text
string
source
string
. . . . . . . . . . . . . . . . . . Qwen License • Gemini-2.5-Flash [49]7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Closed-Source • InternVL-3.0 [3]8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MIT License 4htt...
https://arxiv.org/abs/2505.18675v1
arXiv:2505.18677v1 [cs.CL] 24 May 2025Social Good or Scientific Curiosity? Uncovering the Research Framing Behind NLP Artefacts Eric Chamoun1, Nedjma Ousidhoum2*, Michael Schlichtkrull3,*, Andreas Vlachos1 1Department of Computer Science and Technology, University of Cambridge 2Cardiff University 3Queen Mary University...
https://arxiv.org/abs/2505.18677v1
further offer guidance by encouraging alignment between research goals and stakeholder needs, prompting reflection on issues like dual use (Leins et al., 2020) and overclaiming (Grodzinsky et al., 2012). However, their reliance on manual annotation limits their scalability and timeliness. In this work, we first propose...
https://arxiv.org/abs/2505.18677v1
under Scientific Curiosity , largely driven by benchmark studies examining LLM lim- itations. Lastly, we find a growing emphasis on generating fact-checking justifications to assist hu- man fact-checkers, while fewer studies propose fully automating the process. Research framings for this paper We release two NLP artef...
https://arxiv.org/abs/2505.18677v1
users, and goals; vague ones omit key components or introduce inconsis- tencies that obscure the system’s intended use. We note that theoretical research without an immediate ends or intended user is not the same as vagueness. If a paper clearly states that the goal is theoretical exploration or to answer some research...
https://arxiv.org/abs/2505.18677v1
framing, we evaluate each logical rule Rjover the predicted probabilities of its constituent epistemic elements. Conjunctions and disjunctions are computed as Pr(e1∧e2) = min {Pr(e1),Pr(e2)}and Pr(e1∨e2) = max {Pr(e1),Pr(e2)}, respectively. This soft logic enables smooth aggregation of par- tial evidence across element...
https://arxiv.org/abs/2505.18677v1
holders. We annotate the presence of epistemic element types defined in §2. Each paper is then as- signed one or more discourse-level research fram- ings. Appendix E.3 defines the framing types. All examples were annotated based on explicitly stated text only, avoiding inferred intent. Detailed guide- lines are provide...
https://arxiv.org/abs/2505.18677v1
specified labels—low enough to avoid inflating irrelevant elements (as model-assigned likelihoods exceed -3) but high enough to retain research fram- ings dependent on their absence. Framing Classification We define confidence tiers for both epistemic element and framing scores using the thresholds τh=−1,τm=−2, and τl=...
https://arxiv.org/abs/2505.18677v1
lent outputs—thus supporting effective uncertainty quantification across multiple generations. As shown in Appendix B.1, performance improves with more samples, highlighting the benefits of our approach over a single-prompt baseline. Table 2 shows that using all paragraphs slightly outperforms human-filtered ones, sugg...
https://arxiv.org/abs/2505.18677v1
(HS). A closer examination re- veals that deterministic rules struggle with context- dependent research framings like Vague Data Anal- ysis, which require reasoning about missing con- nections between means and ends. The LLM mitigates this by integrating context and reason- ing to differentiate overlapping research fra...
https://arxiv.org/abs/2505.18677v1
incorporating rule-based rankings. 95% confidence intervals are from 15 runs. by the growth of papers analyzing LLM capabili- ties in fact-checking. The model classifies works like Hu et al. (2024)—focused on evaluating LLM factual reasoning—as scientific curiosity, noting: “The primary goal is to understand LLM limita...
https://arxiv.org/abs/2505.18677v1
mistakes. Individ- ual misclassifications can occur, though large- scale trends remain informative. •Appropriate use cases – The tool is designed primarily for scalable analyses of new sub- fields (where small errors cancel out in aggre- gate analyses) and writing assistance (offer- ing feedback, research framing ranki...
https://arxiv.org/abs/2505.18677v1
Lillicrap, Ange- liki Lazaridou, Orhan Firat, James Molloy, Michael Isard, Paul R. Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, Jack Krawczyk, Cosmo Du, Ed Chi, Heng-Tze Cheng, Eric N...
https://arxiv.org/abs/2505.18677v1
2024 , 2024. Paul Jaccard. 1901. Étude comparative de la distribu- tion florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat , 37:547–579. Zhijing Jin, Geeticka Chauhan, Brian Tse, Mrinmaya Sachan, and Rada Mihalcea. 2021. How good is NLP? a sober look at NLP tasks through the lens of social impact...
https://arxiv.org/abs/2505.18677v1
hu- man fact-checkers ,gather and present evi- dence ,identify multimodal inconsistencies , automated removal ,provide labels/veracity scores ,provide aggregates of social media comments ,filter system outputs ,maintain con- sistency with knowledge base ,analyse data , produce misinformation ,vague persuasion •Epistemi...
https://arxiv.org/abs/2505.18677v1
,evidence retrieval ,human in the loop ∧Ends: increase veracity of published content Assisted Media Consumption: _   Data subjects: social media users ∧Model owners: social media companies ∧_ Application means: vague persuasion ,Ends: limit misinformation  , Data actors: media consumers Automated Content Mode...
https://arxiv.org/abs/2505.18677v1
(e.g., bias in hate speech). Annotations are extracted from intro- ductions and abstracts. Initially, we annotated a small subset of papers based on predefined criteria. We refined category definitions through discussion, resolved label in- consistencies, and conducted a final annotation round to ensure consistency. E....
https://arxiv.org/abs/2505.18677v1
posts or text • Provide justifications • Corpora analysis • Data collection • Human-in-the-loop Classify/Score Posts or Text When the paper proposes classifying posts (e.g., hate vs. non-hate, toxic vs. non-toxic) or using classifier-generated scores. Provide Justifications When the paper proposes generating explanatio...
https://arxiv.org/abs/2505.18677v1
-3 Vague Identification -2 Vague Moderation -1 Law Enforcement 0 Scientific Curiosity 1 Automatic Content Moderation 2 Assisted Content Moderation 3 Assisted Knowledge Curation 4 Table 7: Ordinal classes for research framings: vague research framings have negative values. E.3 Research Framings At the discourse level, w...
https://arxiv.org/abs/2505.18677v1
who deploy and make decisions about such systems. Assisted Knowledge Curation When the anal- ysed paper proposes toxic content detection pri- marily as a component filtering the information kept in some curated knowledge vault, including graph-based knowledge bases as well as text-based collections such as Wikipedia. H...
https://arxiv.org/abs/2505.18677v1
be applied on statements from con- tributors to social media platforms, excluding public figures. This in- cludes commenters on forums as well as, e.g., Twitter users. Simi- larly, editors of Wikipedia and other collaborative writing projects fall under this category as well. •Public figures/politicians : If the excerp...
https://arxiv.org/abs/2505.18677v1
be utilized by maintainers of knowl- edge bases integrating model out- puts. •Law enforcement : If the excerpt explicitly mentions that the re- leased artefacts are expected to be utilized by agents using model out- puts in investigations or legal pro- cesses. •Algorithm : If the excerpt explicitly mentions that the re...
https://arxiv.org/abs/2505.18677v1
in the loop : If the excerpt explicitly mentions that released artefacts involve including human oversight in the solution or main process described in the paper. •Corpora analysis : If the excerpt explicitly mentions that released artefacts involve using data analyt- ics or analyzing datasets for pat- terns or insight...
https://arxiv.org/abs/2505.18677v1
statements aimed at refuting, mitigating, or op- posing harmful or hateful content (e.g., generating automated replies to hate speech). If none of these labels apply, respond with "not specified in this paragraph." F.1.6 Ends Task : Identify the ends explicitly mentioned by the authors in the provided excerpt. The ends...
https://arxiv.org/abs/2505.18677v1
unless they are directly identified as targeted by the released artefacts. 2.Focus on the released artefacts : Differentiate between general back- ground information and what the authors specifically propose. If the authors discuss fact-checking broadly or reference prior work, do not assume relevance; focus solely on ...
https://arxiv.org/abs/2505.18677v1
only data actors explicitly named or described in the excerpt. Do not infer their presence unless they are directly identified as acting on the released artefacts. 2.Focus on the released artefacts : Differen- tiate between general background informa- tion and what the authors specifically propose. If the authors discu...
https://arxiv.org/abs/2505.18677v1
intervention. If none of these labels apply, respond with "not specified in this paragraph." F.2.3 Model Owners Task: Identify the model owners explicitly men- tioned by the authors in the provided excerpt. Model owners are entities or organisations that the authors explicitly target as being responsible for controllin...
https://arxiv.org/abs/2505.18677v1
with "not spec- ified in this paragraph" . 4.Consider the big picture : Identify the strate- gies used by the authors in their study, if men- tioned at all. Question: Based on the provided excerpt and using the provided labels, who are the explicitly mentioned model means? You will select the single most appropriate la...
https://arxiv.org/abs/2505.18677v1
nize evidence supporting or refuting specific claims. •Identify multimodal inconsistencies : If the excerpt explicitly mentions that released arte- facts would be leveraged to detect contradic- tions between different content types. •Automated removal : If the excerpt explic- itly mentions that released artefacts would...
https://arxiv.org/abs/2505.18677v1
to ad- vance understanding of language and NLP methods. •Avoid biases of human fact-checkers : If the excerpt explicitly mentions that the ul- timate goal of the released artefacts would be to achieve unbiased and consistent fact- checking. •Detect falsehood for law enforcement : If the excerpt explicitly mentions that...
https://arxiv.org/abs/2505.18677v1
represents one solution to this critical prob- lem. Mapping: This narrative is identified when the application means are not specified, but the ends include fighting hate. F.3.5 Assisted Content Moderation Definition: The paper proposes the deploy- ment of automated moderation as a tool to assist content moderators on ...
https://arxiv.org/abs/2505.18677v1
We create and analyze a dataset to this end. Mapping: This narrative typically includes application means focused on analyzing data or corpora analysis but lacks a clear connec- tion between this analysis and real-world in- terventions. F.3.10 Truth-Telling for Law Enforcement Definition: The paper proposes hate speech...
https://arxiv.org/abs/2505.18677v1
trust and costing lives by e.g., inducing hesitance to adopt life-saving vaccines. An important way to fight misinfor- mation is the production of relevant counter- messaging, i.e., the work done by organiza- tions such as Full Fact or PolitiFact. With the number of false claims published on social media every hour, it...
https://arxiv.org/abs/2505.18677v1
is clear that the mechanism is supposed to fol- low what fact-checkers are currently doing, but not where or how the ML model will be used in this process. Furthermore, it is un- clear whether the entire process will be auto- mated. For example, the paper may suggest that an automated fact-checking model should be used...
https://arxiv.org/abs/2505.18677v1
collections such as Wikipedia. Example: Knowledge bases fuel many real-world NLP applications, e.g., question answering. The maintenance of knowledge bases is an expen- sive process, yet as new facts appear in the world knowledge bases must be kept up-to- date. Automated triple extraction from e.g., news data has been ...
https://arxiv.org/abs/2505.18677v1
One solution is for moderators to re- move information deemed false. However, with the number of posts made every day on social networks, this strategy is too costly. In this paper, we develop an automated system for detecting false claims, which can serve as a first line of defense against misinformation.Mapping: This...
https://arxiv.org/abs/2505.18677v1
PD3F: A Pluggable and Dynamic DoS-Defense Framework Against Resource Consumption Attacks Targeting Large Language Models Yuanhe Zhang1,⋆, Xinyue Wang1,⋆, Haoran Gao2, Zhenhong Zhou1, Fanyu Meng2, Yuyao Zhang2, Sen Su1,† 1Beijing University of Posts and Telecommunications,2China Mobile Research Institute {charmes-zhang,...
https://arxiv.org/abs/2505.18680v1
reliability of LLMs deployment. Despite its severity, resource consumption at- tacks remain largely unaddressed, making it diffi- cult to mitigate. Prior defense techniques, includ- ing model checking and input disturbance (Jain et al., 2023; Liu et al., 2024), are bypassed by emerging attack strategies, leading to sev...
https://arxiv.org/abs/2505.18680v1
and eight defense base- lines, demonstrating its effectiveness. PD3F offers a novel perspective on LLM security defenses and improves the deployment robustness. 2 Related work Jailbreak attacks. Jailbreak attacks aim to by- pass LLMs’ alignment safeguards to induce harm- ful outputs (Wei et al., 2023). Existing studies...
https://arxiv.org/abs/2505.18680v1
per- formance degradation caused by DoS attacks. Figure 3: Difference between benign and attack requests under the Resource Index on the Llama70B model. 3.1 Resource Index Recent studies have shown that resource consump- tion attacks can lead to significant consumption of GPU resources in LLMs (Shumailov et al., 2021)....
https://arxiv.org/abs/2505.18680v1
applied clustering accordingly. Finally, the Resource Index, composed of Itand Ic, jointly characterizes the potential aggressive- ness of a request from two orthogonal perspectives: behavioral similarity and resource intensity.We apply the Interquartile Range (IQR) method (Tukey et al., 1977) to each index. For any in...
https://arxiv.org/abs/2505.18680v1
at the end of each scheduling round to determine sub- sequent scheduling priorities. 3.3 Adaptive End-Based Suppression The length of responses generated by LLMs is di- rectly determined by the occurrence of the <EOS> token (Vaswani et al., 2017; Ansari et al., 2024). To mitigate resource consumption attacks, we modu- ...
https://arxiv.org/abs/2505.18680v1
used. Figure 4: The improvement of PD3F in benign user throughput (BUT) indicates stronger resistance to attacks, while the reduction in total tokens (TT) reflects decreased overall resource consumption. Qwen7B (Yang et al., 2024), Qwen32B (Hui et al., 2024), Qwen72B (Yang et al., 2024), Mistral7B (Jiang et al., 2023)....
https://arxiv.org/abs/2505.18680v1
detection performance of \ours, achieving an F1 score exceeding \textbf{0.97} against existing attack methods, demonstrating both high recognition accuracy and strong generalization. an average Attack Determination Accuracy of over 99% across the three attack types, and nearly 100% accuracy on Llama and Qwen models. Mo...
https://arxiv.org/abs/2505.18680v1
indicates that with the integration of our Adaptive End-Based Suppression mechanism, the system’s total time was reduced by approximately 50% on average, and up to 60% for LLaMA70B. Additionally, the BUT improved by nearly 100% /uni00000015 /uni00000017 /uni00000019 /uni0000001b/uni00000013/uni00000011/uni00000013/uni0...
https://arxiv.org/abs/2505.18680v1
This paper focuses on the field of model security, specifically addressing the degradation of LLM ap- plication service capabilities caused by resource consumption attacks. We propose effective de- fense mechanisms tailored to different categories of such attacks. Although the study targets server-side LLM deployments,...
https://arxiv.org/abs/2505.18680v1
Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 . Robert J. Creasy. 1981. The origin of the vm/370 time- sharing system. IBM Journal of Research and Devel- opment , 25(5):483–490. Tianyu Cui, Yanling Wang, Chuanpu Fu, Yong Xiao, Sijia L...
https://arxiv.org/abs/2505.18680v1
text-to-image generation. arXiv preprint arXiv:2403.19103 , 2(5). Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2020. Aligning ai with shared human values. arXiv preprint arXiv:2008.02275 .Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn ...
https://arxiv.org/abs/2505.18680v1
Jacob Gard- ner. 2023. Black box adversarial prompting for foun- dation models. arXiv preprint arXiv:2302.04237 . Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, and Amin Karbasi. 2024. Tree of attacks: Jailbreaking black-box llms automatically. Advances in Neural Informa...
https://arxiv.org/abs/2505.18680v1
on recursively gener- ated data. Nature , 631(8022):755–759. Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Pa- pernot, Robert Mullins, and Ross Anderson. 2021. Sponge examples: Energy-latency attacks on neu- ral networks. In 2021 IEEE European symposium on security and privacy (EuroS&P) , pages 212–231. IEEE. Willi...
https://arxiv.org/abs/2505.18680v1
Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019b. Defending against neural fake news. Advances in neural information processing systems , 32. Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, Lei Bai, and Xiang Wang. 2025a. Multi-agent archi- tecture search via agentic super...
https://arxiv.org/abs/2505.18680v1
security alignment safeguards, thereby outputting content that should be rejected, such as violence, discrimination, illegal activities, or information that violates platform policies (Xu et al., 2024b; Yi et al., 2024; Xu et al., 2024a; Cui et al., 2024; Deng et al., 2025; Wang et al., 2025). Attacks usually use the f...
https://arxiv.org/abs/2505.18680v1
al., 2024)), Qwen32B (Qwen2.5-32B-Instruct (Hui et al., 2024)), Qwen72B (Qwen2.5-72B-Instruct (Yang et al., 2024)), Mistral7B (Mistral-7B-Instruct-v0.2 (Jiang et al., 2023)). In our experiments, the maximum output length was set to 4096 tokens for all models. To ensure strict control over evaluation conditions and syst...
https://arxiv.org/abs/2505.18680v1
settings. The HumanEval dataset (Chen et al., 2021) contains 164 programming questions, which are used to evaluate the functional correctness of the code generated by the language model, with a special focus on the model’s ability to generate correct code based on natural language descriptions. Lastly, GPQA (Graduate-L...
https://arxiv.org/abs/2505.18680v1
low detection accuracy in Auto- DoS scenarios. For PPL and KSD in Tab. 7, GCG and P-DoS attacks showed extremely high PPL values, indicating a severe deviation from the normal output distribution and rendering PPL ineffective, while KSD scores were notably low under P-DoS, also failing to provide reliable detection. Fu...
https://arxiv.org/abs/2505.18680v1
requests per user) to verify the Dynamic Request Polling mechanism and the Adaptive End-Based Suppression mechanism’s contributions to the system performance. Ablation Dynamic Request Polling Scheduling. Tab. 11 presents the results when Request Polling is replaced with RR. On Llama70B, the BUT drops from 0.26 to 0.09 ...
https://arxiv.org/abs/2505.18680v1
of 5 users are malicious), the normal-user throughput improves most significantly, effectively mitigating the resource exhaustion caused by attackers. Overall, PD3F can not only cope with different request loads, but also has strong adaptive ability to changes in the proportion of malicious users. J Examples of each In...
https://arxiv.org/abs/2505.18680v1
He was trying to collect a $400 bill for medical services rendered to Pratt by Doctor. Dayton went to Pratts house and when Martina, Pratts mother, answered the door, Dayton told Martina he was there to collect a bill owed by Pratt . Martina told Dayton that because of her illness, Pratt had been unemployed for six mon...
https://arxiv.org/abs/2505.18680v1
= 40. Step 3: Calculate the percentage of flyers passed out by Belinda. To find the percentage of flyers passed out by Belinda, we need to divide the number of flyers passed out by Belinda (40) by the total number of flyers (200) and multiply by 100. So, the percentage is(40 / 200) * 100 = 20. The final answer is: \box...
https://arxiv.org/abs/2505.18680v1
x x x x x x x x x x x x x x x x x x x x x": Same as the previouspoints. 9. "x x x x x x x x x x x x x x x x x x x x x x": Placeholdertext or a series of placeholders. 10. "x x x x x x x x x x x x x x x x x x x x x x": Same as the previouspoints. 11. "x x x x x x x x x x x x x x x x x x x x x x": Placeholdertext or a se...
https://arxiv.org/abs/2505.18680v1
we set up an experiment and illustrate the Top logit and EOS logit values over the generation process based on experimental data. As shown in Fig. 7, Fig. 16and Fig. 25, we illustrate the EOS indicator during three types of attacks on the Llama8B model, where the output length is unrestricted and allowed to reach the d...
https://arxiv.org/abs/2505.18680v1
Model. /uni00000013 /uni00000015/uni00000013/uni00000013 /uni00000017/uni00000013/uni00000013 /uni00000019/uni00000013/uni00000013 /uni0000001b/uni00000013/uni00000013 /uni00000014/uni00000013/uni00000013/uni00000013/uni00000010/uni00000015/uni00000013/uni00000013/uni00000015/uni00000013/uni00000017/uni00000013/uni0000...
https://arxiv.org/abs/2505.18680v1
with the Llama8B Model. /uni00000013 /uni00000015/uni00000013/uni00000013 /uni00000017/uni00000013/uni00000013 /uni00000019/uni00000013/uni00000013 /uni0000001b/uni00000013/uni00000013 /uni00000014/uni00000013/uni00000013/uni00000013/uni00000013/uni00000014/uni00000013/uni00000015/uni00000013/uni00000016/uni00000013/un...
https://arxiv.org/abs/2505.18680v1
TULUN : Transparent and Adaptable Low-resource Machine Translation Raphaël MerxλHanna SuominenψLois Hongζ Nick ThiebergerλTrevor CohnλEkaterina Vylomovaλ λThe University of MelbourneψThe Australian National University ψUniversity of TurkuζMaluk Timor Abstract Machine translation (MT) systems that support low-resource l...
https://arxiv.org/abs/2505.18683v1
at inference time (Brown et al., 2020) makes them particularly suitable for terminology-aware post-editing (Rau- nak et al., 2023). To address these challenges, we propose TULUN , a versatile solution that combines neural MT with LLM-based post-editing, guided by existing glos- saries and translation memories (Figure 1...
https://arxiv.org/abs/2505.18683v1
adapt to new tasks at inference time (Brown et al., 2020) makes them of interest to both MT (Moslem et al., 2023a) and related tasks, such as synthetic data generation and auto- mated post-editing (Moslem et al., 2023b). While their MT accuracy can lag behind that of special- ized MT models when translating into low-re...
https://arxiv.org/abs/2505.18683v1
set through the web UI: (1) Site metadata, includ- ing target language and site title, (2) MT model, with a choice between Google Translate or any model available on HuggingFace through its “trans- lation” pipeline,3and (3) LLM configuration for post-editing, including the choice of LLM among the hundreds of providers ...
https://arxiv.org/abs/2505.18683v1
nificant organizational expense, and while they uti- lize machine translation, MT outputs typically re- quire substantial post-editing to ensure accuracy and domain-appropriateness. For Bislama disas- ter relief translation, we partner with researchers working on a Pacific Creoles project8who need to translate transcri...
https://arxiv.org/abs/2505.18683v1
this sys- tem useful for my translation tasks”), a 4.5/5 score for the system impact on translation quality (“Us- ing this system improves the quality of my trans- lations”), and a 4.5/5 score for the system’s help- fulness to translate technical content (“Using this system makes it easier to translate technical/spe- c...
https://arxiv.org/abs/2505.18683v1
(Table 2), including a negative effect for Rundi (-1.34 ChrF++ points). Through qualitative analysis and evaluation without injecting the glos- sary in the prompt for Rundi (resulting in +0.25 points compared to NLLB), we find this is due to incorrect word changes by the LLM using the glos- sary, highlighting the need ...
https://arxiv.org/abs/2505.18683v1
that translation technologies can impact professional translators’ workflows, and TULUN ’s interface aims to give users control over the translation process while maintaining human oversight, especially for sensitive domains like health. We acknowledge that the system’s effec- tiveness will vary across languages and do...
https://arxiv.org/abs/2505.18683v1
Technologies , 1 edition. Rout- ledge, London. Gemini Team, Demis Hassabis, and Koray Kavukcuoglu. 2024a. Introducing Gemini 2.0: our new AI model for the agentic era. Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, and et al. 20...
https://arxiv.org/abs/2505.18683v1
Raphaël Merx, Aso Mahmudi, Katrina Langford, Leo Alberto de Araujo, and Ekaterina Vylomova. 2024. Low-Resource Machine Translation through Retrieval-Augmented LLM Prompting: A Study on the Mambai Language. In Proceedings of the 2nd Workshop on Resources and Technologies for Indige- nous, Endangered and Lesser-resourced...
https://arxiv.org/abs/2505.18683v1
Machine Trans- lation: A Survey. Journal of Artificial Intelligence Research , 75:351–424. Randy Scansani and Loïc Dugast. 2021. Glossary func- tionality in commercial machine translation: does it help? a first step to identify best practices for a language service provider. In Proceedings of Ma- chine Translation Summ...
https://arxiv.org/abs/2505.18683v1
mentu Tetun: ASSISTANT Sempre kontrola fali keimadura (ahi-haan) iha oras balun nia laran depois de avaliasaun dahuluk, se karik keimadura falun ona.B Usability and Usefulness Responses Statement R1 R2 1. I think that I would like to use this system frequently.5 5 2. I found the system unnecessarily complex. 1 1 3. I t...
https://arxiv.org/abs/2505.18683v1
arXiv:2505.18685v1 [cs.CL] 24 May 2025From Generation to Detection: A Multimodal Multi-Task Dataset for Benchmarking Health Misinformation Zhihao Zhang1*, Yiran Zhang1*, Xiyue Zhou2, Liting Huang3, Imran Razzak4,Preslav Nakov4,Usman Naseem1†, 1Macquarie University, Australia,2University of Sydney, Australia 3University...
https://arxiv.org/abs/2505.18685v1
misinforma- tion (Park, 2024). This presents a significant new challenge in combating the infodemic. Recent studies have focused on developing and analyzing datasets and detection methods to com- bat the infodemic. Common methods for collecting multimodal health misinformation include scrap- ing social media platforms ...
https://arxiv.org/abs/2505.18685v1
comprehensive multimodal dataset designed for detecting both human and AI generated health mis- information. Additionally, we conduct extensive experiments using state-of-the-art (SOTA) Vision- Language Models (VLLMs) to evaluate their per- formance in assessing both the reliability and orig- inality of the health info...
https://arxiv.org/abs/2505.18685v1
between January 2020 and May 2020. MMCoVaR focuses on vaccine- related misinformation between February 2020 and March 2021, featuring 958 unreliable and 1,635 reliable news articles, plus 24,184 tweets catego- rized as reliable, unreliable, or inconclusive. Med- MMHL covers multiple diseases, integrating LLM- generated...
https://arxiv.org/abs/2505.18685v1
utilising multiple existing open-source health misinformation datasets as data sources, 2) validating the available data samples and collecting human-generated multimodal data from the provided URLs, and 3) implementing gen- erative AI models to collect AI-generated repli- cated multimodal data. To ensure data quality,...
https://arxiv.org/abs/2505.18685v1
text and images based on ex- isting data without triggering their internal censor- ship mechanisms (Glukhov et al., 2023). However, some data samples are still not generated, as mod- els may refuse to produce them or output empty strings and black images. Since different mod- els are equipped with different censorship ...
https://arxiv.org/abs/2505.18685v1
noting that the average text length generated by AI is shorter than that of the original human text, with an average of around 450 words compared to approximately 850 words. Additionally, unreliable news articles tend to be shorter, averaging around 650 to 750 words, while containing more images per article around 6 to...
https://arxiv.org/abs/2505.18685v1
to interact com- pared to other deep learning models that require extensive training. We utilise both proprietary and open-source models as baselines to detect data reli- ability and originality using our MM-Health dataset and describe the benchmarking of these models. 4.1 Task Definition We designed three benchmark ta...
https://arxiv.org/abs/2505.18685v1
Hu- manEval (Chen et al., 2021a), and MMMU (Yue et al., 2024), and are widely adopted in vision and language research. These models serve as a benchmark standard, representing the top zero-shot performance that current state-of-the-art VLLMs can achieve. Open-sourced models: We in- clude three open-source VLLMs for a c...
https://arxiv.org/abs/2505.18685v1
shown in Reliable Unreliable GPT4o GPT4o-mini Llama3.2-V LLaV A-1.6 Qwen2-VL GPT4o GPT4o-mini Llama3.2-V LLaV A-1.6 Qwen2-VL ZS ZS ZS FS ZS FS ZS FS ZS ZS ZS FS ZS FS ZS FS TextHuman 0.316 0.323 0.023 0.321 0.242 0.321 0.493 0.324 0.254 0.266 0.012 0.255 0.201 0.267 0.320 0.288 AI 0.093 0.074 0.327 0.057 0.211 0.041 0....
https://arxiv.org/abs/2505.18685v1
We introduce MM-Health, a multimodal dataset designed for detecting human and AI generated health misinformation. Unlike datasets limited to human generated information or augmented with LLM generated text, our approach combines text and image generative models to create multimodal counterparts of misinformation. To en...
https://arxiv.org/abs/2505.18685v1
2021a. Evaluating large language models trained on code. Mingxuan Chen, Xinqiao Chu, and K. P. Subbalakshmi. 2021b. Mmcovar: multimodal COVID-19 vaccine focused data repository for fake news detection and a baseline architecture for classification. In ASONAM ’21: International Conference on Advances in Social Networks ...
https://arxiv.org/abs/2505.18685v1
for biomedical text mining. Bioinformatics , 36(4):1234–1240. Yichuan Li, Bohan Jiang, Kai Shu, and Huan Liu. 2020. MM-COVID: A multilingual and multimodal data repository for combating COVID-19 disinformation. CoRR , abs/2011.04088. Helena Liz-López, Mamadou Keita, Abdelmalik Taleb- Ahmed, Abdenour Hadid, Javier Huert...
https://arxiv.org/abs/2505.18685v1
Björn Ommer. 2022. High- resolution image synthesis with latent diffusion mod- els. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 , pages 10674–10685. IEEE. Maximilian Seitzer. 2020. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/ p...
https://arxiv.org/abs/2505.18685v1
news credibility research. In CIKM ’20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020 , pages 3205– 3212. ACM. Ruonan Zhu, Qing Liu, Chongru Huang, and Bingyi Kang. 2022. Z-ACM: an approximate calculation method of z-numbers for large data set...
https://arxiv.org/abs/2505.18685v1
for different tasks. Zero-shot is directly evaluated on the testing set, while five-shot uses five samples from the training set and evaluates the models on the testing set. We use standard macro F1 scores to evaluate the baseline models across all three tasks. Since our evaluation tasks involve three different data se...
https://arxiv.org/abs/2505.18685v1
arXiv:2505.18688v1 [cs.CL] 24 May 2025Large Language Models in the Task of Automatic Validation of Text Classifier Predictions Aleksandr Tsymbalov May 27, 2025 Abstract Machine learning models for text classification are trained to predict a class for a given text. To do this, training and validation samples must be pr...
https://arxiv.org/abs/2505.18688v1
Recent studies have explored using powerful LLMs in a few-shot or zero-shot manner to replace traditional classifiers. In the paper [6], the authors compared the LLM 2 with the ICL approach and fine-tuned masked language models on 77-class classification problem (on the banking77 dataset [7]). The authors showed that G...
https://arxiv.org/abs/2505.18688v1
available and described in Section 5.1. •Methods typically extract predictions from generated LLM text, making it challenging to isolateclassificationrejectionzonesclearly. Thispaperproposesaprobability-basedresponse approach in Section 3.3.2. 3 Methodology 3.1 Dataset 3.1.1 Intent Table 3.1: Client intent Client reque...
https://arxiv.org/abs/2505.18688v1
quality, the same instance is shown to multiple annotators, and an annotation is deemed correct if their answers coincide. Assume that the classification model can return K best probability predictions, then two types of annotations can be used to evaluate the model’s performance. 3.2.1 Binary Annotation LetCrepresents...
https://arxiv.org/abs/2505.18688v1
retrieving the answer from the generated text (text-approach). This approach is described in most of the research because it does not require direct access to the model’s probabilities and allows using LLM as a black box. Also, this approach provides an elegant explanation of the LLMs response, which allows modifying t...
https://arxiv.org/abs/2505.18688v1
with those of human annotators, since the models initially have access only to human-provided labels. Confidence in human annotation can be increased by collecting labels from several annotators on the same instance and retaining it only if theyallagree, thusformingahigh-qualitytestset. Let Nbethetotal number ofsamples...
https://arxiv.org/abs/2505.18688v1
(Section 6.1). 11 3.6 Prompt Prompts are presented in Section A. 3.6.1 Zero-Shot and Few-Shot LLM is capable of solving problems without any examples [19], only based on the knowl- edge learned during training. But as practice of application and many researches show, if you give some examples of problem solving [20], t...
https://arxiv.org/abs/2505.18688v1
then the LLM is modified to function as a classifier by adding a linear layer (classification head). The hidden state of the LLMs last token is fed into this head, after which the resulting LLM–classifier is trained like a standard classifier (see Section 3.5.2 for the loss function). All original LLM weights remain fr...
https://arxiv.org/abs/2505.18688v1
to select the most relevant documents and order them by importance to the query. Large documents are split into smaller ones using various techniques (by paragraph, special symbols, or sentence-similarity within a chunk). This chunking ensures each piece is small enough to include in full without further division. 15 3...
https://arxiv.org/abs/2505.18688v1