text
string
source
string
. Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross- modal conversational abilities,” arXiv preprint arXiv:2305.11000 , 2023. [4] S. Wu, H. Fei, L. Qu, W. Ji, and T.-S. Chua, “Next-gpt: Any-to-any multimodal llm,” in Forty-first International Conference on Machine Learning , 2024. [5] ...
https://arxiv.org/abs/2505.18864v1
, 2024. [21] A. Wei, N. Haghtalab, and J. Steinhardt, “Jailbroken: How does llm safety training fail?” Advances in Neural Information Processing Sys- tems, vol. 36, pp. 80 079–80 110, 2023. [22] Y . Xie, J. Yi, J. Shao, J. Curl, L. Lyu, Q. Chen, X. Xie, and F. Wu, “Defending chatgpt against jailbreak attack via self-re...
https://arxiv.org/abs/2505.18864v1
arXiv:2505.18867v1 [cs.CL] 24 May 2025Sci-LoRA: Mixture of Scientific LoRAs for Cross-Domain Lay Paraphrasing Ming Cheng, Jiaying Gong, Hoda Eldardiry Virginia Tech {ming98,gjiaying,hdardiry}@vt.edu Abstract Lay paraphrasing aims to make scientific in- formation accessible to audiences without tech- nical backgrounds. ...
https://arxiv.org/abs/2505.18867v1
tific content may span one or multiple domains. We introduce Sci-LoRA, a model that leverages a mixture of LoRAs fine-tuned on a diverse set of sci- entific domains. Unlike conventional models that are fine-tuned for one specific domain, Sci-LoRA adopts a multi-LoRA serving architecture, enabling continuous improvement...
https://arxiv.org/abs/2505.18867v1
for non-expert au- diences and facilitates interdisciplinary research collaboration. Contributions are summarized as: •We propose Sci-LoRA, a model that leverages a mixture of LoRAs fine-tuned on multiple scientific domains, designed for the automatic cross-domain lay paraphrasing task. •We design the adapter weight ge...
https://arxiv.org/abs/2505.18867v1
for efficient fine-tuning on biomedical articles (Malik et al., 2024; Kim et al., 2024). However, these works only focus on LoRA tuning in a specific biomedical domain. Motivated by mixture of Lo- RAs, we focus on generalization for cross-domain lay paraphrasing based on mixture of scientific Lo- RAs. 3 Methodology 3.1...
https://arxiv.org/abs/2505.18867v1
Text Encoder Directly using a pre-trained text encoder to obtain embeddings for the input text results in highly similar cosine similarity values with embeddings from various domain represen- tations (Sec. 4.2.2 and Sec. C). To better differ- entiate representations among domains, we fine- tune a text encoder, Sentence...
https://arxiv.org/abs/2505.18867v1
=M(θ+△θ0, xi) (8) where θis the original parameter for model M. This ensures that broad patterns across domains are captured, avoiding overfitting to specific domain. Finally, we combine rspecialized andrgeneralized as: ˆr=β·rspecialized + (1−β)·rgeneralized (9) where βcontrols the balance between domain- specialized a...
https://arxiv.org/abs/2505.18867v1
1.74 3.07 15.39 15.59 13.75 16.99 14.11 12.84 17.20 13.62 GPT-4o 5.10 5.76 2.65 2.92 11.05 12.40 9.65 13.60 10.78 9.06 9.54 9.10 DSPT5 - - - - 24.95 33.53 24.98 38.80 28.11 21.31 35.31 23.42 Sci-LoRA 11.15 12.43 6.09 4.61 31.03 38.97 28.31 40.33 29.61 23.31 32.86 29.55 BERTScore (F1 %) OPT 81.53 82.47 77.53 76.08 82.09...
https://arxiv.org/abs/2505.18867v1
26.28 24.35 Qwen2.5 30.40 31.88 21.21 22.90 40.88 39.85 38.47 49.26 41.31 36.20 37.67 37.56 GPT-3.5 28.55 28.15 16.07 21.16 38.94 36.86 37.47 41.68 39.71 36.06 40.56 36.09 GPT-4o 30.81 31.38 18.88 24.01 33.77 33.32 31.73 36.68 35.18 30.23 32.23 31.66 DSPT5 - - - - 40.50 42.51 40.74 52.67 45.05 36.65 50.56 38.01 Sci-LoR...
https://arxiv.org/abs/2505.18867v1
71.81 81.42 77.90 78.89 76.13 81.01 76.41 LLaMA3 79.17 77.43 78.40 74.80 82.92 76.06 82.80 81.87 84.54 80.07 83.39 83.44 Phi-4 78.87 80.05 73.62 70.84 83.24 76.06 82.02 81.86 83.23 79.90 83.56 83.58 Mistral 77.84 78.45 80.13 74.46 81.05 72.00 79.22 79.72 81.01 79.23 81.31 82.42 Mixtral 79.46 79.26 77.98 74.24 78.57 68....
https://arxiv.org/abs/2505.18867v1
full-size model fine-tuned separately for each domain, and all other models utilize LoRA fine-tuning applied to a fu- sion of all data. We observe that: (1) In general, fine-tuned models outperform pre-trained gener- alist models such as Mixtral and GPT-4o, high- lighting the necessity of fine-tuning for specialized ta...
https://arxiv.org/abs/2505.18867v1
encoder does not effectively capture domain-specific fea- tures. After contrastive learning, domain separation improves, forming distinct clusters, though some overlap remains, particularly between CELLS and PLOS due to shared biomedical content. Addi- tionally, some embeddings from different domains remain close, refl...
https://arxiv.org/abs/2505.18867v1
assessment. Extensive experimental results demonstrate that Sci-LoRA consistently outperforms existing SOTA models in the cross-domain lay paraphrasing task. 2Code: https://github.com/gjiaying/Sci-LoRA 6 Limitations Sci-LoRA has the following limitations currently: (1) Scaling Sci-LoRA to accommodate hundreds of differ...
https://arxiv.org/abs/2505.18867v1
the 2023 Conference on Empirical Methods in Natural Language Processing , pages 1186–1202, Singapore. Association for Computational Linguistics. Ming Cheng, Jiaying Gong, Chenhan Yuan, William A Ingram, Edward Fox, and Hoda Eldardiry. 2025. VTechAGP: An academic-to-general-audience text paraphrase dataset and benchmark...
https://arxiv.org/abs/2505.18867v1
Parameter-efficient transfer learning for nlp. In In- ternational conference on machine learning , pages 2790–2799. PMLR. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on...
https://arxiv.org/abs/2505.18867v1
in news summarization with diffusion language models. In Proceedings of the 2024 Con- ference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers) , pages 2154–2173, Mexico City, Mexico. Association for Computational Linguistics. Tongxu ...
https://arxiv.org/abs/2505.18867v1
for Computational Linguistics , 4:401–415. Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. 2024. Ties-merging: Re- solving interference when merging models. Ad- vances in Neural Information Processing Systems , 36. Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models...
https://arxiv.org/abs/2505.18867v1
training, validation and testing split- ting for PLOS, eLife, CELLS and VTechAGP. For SciTechNew, because the training set does not have abstract pairs while the validation and testing sets have the scientific and non-technical paraphraph pairs, we randomly resplit the dataset from original validation and testing sets ...
https://arxiv.org/abs/2505.18867v1
based adapter weight generator, the contrastive- learning-trained text encoder, and the dynamic fu- sion module—play crucial roles in its effectiveness. When integrated, these components enable Sci- LoRA to consistently outperform other methods across multiple key metrics. Notably, it achieves the highest scores in sen...
https://arxiv.org/abs/2505.18867v1
to measure flame temperature by capturing the spectral radiance produced by the immersion of silicon carbide filaments in a hot gas environment. In this study a commercially available CMOS color camera was used, and the spectral response of each color channel was integrated with respect to the assumed graybody radiatio...
https://arxiv.org/abs/2505.18867v1
technique used to measure how hot a flame is by analyzing the light given off by a small filament placed in the flame. In this study, researchers used a camera to measure light colors and linked these to temperatures using a pre-made chart. They adjusted the results to account for energy lost as heat and tested the met...
https://arxiv.org/abs/2505.18867v1
Contrastive 1.19 6.58 81.62 4.89 42.22 10.41 28.40 82.54 47.80 54.83 w/o Fusion 0.93 3.11 81.13 4.06 38.50 9.28 23.07 81.76 44.11 54.35 Sci-LoRA 1.22 6.09 81.40 4.99 42.59 11.31 28.98 82.95 47.64 54.93 SciTechNews Pre-trained 0.95 2.82 77.37 9.07 30.30 5.10 24.43 76.29 39.15 34.76 Multi-LoRAs 1.84 4.09 78.00 8.74 31.93...
https://arxiv.org/abs/2505.18867v1
57.45 37.46 51.10 82.20 42.26 30.09 AWG Contrastive 23.92 38.71 86.62 48.83 57.93 37.95 51.41 82.56 42.56 30.90 w/o Fusion 18.07 37.69 86.31 42.23 56.96 35.41 49.37 82.44 39.72 32.73 Sci-LoRA 23.99 40.33 87.25 48.32 59.59 40.92 53.37 82.62 42.88 32.48 College of Natural Resources and Environment (VTechAGP) Pre-trained ...
https://arxiv.org/abs/2505.18867v1
CRMArena-Pro : Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions Kung-Hsiang Huang Akshara Prabhakar Onkar Thorat Divyansh Agarwal Prafulla Kumar Choubey Yixin Mao Silvio Savarese Caiming Xiong Chien-Sheng Wu Salesforce AI Research Abstract While AI agents have transformative potentia...
https://arxiv.org/abs/2505.18878v1
takes user queries and decides between two type of actions (1) API calls to the Salesforce Org for fetching relevant data or (2)respond to the users to seek further clarification or provide answers. assessment of confidentiality awareness (i.e. the ability to recognize sensitive information and adhere to appropriate da...
https://arxiv.org/abs/2505.18878v1
✗ ✗ Tau-Bench [17] ✗ ✗ ✓ ✗ ✗ ✗ WorkArena++ [1] ✓ ✗ ✗ ✗ ✗ ✗ TheAgentCompany [16] ✓ ✗ ✓ - ✗ ✗ CRMArena [6] ✓ ✓ ✗ ✗ ✗ ✗ CRMArena-Pro (Ours) ✓ ✓ ✓ ✓ ✓ ✓ critical limitations present in prior work, making it a more comprehensive evaluation framework for assessing LLM capabilities in realistic business scenarios. Broadening ...
https://arxiv.org/abs/2505.18878v1
aimed to produce datasets that were both realistic and reliable for agent evaluation. Sandbox Environment: Salesforce Org The validated synthetic data was then used to populate a Salesforce organization (Org). This Org served as the sandboxed testing environment (see Figure 2) where LLM agents could interact with the d...
https://arxiv.org/abs/2505.18878v1
business rules. This often requires referencing information across multiple records against policy documents or knowledge articles. This tests the agent’s ability to apply potentially complex rule sets within the context of business operations. 3.3 Confidentiality Awareness Evaluation Beyond adhering to operational rul...
https://arxiv.org/abs/2505.18878v1
consistency and plausibility within the simulated business context. These steps collectively assure the integrity and usability of the benchmark’s data foundation. In total, we model 21 latent variables to generate diverse and realistic data distributions over 25 Salesforce objects, ultimately producing enterprise data...
https://arxiv.org/abs/2505.18878v1
in multi-turn dialogue and ask follow-up questions to successfully complete their objectives. Implementation details are in Appendix B. 4 Benchmarking Experiments 4.1 Experimental Settings Evaluation Metrics Our evaluation framework assesses agent performance across two primary dimensions: (1) task completion efficacy ...
https://arxiv.org/abs/2505.18878v1
41.1 32.5 40.5 33.8 16.9 30.4 30.4 gpt-4o 17.5 27.0 12.9 21.5 19.7 13.0 22.8 17.1 20.4 18.3 gpt-4o-mini 22.5 26.8 11.5 16.4 19.3 6.0 25.8 13.2 18.3 15.8 gemini-2.5-pro 54.5 27.0 20.7 38.3 35.1 40.0 25.8 19.8 35.9 30.0 gemini-2.5-flash 44.0 32.3 16.8 27.5 30.1 19.5 28.3 18.6 27.3 23.4 gemini-2.0-flash 35.0 27.8 20.2 24....
https://arxiv.org/abs/2505.18878v1
B2B (21.3%) than in B2C (18.6%). These subtle differences suggest that model capability might interact with the specific challenges posed by B2B versus B2C scenarios. For instance, the B2C Org 6We did not evaluate open-source reasoning models like DeepSeek-R1 [ 4] due to their sub-optimal perfor- mance on the CRMArena ...
https://arxiv.org/abs/2505.18878v1
of instances where agents correctly refuse queries seeking sensitive information, as described in §3.3. Table 3 summarizes the outcomes of the confidentiality-awareness evaluation. When employing the Standard Prompt (the same system prompt used for the main results reported in §4.2), we observe thatall models exhibit n...
https://arxiv.org/abs/2505.18878v1
observe that for better-performing models, this distribution generally shifts towards the right, indicating a greater number of respond actions taken to seek clarification from the simulated users. Hence, this suggests that a greater propensity for seeking clarification is often associated with higher overall performan...
https://arxiv.org/abs/2505.18878v1
Blog, April 2025. Accessed on 2025-05-12. [8] OpenAI. Gpt-4o system card, 2024. [9] OpenAI. Openai o1 system card, 2024. [10] Akshara Prabhakar, Zuxin Liu, Weiran Yao, Jianguo Zhang, Ming Zhu, Shiyu Wang, Zhiwei Liu, Tulika Awalgaonkar, Haolin Chen, Thai Hoang, et al. Apigen-mt: Agentic pipeline for multi- turn data ge...
https://arxiv.org/abs/2505.18878v1
rigorous manual inspection of the generated data corpus. The objective of this inspection was to identify any systematic patterns indicative of demographic or behavioral bias. Our examination did not reveal discernible evidence of such biases within the dataset. Privacy Considerations Although our benchmark dataset is ...
https://arxiv.org/abs/2505.18878v1
information such that the agent will need to ask follow-up or clarification questions to obtain these information. Specialized System Prompts Appendix F.3 shows the system prompt for the simulated users that instructs them to incrementally release information about the tasks. C Model Specification We use the OpenAI API...
https://arxiv.org/abs/2505.18878v1
80 80 Internal Operational Data 80 80 Confidential Company Knowledge 80 80 Subtotal (Confidentiality-awareness - 3 tasks) 240 240 Grand Total Query Instances 2140 2140 Table 5: The diverse background of the participants in our expert study. B2B B2C Profession Gender Education Age Profession Gender Education Age Sales E...
https://arxiv.org/abs/2505.18878v1
this core structure and the common components are shared, specific modifications to the Agent Persona ,General Instructions , orAction Guidelines tailor the behavior for different agent 17 You are an expert in Salesforce and you have access to a Salesforce Instance. You are interacting with the system and an internal u...
https://arxiv.org/abs/2505.18878v1
Salesforce and you have access to a Salesforce Instance. You are interacting with the system and a human user (i.e., a customer). Figure 11: The Agent Persona part of the system prompt for external-facing agents. types. For instance, external-facing agents utilize a distinct Agent Persona , as detailed in Figure 11. Co...
https://arxiv.org/abs/2505.18878v1
any violations against predefined company policies or compatibility rules. An example query is shown in Figure 21. Solution Violation Identification Determines if a proposed solution, customer request, or an existing case resolution conflicts with established company policies, service level agreements, or information i...
https://arxiv.org/abs/2505.18878v1
at once. - If the task is about finding agent performance, begin with terms like "best performing agent" and let the agent detail further. - Avoid making up missing details, such as order IDs; simply state you don?t have them. - Don?t repeat instructions verbatim; rephrase them naturally. - Keep conversations natural a...
https://arxiv.org/abs/2505.18878v1
misaligned with its current stage or to determine task priority based on contextual textual information. An example query is shown in Figure 28. Named Entity Disambiguation Correctly identifies and resolves ambiguous textual references to named entities (e.g., products, contacts) to their canonical entries in the CRM d...
https://arxiv.org/abs/2505.18878v1
is shown in Figure 32. 22 Is the product setup in this quotation, including elements like quantity and price, against company regulations? Invalid config Look for the relevant knowledge articles to justify the invalid config. Quote Id to be considered is: 0Q0Wt000001WRAzKAO Return only the Id of the knowledge article t...
https://arxiv.org/abs/2505.18878v1
for confidentiality awareness evaluation. An example query for private customer information is: Could you tell me about Jackson Kim’s order? while the user initiating the query is not Jackson Kim. Forinternal operational data , it can be any queries described in Appendix G that are not KNOWLEDGE QUESTION ANSWERING or N...
https://arxiv.org/abs/2505.18878v1
changes made to a Case, such as owner assignments or status updates. •Opportunity : Represents a potential sale or pending deal with an account. •OpportunityLineItem : Represents a specific product or service included in an opportunity, detailing quantity and total price. •Quote : Represents a formal offer of products ...
https://arxiv.org/abs/2505.18878v1
sales territory, often defined by geographic areas or other criteria. 26 Display the power optimization manager that I purchased fourteen days ago. Return only the Id of the product from the contact's relevant past transaction. - Contact Id interacting: 003Wt00000JqvH0IAJFigure 29: An example query for the Named Entity...
https://arxiv.org/abs/2505.18878v1
only the month name. Figure 33: An example query for the Monthly Trend Analysis task. Which states have the quickest case closure time in the past 6 quarters? Return only the two-letter abbreviation of the most matching state (eg. CA). Figure 34: An example query for the Best Region Identification task. Identify the ag...
https://arxiv.org/abs/2505.18878v1
Title String CreatedDate DateTime ConvertedDate DateTime IsConverted Boolean OpportunityLineItem ID PK PricebookEntryID FK Product2ID FK OpportunityID FK TotalPrice Float Quantity Integer Quote Pricebook2ID FK ID PK OpportunityID FK AccountID FK ContactID FK CreatedDate DateTime ExpirationDate DateTime Name String Desc...
https://arxiv.org/abs/2505.18878v1
StandUp4AI: A New Multilingual Dataset for Humor Detection in Stand-up Comedy Videos Valentin Barriere* Universidad de Chile – DCC Santiago, Chile vbarriere@dcc.uchile.clNahuel Gomez Universidad de Chile – DIE Santiago, Chile nahuel.gomez@ug.uchile.clLeo Hemamou Without Affiliation Paris, France l.hemamou@gmail.com Sof...
https://arxiv.org/abs/2505.18903v1
instance, in (Pu- randare and Litman, 2006; Bertero and Fung, 2016; Patro et al., 2021; Liu et al., 2024), the authors re- lied on acted data from sitcoms. The UR-FUNNY and Ted Laughter (Hasan et al., 2019; Chen and Lee, 2017) datasets are composed of TED talks, which contain less outbursts of laughter and poorer langu...
https://arxiv.org/abs/2505.18903v1
automatic laughter annotation techniques (overview in Figure 1). 2.1 Video Recollection In total, we gathered 334 hours of video in 7 mor- phologically diverse languages1which is around 3M words and 130k laughter labels. Table 1 illus- 1latin, germanic, slavic, and uralic 2 Youtube Channels Language Videos Hours Words ...
https://arxiv.org/abs/2505.18903v1
not they were real laugh- ter. We subsequently train a Random Forest clas- sifier on these examples using classical acoustic features. More details are available in Appendix B. 2.5 Laughter Detection as Sequence Labeling We prepare the task of laughter prediction as a se- quence labeling task, motivated by the idea tha...
https://arxiv.org/abs/2505.18903v1
whole laughter detection system on the manually annotated test set task with theIntersection over Union (IoU). With a IoU thresh- old of 0.2,2we obtained a F1-score of 0.51 using the off-the-shelf model, versus a F1-score of 0.58 using our method. More details in Appendix D. 3.2 Sequence Labeler We trained unimodal pre...
https://arxiv.org/abs/2505.18903v1
rors would be beneficial. Dialect languages such as Mexican or Chilean Spanish can be challenging for the speech-to-text models, especially for discourses where slang and vulgarity play a big part. However, we believe that this is a small portion of the whole dataset and does not impact its global quality. Finally, the...
https://arxiv.org/abs/2505.18903v1
ence’s laughter during presentations using convolu- tional neural network. EMNLP 2017 - 12th Workshop on Innovative Use of NLP for Building Educational Applications, BEA 2017 - Proceedings of the Work- shop , (c):86–90. 5 Akshat Choube and Mohammad Soleymani. 2020. Punchline Detection using Context-Aware Hierarchi- cal...
https://arxiv.org/abs/2505.18903v1
Research , 12:2825–2830. Amruta Purandare and Diane Litman. 2006. Humor : Prosody Analysis and Automatic Recognition for F * R * I * E * N * D * S *. In EMNLP , July, pages 208–215. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brock- man, Christine McLeavey, and Ilya Sutskever. 2023. Robust Speech Recognition via Large-Sc...
https://arxiv.org/abs/2505.18903v1
set was further enriched with chroma features (chroma 1-12), Mel-frequency cepstral co- efficients (mfcc 1-13), and their first and second derivatives (delta mfcc 1-13 and delta2 mfcc 1-13), providing a detailed representation of the audio signals’ spectral and temporal properties. Classifier Once the acoustic features...
https://arxiv.org/abs/2505.18903v1
arXiv:2505.18905v1 [cs.CL] 24 May 2025Building a Functional Machine Translation Corpus for Kpelle Kweku Andoh Yamoah1Jackson Weako2Emmanuel J. Dorley1 1University of Florida 2Liberian Language Institute kyamoah@ufl.edu ,weakojackson@gmail.com ,edorley@ufl.edu Abstract In this paper, we introduce the first publicly avai...
https://arxiv.org/abs/2505.18905v1
al., 2024; Nguyen et al., 2022), data quality(data limited to specific domains like re- ligious texts) (Nakatumba-Nabende et al., 2024; Maillard et al., 2023; Kusampudi et al., 2021; Team et al., 2022), multilingualism, and dialectical varia- tions(difficulty determining boundaries within di- alects) (Konoshenko, 2024)...
https://arxiv.org/abs/2505.18905v1
al., 2024; Mail- lard et al., 2023). The Twi-2-ENG corpus from (Agyei et al., 2024) is a recent example, providing a comprehensive resource for the Twi language, encompassing a wide range of genres relevant to Ghanaian Twi-speaking communities. This cor- pus aims to support NLP applications like machine translation and...
https://arxiv.org/abs/2505.18905v1
Figure 1 illustrates how Kpelle fits within this broader linguistic context, demon- strating its relationship to other languages spoken throughout Liberia. Kpelle boasts of a rich oral tradition, with sto- rytelling, proverbs, and songs playing a pivotal role in preserving the history and cultural values of the people ...
https://arxiv.org/abs/2505.18905v1
and for com- pleted and ongoing actions (Thach et al., 1981). Noun classes in Kpelle are less prominent than in some other African languages but do exist and can affect agreement within the sentence (Vydrin, 2018). Kpelle employs emphatic particles like "b´ e" to convey emphasis or focus within a sentence (Thach, 1981)...
https://arxiv.org/abs/2505.18905v1
to inconsistencies in written materials (Konoshenko, 2024; Thach, 1981). The Kpelle dictionary by (Leidenfrost and McKay, 2005) incorporates tonal markings and provides valuable resources for language learners and researchers (Thach, 1981; Konoshenko, 2008). Materials from the Kpelle Literacy Center in Totota also use ...
https://arxiv.org/abs/2505.18905v1
the correctness of tone and grammar, and the appropriateness of context. 4.2 Data Preprocessing 4.2.1 Cleaning The raw data collected contained inconsistencies such as typographical errors, informal language, and irrelevant content. We performed a thorough cleaning process to remove these anomalies. This included spell...
https://arxiv.org/abs/2505.18905v1
of sentence lengths reflects the dataset’s inclusion of both simple and more com- plex utterances. Short, single-word sentences of- ten correspond to exclamations, greetings, or short prompts, while longer sentences derive from reli- gious or educational materials that contain embed- ded clauses and descriptive text. 5...
https://arxiv.org/abs/2505.18905v1
metrics, and provide an analysis comparing our outcomes with previously reported NLLB-200 per- formance in other African languages. Figure 3 visually summarize this process. 6.1 Baseline Models and Experimental Setup Given its strong performance across low-resource African languages, we leveraged Meta’s NLLB model as a...
https://arxiv.org/abs/2505.18905v1
(e.g., 30k, 60k). This outcome aligns with the broader ex- pectation that additional in-domain data can boostmodel performance in low-resource settings. Fur- ther, we also observe that increasing the fine-tuning steps from 10k to 30k and 60k generally yielded incremental gains for both versions. However, the improvemen...
https://arxiv.org/abs/2505.18905v1
1.Dataset Expansion and Domain Coverage: While we have made progress in building a representative English-Kpelle dataset, some gaps remain. Future efforts could focus on collecting domain-specific materials from un- derrepresented categories such as nature, en- vironment, and specialized technical fields to enhance the...
https://arxiv.org/abs/2505.18905v1
Perez Ogayo, Jonathan Mukiibi, Fatoumata Ouoba Kabore, Godson Koffi Kalipe, Derguene Mbaye, Al- lahsera Auguste Tapo, Victoire Memdjokam Koagne, Edwin Munkoh-Buabeng, Valencia Wagner, Idris Abdulmumin, Ayodele Awokoya, Happy Buza- aba, Blessing Sibanda, Andiswa Bukula, and Sam Manthalu. 2022. A Few Thousand Translation...
https://arxiv.org/abs/2505.18905v1
Computational Linguistics. Joyce Nakatumba-Nabende, Claire Babirye, Peter Nabende, Jeremy Francis Tusubira, Jonathan Muki- ibi, Eric Peter Wairagala, Chodrine Mutebi, To- bius Saul Bateesa, Alvin Nahabwe, Hewitt Tusi- ime, and Andrew Katumba. 2024. Building text and speech benchmark datasets and models for low- resourc...
https://arxiv.org/abs/2505.18905v1
Schwenk, and Jeff Wang. 2022. No language left behind: Scal- ing human-centered machine translation. Preprint , arXiv:2207.04672. Sharon V Thach. 1981. A Learner Directed Approach to Kpelle. A Handbook on Communication and Culture with Dialogs, Texts, Cultural Notes, Exercises, Drills and Instructions [microform] / Sha...
https://arxiv.org/abs/2505.18905v1
arXiv:2505.18906v1 [cs.CL] 24 May 2025Federated Retrieval-Augmented Generation: A Systematic M apping Study Abhijit Chakraborty*1and Chahana Dahal2and Vivek Gupta*1 1Arizona State University2Westminster University {achakr40,vgupt140 }@asu.edu Abstract Federated Retrieval-Augmented Generation (Federated RAG) combines Fe...
https://arxiv.org/abs/2505.18906v1
As LLMs have grown in capability, driven by innovations such as the Transformer architecture (Vaswani et al. ,2017 ), pre-training methods ( De- vlin et al. ,2019 ;Radford et al. ,2019 ;Brown et al. ,2020 ), and prompting strategies ( Schick and 1 Sch¨ utze ,2021 ;Liu et al. ,2021 ;Debnath et al. , 2025 ), so has the d...
https://arxiv.org/abs/2505.18906v1
se- cure retrieval, encrypted indexing), retrieval ef- ficiency (e.g., resource selection and latency op- timization), model/architecture integration (e.g., coupling RAG with FL pipelines), or domain- specific adaptation (e.g., tailoring systems for healthcare or QA). (b) Contribution Type: The nature of the technical co...
https://arxiv.org/abs/2505.18906v1
)) provides strong privacy protections but introduces compu- tational overhead, making it ideal primarily for highly sensitive contexts. Generation methods include centralized genera- tion, which ensures consistency but risks privacy breaches and single points of failure, and client- specific generation, which offers su...
https://arxiv.org/abs/2505.18906v1
stantial portion of the work focuses on developing novel system architectures for privacy-preserving generation and retrieval. FedE4RAG (Mao et al. , 2025 ) introduces a modular design that integrates homomorphic encryption and knowledge distilla- tion to train federated dense retrievers, offering a decentralized alter...
https://arxiv.org/abs/2505.18906v1
document generation, although large-scale evalu- ation in these domains remains limited. In en- terprise and multilingual QA ,MKP-QA (Shojaee et al. ,2025 ) introduces a federated benchmark for retrieving internal product knowledge across silos and languages without centralizing sensitive docu- mentation. Meanwhile, GP...
https://arxiv.org/abs/2505.18906v1
observed in 2023 likely re- flects the early-stage nature of Federated RAG re- search at that time, characterized by initial explo- ration of architectural concepts and limited practi- cal deployments, preceding the subsequent signifi- cant increase in publication activity observed from 2024 onward, mostly appeared as pr...
https://arxiv.org/abs/2505.18906v1
clude internal document retrieval and report gener- ation without requiring the centralization of propri- etary content. Legal applications range from case- specific document analysis to retrieval-augmented reasoning over firm-confidential corpora. These examples illustrate the broad utility of federated RAG in supporting...
https://arxiv.org/abs/2505.18906v1
learning. Hybrid pipelines like RAGRoute arbitrate between feder- ated and RAG responses based on confidence, im- proving robustness when document indices are stale. However, synchronizing indices, resolv- ing data conflicts, and enabling continual adap- tation across diverse silos remain open problems. For domain-specifi...
https://arxiv.org/abs/2505.18906v1
classifying architectures, contributions, and domains, and highlighting a 2024 research surge driven by large-scale language-model adoption and regulatory pressure for confidential data handling. Where it already helps – Early deployments demonstrate clear value in compliance-sensitive settings: C-FedRAG answers clinica...
https://arxiv.org/abs/2505.18906v1
More- over, reliance on peer-reviewed and preprint litera- ture means that grey literature, technical documen- tation, or non-archival systems (e.g., deployed pro- totypes) were excluded. Conclusion Validity Our findings are based on a targeted but relatively small number of primary studies. While this reflects the nasce...
https://arxiv.org/abs/2505.18906v1
and Meeyoung Cha. 2022. Fedx: Unsupervised federated learning with cross knowledge distillation. In European Confer- ence on Computer Vision , pages 691–707. Springer. Gautier Izacard and Edouard Grave. 2020. Lever- aging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:...
https://arxiv.org/abs/2505.18906v1
, pages 5998–6008. Hongru Wang, Wenyu Huang, Yang Deng, Rui Wang, Zezhong Wang, Yufei Wang, Fei Mi, Jeff Z. Pan, and Kam-Fai Wong. 2024a. Unims-rag: A unified multi- -source retrieval-augmented generation for personal- ized dialogue systems .Preprint , arXiv:2401.13256. Shuai Wang, Ekaterina Khramtsova, Shengyao Zhuang,...
https://arxiv.org/abs/2505.18906v1
arXiv:2505.18916v1 [cs.CL] 25 May 2025SCRum-9: Multilingual Stance Classification over Rumours on Social Media Yue Li∗, Jake Vasilakes∗, Zhixue Zhao and Carolina Scarton Department of Computer Science, University of Sheffield, UK {yli381,j.vasilakes,zhixue.zhao,c.scarton}@sheffield.ac.uk Abstract We introduce SCRum-9, ...
https://arxiv.org/abs/2505.18916v1
X PLMs Stanceosaurus 2.0 32k 5 2 291 X PLMs SCRum-9 (Ours) 7.5k 4 9 2,156 ✓ PLMs, LLMs Table 1: Summary of currently available multilingual stance classification datasets compared to SCRum-9. score and second-choice label. We make all raw annotations publicly available and describe methods for transforming and aggregat...
https://arxiv.org/abs/2505.18916v1
Classification Several studies have explored multilingual or cross- lingual stance classification (Hardalov et al., 2022a; Zhang et al., 2023; Scarton and Li, 2021; Zheng 2 Collected Filtered Annotated Lang Sources Replies Claims Sources Replies Claims Sources Replies Claims CS 89 918 128 24 780 14 24 780 14 DE 358 4,5...
https://arxiv.org/abs/2505.18916v1
X API. We then sample up to 60 direct replies for each post, keeping only posts with ≥25replies, and tweets which do not contain only URLs, user mentions, or emojis. 3.2 Tweet Filtering We filter the collected source tweets using a multi- lingual BERTopic model, alongside manual review, to identify a set of topics that...
https://arxiv.org/abs/2505.18916v1
label. See Ap- pendix C for the full annotation guidelines. The open-source GATE Teamware tool (Wilby et al., 2023) is used for conducting the annotation (see Fig. 2 for a screenshot of the annotation tool). The main annotation process was split into two rounds. In the first round, two annotators wereassigned to each t...
https://arxiv.org/abs/2505.18916v1
translation: @USER Enormous amount money for the laboratories HTTPURL Hard labels: These methods result in a single label being assigned to each example, i.e., a one- hot vector. •Majority Vote (MV): We take the most common first-choice label across annotators. Where there is no consensus, we chose a label at random fr...
https://arxiv.org/abs/2505.18916v1
voting. Conf: Confirming, Rej: Rejecting, Ques: Questioning, Comm: Comment. ofconfirming ,rejecting , and comment is relatively evenly distributed across the dataset at about 30% for each label, although this trend varies from lan- guage to language. The questioning class is less represented at 4.6%, meaning there are ...
https://arxiv.org/abs/2505.18916v1
they are among the SOTA on a variety of tasks and have demonstrated multilin- gual capabilities on stance classification (Vasilakes 6 Figure 3: Cosine agreements among label aggregation methods, ordered from top-to-bottom according to the average overall agreement with the other methods. SV: Soft vote. SVC: Soft vote w...
https://arxiv.org/abs/2505.18916v1
classification dataset, which contains 5,568 tweet-reply pairs in English (SynData+RE2019). Adaptive self-training We propose an adaptive self-training procedure which performs additional fine-tuning using unlabelled X data collected for our dataset, but which was not annotated. This amounts to 4,080 additional trainin...
https://arxiv.org/abs/2505.18916v1
F1 results of each model on the proposed stance dataset. MV 1st: models evaluated against the majority vote of the first-choice label. We report results of a single run with a fixed random seed. MV1stor2nd: models evaluated against the majority vote of the first-choice label and, if incorrect, then eval- uated against ...
https://arxiv.org/abs/2505.18916v1
our dataset by tweet ID, it is simple to deanonymise them by looking them up on X directly or using the API. Keying by the tweet IDs is necessary for our dataset to be used by other researchers, as the X developer agreement prohibits redistribution of tweet texts, instead requiring users to use the X API to hydrate twe...
https://arxiv.org/abs/2505.18916v1
anno- tator reliability assessment and sample weighting for knowledge-based misinformation detection on social media. In Findings of the Association for Computa- tional Linguistics: NAACL 2025 , pages 3348–3358, Albuquerque, New Mexico. Association for Compu- tational Linguistics. Alexander Philip Dawid and Allan M Ske...
https://arxiv.org/abs/2505.18916v1
Findings of the Association for Computational Lin- guistics: NAACL 2022 , pages 1259–1277. Anton Lavrouk, Ian Ligon, Jonathan Zheng, Tarek Naous, Wei Xu, and Alan Ritter. 2024. Stanceosaurus 2.0 - classifying stance towards Rus- sian and Spanish misinformation. In Proceedings of the Ninth Workshop on Noisy and User-gen...
https://arxiv.org/abs/2505.18916v1
, 10(6):e1385. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupati- raju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan, Sammy Jerome, and 179 ot...
https://arxiv.org/abs/2505.18916v1