complexity impact system performance. Additionally, we consider fine-tuning small stateof-the-art classification models for estimating the upper bound of predictive performance extractable from diferent components of the data points in the collected dataset and assess the performance of LLMs with a few-shots relative to this limit. Thus, our findings contribute to the development of more robust and transparent AFC systems using LLMs.
Our work builds on the existing body of research in fact-checking and retrieval-augmented generation
while addressing several gaps in the literature. Prior studies have established the value of
transformerbased architectures such as BERT and GPT for tasks ranging from sequence classification to text
generation [
Our approach unifies the components of claim verification into a single structured framework. However,
unlike previous works that often rely on fine-tuned models or separate stages for classification and
explanation [
Additionally, we evaluate our approach on open-source LLMs of diferent scale, in contrast to similar
prior work that utilizes only closed-source models such as ChatGPT [
Fact-checking organizations, that document their eforts and share them publicly, ofer a great
opportunity to analyze relevant misinformation and model the verification process. Moreover, by providing the
initial judgment on what is check-worthy or not, fact-checking experts greatly reduce the complexity of
the task at hand [
Background: The editorial board is made up of 16 journalists ...
Context: ... stated on June 14, 2017 in a New York Times editorial Claim: ”A political map circulated by Sarah Palin’s 2019s PAC incited Rep. Gabby Gifords’s 2019 shooting”
Label: False
PolitiFact’s rating system follows an ordinal six-class labeling scheme. Table 1 provides the oficial descriptions of these six classes.
MOSTLY TRUE ... is accurate but needs clarification or additional information.
HALF TRUE ... is partially accurate but leaves out important details or takes things out of context. MOSTLY FALSE ... contains an element of truth but ignores critical facts [...].
FALSE ... is not accurate.
PANTS ON FIRE ... is not accurate (thus false) and makes a ridiculous claim.
While PolitiFact assigns a separate PANTS ON FIRE label to document the characteristic of ridiculousness in claims, we are only interested in the dimension of truthfulness and therefore treat this special label as a sub-case of False. Thus, we merge the classes, discard the sixth label, and reduce the overall set of labels to five. Table 2 illustrates the resulting distribution of classes.
This Section outlines the methodology used to design and evaluate our framework for automated fact-checking with LLMs. Following the description of the data collection from PolitiFact, we formulate the problem and the experimental setup. Specifically, we discuss model selection, labeling scheme choices, and evidence retrieval.
The approach in this study is motivated by the need to enhance coherence, consistency, and interpretability in automated fact-checking systems. By combining reasoning, classification, and explanation as justification within a single framework, we aim to leverage intermediate analysis to improve performance and ensure consistency between outputs. This study approaches automated fact-checking as a multi-component task with three key objectives: 1. Reasoning: Producing a detailed, step-by-step analysis of the claim using the available information. 2. Verdict: Assigning a veracity label to the claim based on a predefined set of categories. 3. Explanation: Providing a clear and concise explanation in natural language to support the assigned verdict.
The reasoning task follows the idea of chain-of-thought reasoning [23] by constructing a step-by-step analysis of the available information as a natural language explanation [27]. Thus, the verdict classification is integrated with both preceding analysis and subsequent explanation to enhance performance, building on insights from existing research. Zhang et al. [28] demonstrate that jointly generating explanations and predictions outperforms explain-then-predict models. Similarly, Atanasova et al. [29] find that generating fact-checking explanations alongside veracity predictions improves both the performance and the quality of the explanations. These tasks are addressed within a few-shot classification framework, utilizing instruction-based prompts to guide LLMs in generating structured outputs.
We design the prompts based on the previously outlined problem formulation and established principles of prompt engineering [30]. Each prompt is composed of three main components: system, user, and assistant. The system message sets the model’s context and provides the instructions, including the selected labeling scheme. The user message specifies the speaker, context, and claim, with evidence included when available. The assistant message contains the model’s response to the input. To simulate a chat history with desired outputs, few-shot examples, one per label, are included as user and assistant message pairs following the system message and preceding the actual input.
SYSTEM: You are an intelligent decision support system for automated fact-checking. Your tasks are: 1. Analyze the claim step-by-step. 2. Classify the claim’s veracity based on your analysis. [LABELS] 3. Provide a concise natural language explanation for the verdict prediction.
USER: [SPEAKER][CONTEXT] the claim [CLAIM]. Evidence: [EVIDENCE] To ensure consistency and enable automated processing, we enforce a structured output format using the vLLM2 and outlines3 libraries. In this context, structure refers to the property of generated output satisfying a constrained syntax [31]. The output is generated as a parsable JSON object with the following properties: reasoning, verdict, and explanation. Reasoning as free-text, step-by-step analysis of the claim. The verdict as the predicted veracity label, constrained to any option of the predefined set of labels. Lastly, the concise natural language explanation arguing the verdict prediction.
To evaluate performance across diferent model scales, we selected a range of LLMs from the Llama 3 series. We choose Llama architecture models due to their state-of-the-art performance and open-source availability, making them well-suited for evaluating automated fact-checking systems. The models used in this study are Llama-3.2-3B, Llama-3.1-8B, Llama-3.1-70B, Llama-3.3-70B in their instructionifnetuned state. The selection covers varying parameter sizes (3B, 8B, 70B) to investigate the relationship between model scale and task performance. Our strategy is to evaluate the most recent model available at each size. The 3.2 line was the first to introduce the 3B size, while the only 8B version is found in the 3.1 line. For the 70B size, checkpoints are available in both the 3.1 and 3.3 lines. All models have a December 2023 knowledge cutof. During pre-training, the 3.2 models processed 9 trillion tokens, whereas the 3.1 and 3.3 models processed 15 trillion tokens. The 3.3 70B Llama model achieves comparable performance to the 3.1 405B model4, making it one of the most performant open source models at this size. This justifies its inclusion as an additional option in model selection. All models are used in their instruction-tuned state to ensure alignment with the task. Instead of further fine-tuning, we rely on the models’ available capabilities to perform few-shot reasoning, classification and explanation.
Fact-checkers adopt varied approaches to labeling schemes, reflecting diferent priorities and methodologies. Some, such as FullFact5, rely solely on justifications without assigning explicit ratings to claims. Others, like PolitiFact and Snopes6, implement labeling systems grounded in the idea of truthfulness. A further extension of these schemes includes labels for scenarios where evidence is incomplete or unavailable. In the AFC community, truthfulness labels are frequently mapped to a conceptual dimension that evaluates factuality based on available ground-truth evidence. Labels such as supported, refuted, cherry-picked, or not enough information (NEI) are commonly used [32, 33, 26], requiring significant human efort for exploration and annotation. While these approaches provide valuable insights, they also introduce complexities related to interpretation and consistency in annotations. We postpone this perspective to future work. Our focus in this study is to assess whether fact-checking can be efectively modeled across diferent granularities of truthfulness on the collected data. Specifically, we aim to evaluate the trade-ofs between simpler and more nuanced labeling schemes in terms of their impact on classification performance and justification quality. To evaluate the impact of label granularity on fact-checking performance, we merge the five original PolitiFact labels ( True, Mostly True, Half True, Mostly False, False) into coarser schemes, progressively reducing complexity while preserving interpretability. In the three-class scheme, the original labels true and false are grouped into mostly true and mostly false, respectively. In the binary scheme, the label half-true is merged into mostly true. We aim to align our label aggregation with PolitiFact’s definitions, as introduced in Table 1. Table 2 illustrates the resulting distributions. 2https://github.com/vllm-project/vllm 3https://github.com/dottxt-ai/outlines 4https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct#benchmarks 5https://fullfact.org/ 6https://snopes.com/ 14.18% 18.75% 19.79% 17.99% 29.30% The PolitiFact label definitions, as specified in Section 3, are consistent across schemes. By evaluating these schemes, we aim to understand how diferent levels of granularity influence the model’s ability to classify claims and provide useful explanations.
Although PolitiFact’s fact-checking articles provide human-collected evidence that informs the justification and final verdict, extracting and decontextualizing these evidences is not trivial and requires additional specialized modeling and annotation. Consequently, in this study we focus on web-based fact-checking to gather relevant information. We collect the evidence by querying a web search API7 for each claim and retrieve the top 10 search results. We do not apply any query optimization or re-ranking of results. We restrict the search to exclude a list of well-known US fact-checking sites as well as snippets that mention keywords such as ”PolitiFact”, ”fact-check”, or ”debunk” to exclude fact-checking articles or direct references. This way, we aim to reduce information leaking in from pages reporting the actual verification results, rather than evidence. Due to these constrains, we were not able to retrieve evidences for 667 claims. Table 3 lists three search results for the claim presented in Figure 1 where we shortened the snippet text and removed the title and source URLs.
To assess the performance of our automated fact-checking approach, we utilize a combination of classification and generation evaluation metrics. These metrics evaluate both the performance of verdict classification and the quality of generated outputs, ensuring a comprehensive analysis of system performance. We report accuracy and F1-Scores in diferent aggregations strategies to observe diferent aspects of the classification results. To evaluate the quality of generated outputs, we use TIGERScore, a reference-free metric that has been fine-tuned to assess generated text quality based on a set of criteria and assign penalties to mistakes [34]. Specifically, comprehension, accuracy, informativeness, and coherence are evaluated. TIGERScore provides an error evaluation of the generated outputs and assigns penalty scores between [−5, −0.5] for each error without relying on ground truth references. The penalty scores are added up and reported for each case. Thus, a score close to 0 shows higher quality output. In this study, we utilize the 13B TIGERScore model with default hyperparameters to evaluate generated outputs. The evaluation prompt design follows our task prompt as described in Section 4.2.
Due to the stochastic nature of LLMs, evaluation is often not trivial. Thus, we run each fact-checking task three times and report the majority vote for the classification performance evaluation. Additionally, as TIGERScore is a generative evaluation metric, we also run it three times and report the average metric for the justification quality assessment.
The evaluation section presents a detailed analysis of our automated fact-checking approach. We assess the task performance based on model size, labeling scheme, and the impact of evidence retrieval on both classification performance and the quality of generated outputs. This evaluation is structured around our predefined hypotheses and utilizes the previously introduced range of metrics to ensure a robust assessment. Additionally, statistical analyses are conducted to determine the significance of observed performance diferences.
Our evaluation focuses on several fundamental questions regarding the introduced problem setting. We examine whether models can reliably distinguish between the original truthfulness labels, or if alternative approaches to claim annotation and the fact-checking task formulation are required. We also consider potential limitations on the granularity of truthfulness labels that models can efectively handle. Additionally, we assess the role of parametric knowledge in task performance, specifically whether model size yields the expected efect of better performance. Finally, we investigate the impact of evidence integration on task performance. Based on these research questions, our evaluation is structured around the following hypotheses: Hypothesis 1: Classification task performance decreases as label complexity increases. Hypothesis 2: Justification quality decreases as label complexity increases.
Hypothesis 3: Retrieving and incorporating evidence improves both classification accuracy and the quality of generated justifications.
Hypothesis 4: Larger models perform better in the classification task and produce higher quality justifications.
Hypothesis 5: Smaller models benefit more significantly from evidence integration than larger models due to less parametric knowledge being available.
Previously, we introduced a claim involving the New York Times editorial and Sarah Palin in Table 1 and showcased examples of retrieved web evidence in Table 3. In Figure 2, we now present an actual output generated by the Llama3.3-70B model under the evidence-augmented setting with the five-class labeling scheme.
The output in Table 2 demonstrates good justification quality. The verdict is correctly classified as False, aligning with the evidence and reasoning provided. The reasoning section efectively incorporates the retrieved evidence, presenting a detailed analysis of the claim and referencing the correction issued by the New York Times. It also mentions the court’s ruling in favor of the publication, which is not directly relevant to the claim verification. The explanation is concise and supports the verdict, accurately summarizing the key points without introducing ambiguity. This example highlights the potential of retrieval-augmented generation to improve classification accuracy and justification quality. The results presented in Tables 4,5, and 6 illustrate the results in classification performance across diferent labeling schemes and model sizes, with and without evidence retrieval. For the five-class setup (Table 4), evidence retrieval consistently enhances model performance, as seen in higher 1 scores and TIGERScore improvements. However, the 3B model struggles to outperform the baseline significantly, indicating limited capacity in handling a complex task such as automated fact-checking. Verdict
Explanation
Content The New York Times editorial board claimed that a political map circulated by Sarah Palin’s PAC incited Rep. Gabby Gifords’ shooting. However, there is no established link between the two events, and the Times subsequently issued a correction. The case was later brought to court, and the New York Times was sued for defamation. The court ultimately ruled in favor of the New York Times, stating that the editorial did not defame Palin.
False The statement is not accurate. There is no established link between the political map circulated by Sarah Palin’s PAC and the shooting of Rep. Gabby Gifords, and the New York Times issued a correction after publishing the claim.
In the three-class classification scheme (Table 5), evidence retrieval again provides a notable performance boost across all models, with improvements becoming more pronounced in larger models. This indicates that as label complexity decreases, models are better able to leverage evidence to enhance classification accuracy and justifications. The 3.3-70B-Instruct model achieves the highest scores, emphasizing the advantage of size when combined with external knowledge. Table 5 presents the classification metrics for the three-class scheme. Similar to the five-class results, evidence retrieval enhances performance across all models.
For binary classification results in Table 6, the reduced complexity of the task yields the highest overall performance across all models. Evidence retrieval continues to provide a measurable benefit, particularly in the largest models, where the highest 1 scores and TIGERScore improvements are observed. In the following, we investigate the indications we have described earlier with statistical analyses to draw conclusions about the hypotheses we specified in Section 5.1.
We conducted a Friedman test on 1 across the three classification schemes, with and without evidence. The result indicates that at least one of the schemes difers significantly ( < 0.05 ) in terms of classification performance. These findings support hypothesis 1 that as labeling becomes more complex, classification performance tends to decrease, potentially due to more nuanced distinctions between labels that increase the quantity of prediction errors. For TIGERScore evaluation, the Friedman test was significant ( < 0.05 ) for the setting with evidences, but a subsequent Conover’s test revealed no significant pairwise diferences. Additionally, the Friedman test reveals no significance for the setting without evidence. This result suggests that there is no measurable diference in justification quality across the three schemes, with or without evidence. These ifndings reject hypothesis 2 that more complex label sets negatively afect overall justification quality. This may be in part due to claim analysis and explanation being dificult enough, regardless of whether the label schemes are more or less complex.
To determine the statistical significance of performance when including evidence, we conducted paired t-tests comparing models with and without evidence across all classification schemes for the 1 and TIGERScore. The results indicate a statically significant diference (p < 0.01) for both metrics when evidence retrieval is included during the fact-checking task. This supports hypothesis 3. Thus, external evidence helps the model to disambiguate classes and produce more useful justifications for the fact-checking task.
To evaluate whether larger models outperform their smaller counterparts, we performed a Friedman test on both 1 and TIGERScore with and without evidence across four diferent model sizes. The results indicate a significant diference ( < 0.05 ), confirming that model size has a measurable impact on performance. These findings support hypothesis 4.
Finally, to investigate whether smaller models benefit more from evidence integration than larger models, we examined the performance gains by subtracting the no evidence scores from the with evidence scores for both 1 and TIGERScore across all model sizes. For 1 gains, the Friedman test showed no significant diference ( = 0.167 ), whereas the TIGERScore gains were statistically significant ( < 0.05 ). Thus, we partially reject hypothesis 5. This implies that larger models benefit even more from external evidence, presumably due to their ability to reason efectively across long context sizes, whereas smaller models exhibit relatively limited improvements. We expect that integrating more credible and complete information sources, could enhance overall performance for both smaller and larger models even further.
Since we consider fine-tuning approaches impractical for real-world automated fact-checking, due to the dynamic and fast-changing nature of misinformation, which limits the usefulness of models trained on static datasets, our primary focus in this study has been on few-shot inference using large language models. Earlier encoder-based architectures, such as BERT, were constrained by a maximum sequence length of 512 tokens, which restricted their ability to incorporate additional context. Recent advancements, such as ModernBERT [35], adjust the original BERT architecture and are able to support sequence lengths of up to 8192 tokens. This allows the integration of more contextual information and retrieved evidence directly into the classification process, enabling the evaluation of their utility for veracity prediction.
To complement our few-shot evaluation and to better understand how diferent input signals contribute to classification outcomes, we conduct an ablation study using the ModernBERT-large architecture. Specifically, we fine-tune the model across a series of input configurations to assess how predictive performance changes when incrementally adding additional contextual information. We begin with the claim alone as input. We then add information about the surrounding context in which the claim appeared, such as a speech, interview, or social media post. Next, we incorporate the speaker who issued the claim. Finally, we include retrieved web evidence that provides external factual grounding. This study helps quantify the individual impact of each component and provides an empirical upper predict performance bound for fine-tuning on the dataset, enabling a more informed comparison with few-shot LLM performance. The results in Table 7 show that incorporating evidence consistently produces the most significant gains across all label granularities. In the five-class setting, starting with only the claim results in the lowest performance. Adding context leads to modest improvements, suggesting that surrounding details help disambiguate some claims. For example, knowing whether a statement was made during a campaign rally or in an oficial policy document can influence its interpretation. Speaker information further improves performance, which may be attributed to prior knowledge about the speaker’s reliability, role, or political alignment that implicitly guides veracity estimation. In the binary setting, adding context does not improve performance and even reduces it slightly. This outcome likely stems from the way binary labels are constructed by merging more nuanced classes. As a result, diferent claims with dissimilar contexts may be grouped under the same binary label, making context a noisy feature. In contrast, speaker information helps more consistently. This may reflect the fact that in coarse-grained tasks, speaker identity acts as a high-level signal about the probable factuality of a claim. The classification results align closely with the previously presented LLM few-shot inference results, showing that evidence consistently provides significant performance improvements across all label schemes. For both classifiers and LLMs, the inclusion of evidence enables better disambiguation and enhances predictive performance, particularly in more complex multi-class tasks. While smaller LLMs provide comparable performance across tasks in few-shot inference, larger LLMs consistently surpass the fine-tuned SLMs without requiring fine-tuning and provide the additional advantage of generating reasoning and detailed justifications across more extensive contexts. This highlights the general utility of LLMs for AFC.
This study investigated AFC of real-world claims using LLMs in a few-shot inference scenario. By evaluating task performance across three labeling schemes and multiple LLM sizes of the same architecture, we demonstrated the importance of evidence integration, model scale, and labeling complexity in determining system efectiveness. Evidence retrieval consistently improved classification accuracy and justification quality, with larger models showing the most significant gains. In contrast, smaller models struggled to perform or benefit as much from evidence integration, highlighting the need for further optimization in computationally constrained environments. While more coarse-grained labels naturally yield higher performance, future work should explore how to integrate alternative labeling strategies and a more nuanced assessment of claims across diferent perspectives to develop a robust AFC approach. Our experiments show that LLMs can efectively perform multi-component tasks by reasoning over presented data and generating detailed justifications. However, our results also indicate that alternative approaches leveraging fine-tuning can be advantageous for specific subtasks or in resource-constrained settings. For instance, while LLMs excel in knowledge-based reasoning and explanation generation in few-shot scenarios, models like ModernBERT can be suficiently efective for classification tasks when supervised training data is available. This suggests the potential for hybrid frameworks where supervised fine-tuning is employed for tasks that rarely change, such as document type or natural language inference classification, while LLMs are reserved for dynamic scenarios that require integration of dynamic facts to produce grounded reports. Moreover, integrating more credible evidence, including human-aggregated sources, could further enhance AFC performance by providing more reliable context for claim evaluation. Furthermore, although our study focused on LLMs from the Llama family, future work could benefit from expanding the comparison to include diferent model families. A broader analysis comparing diverse model families would be valuable, especially for applications where training and inference costs, use cases, and interpretability requirements difer substantially. To extend AFC toward intelligent decision assistance for expert fact-checkers, future research should focus on structuring justifications to align more closely with human verification strategies. This includes presenting concise, faithful explanations that detail key reasoning steps and clearly highlight the integrated evidence. Preliminary observations indicate that LLM-based systems can sufer from hallucinations, underscoring the need for extensive evaluation and user studies to understand how experts interpret and trust the generated explanations. Such studies would not only help refine the presentation of justifications but also identify gaps in current AFC systems and better define their role in supporting, rather than replacing, human fact-checking eforts.
This study is partially funded by the German Federal Ministry of Education and Research (BMBF, reference: 03RU2U151C) in the scope of the research project news-polygraph and by JST AIP Acceleration Research (JPMJCR24U3) and JST CREST Grants (JPMJCR20D3). Linguistics, Online, 2020, pp. 1906–1919. doi:10.18653/v1/2020.acl-main.173. [22] C. Si, N. Goyal, S. T. Wu, C. Zhao, S. Feng, H. Daumé III, J. Boyd-Graber, Large Language Models
Help Humans Verify Truthfulness – Except When They Are Convincingly Wrong, 2024. [23] J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, D. Zhou, Chain-of
Thought Prompting Elicits Reasoning in Large Language Models, 2023. [24] I. Augenstein, C. Lioma, D. Wang, L. Chaves Lima, C. Hansen, C. Hansen, J. G. Simonsen, MultiFC: A Real-World Multi-Domain Dataset for Evidence-Based Fact Checking of Claims, in: Proceedings of 2019 EMNLP-IJCNLP, Association for Computational Linguistics, 2019, pp. 4685–4697. [25] W. Y. Wang, “Liar, Liar Pants on Fire”: A New Benchmark Dataset for Fake News Detection, in: R. Barzilay, M.-Y. Kan (Eds.), Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), Association for Computational Linguistics, Vancouver, Canada, 2017, pp. 422–426. doi:10.18653/v1/P17-2067. [26] B. M. Yao, A. Shah, L. Sun, J.-H. Cho, L. Huang, End-to-End Multimodal Fact-Checking and Explanation Generation: A Challenging Dataset and Models, in: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2023, pp. 2733–2743. doi:10.1145/3539618.3591879. arXiv:2205.12487. [27] N. Kotonya, F. Toni, Towards a Framework for Evaluating Explanations in Automated Fact Verification, 2024. arXiv:2403.20322. [28] Z. Zhang, K. Rudra, A. Anand, Explain and Predict, and then Predict Again, in: Proceedings of the 14th ACM International Conference on Web Search and Data Mining, WSDM ’21, Association for Computing Machinery, New York, NY, USA, 2021, pp. 418–426. doi:10.1145/3437963.3441758. [29] P. Atanasova, J. G. Simonsen, C. Lioma, I. Augenstein, Generating Fact Checking Explanations, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, 2020, pp. 7352–7364. [30] S. Schulhof, M. Ilie, N. Balepur, K. Kahadze, A. Liu, C. Si, Y. Li, A. Gupta, H. Han, S. Schulhof, P. S. Dulepet, S. Vidyadhara, D. Ki, S. Agrawal, C. Pham, G. Kroiz, F. Li, H. Tao, A. Srivastava, H. Da Costa, S. Gupta, M. L. Rogers, I. Goncearenco, G. Sarli, I. Galynker, D. Peskof, M. Carpuat, J. White, S. Anadkat, A. Hoyle, P. Resnik, The Prompt Report: A Systematic Survey of Prompting Techniques, 2024. arXiv:2406.06608. [31] B. T. Willard, R. Louf, Eficient Guided Generation for Large Language Models, 2023. [32] A. Hanselowski, C. Stab, C. Schulz, Z. Li, I. Gurevych, A Richly Annotated Corpus for Diferent
Tasks in Automated Fact-Checking, 2019. doi:10.48550/arXiv.1911.01214. arXiv:1911.01214. [33] M. Schlichtkrull, Z. Guo, A. Vlachos, AVeriTeC: A Dataset for Real-world Claim Verification with
Evidence from the Web, 2023. doi:10.48550/arXiv.2305.13117. arXiv:2305.13117. [34] D. Jiang, Y. Li, G. Zhang, W. Huang, B. Y. Lin, W. Chen, TIGERScore: Towards Building Explainable
Metric for All Text Generation Tasks, 2024. [35] B. Warner, A. Chafin, B. Clavié, O. Weller, O. Hallström, S. Taghadouini, A. Gallagher, R. Biswas, F. Ladhak, T. Aarsen, N. Cooper, G. Adams, J. Howard, I. Poli, Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Eficient, and Long Context Finetuning and Inference, 2024. doi:10.48550/arXiv.2412.13663. arXiv:2412.13663.