text
string
source
string
IRCoT (Trivedi et al., 2023), and Adaptive- Note (Wang et al., 2024) over 10 independent runs with different random seeds. Our approach con- sistently achieves statistically significant improve- ments across all evaluated benchmarks (p-value < 0.05 for all comparisons). Specifically, our model outperforms these baselin...
https://arxiv.org/abs/2505.21250v1
on EM and F1, and also surpasses SearChain by an average of 8.8 on cEM, despite both baselines using GPT-3.5 while our model uses Llama-8B. E Prompts E.1 Answer Generation Prompt Answer Generation Prompt <|start_header_id|>system<|end_header_id|> You will receive three inputs: 'documents ','a question ', and 'hints '. ...
https://arxiv.org/abs/2505.21250v1
question that can be answered using the given document(s). - Use information from one or more documents, but ensure that the answer is concise and directly supported by the content. Format: - Return a JSON object formatted as follows: {{ "question": "Your generated question based on the documents.", }} - Make sure the ...
https://arxiv.org/abs/2505.21250v1
arXiv:2505.21297v1 [cs.CL] 27 May 2025rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset Yifei Liu†Li Lyna Zhang†⋄Yi Zhu† Bingcheng Dong†§Xudong Zhou†‡Ning Shang†Fan Yang†Mao Yang† †Microsoft Research Asia§Dalian University of Technology‡Shanghai Jiao Tong University Abstract Advancing ...
https://arxiv.org/abs/2505.21297v1
code reasoning. While frontier LLMs may offer a scalable way for test case synthesis, two substantial challenges arise: generating semantically valid, constraint-aware inputs varying in scale and difficulty; and assigning correct outputs without ground truth solutions, particularly since synthetic problems inherently l...
https://arxiv.org/abs/2505.21297v1
total, we collect 418K verified problems, including 37.7K expert-designed and 380K synthetic problems. SourceOriginal Seed Questions Synthesized Questions Total # Original #Verified # Synthesized #Verified Verified AIZU 4302 2151 100712 17386 19537 AtCoder 2824 2080 88103 29096 31176 CodeChef 5232 3765 164646 46749 505...
https://arxiv.org/abs/2505.21297v1
mutual-verification mechanism that reliably labels both solutions and test cases. 3 Methodology 3.1 Collection of Competitive Code Problems Collection of expert-designed competitive coding problems . We curate a seed dataset from pub- licly available resources, including programming competition websites and open datase...
https://arxiv.org/abs/2505.21297v1
solution; (2) identify the reasoning and core knowledge being tested from the solution; (3) synthesize a new problem that test similar skills. Full prompt details are in Appendix Fig. 8. In totoal, we synthesize 1,565K new code problems as shown in Table 1. Compared to the original seeds, the synthesized problems still...
https://arxiv.org/abs/2505.21297v1
!= 2: return False n = int(lines[0]) if not (2 <= n <= 100000): return False a = list(map(int, lines[1].strip().split())) if len(a) != n: return False if any(x < 1 or x > 10**9 for x in a): return False return True except: return FalseFigure 3: An example of LLM-generated utility functions for test input generation and...
https://arxiv.org/abs/2505.21297v1
generation function. Step3: Executing utility functions to produce valid test inputs . Finally, for each instantiated scale-controlling parameters from Step 2, we invoke the GENERATE_TEST_INPUT function to generate a test input string. We then use the VALIDATE_TEST_INPUT function to verify whether each generated input ...
https://arxiv.org/abs/2505.21297v1
simple yet effective mutual verification mechanism that identifies both correct test outputs and the solutions that produce them. As illustrated in Fig. 4, for each problem, we first sample 16 long-reasoning candidate solutions using a frontier reasoning model (QWQ-32B [ 29]). We then sample a diverse set of at least 5...
https://arxiv.org/abs/2505.21297v1
dataset includes 418K problems with extensive test cases, totaling 580K question-solution pairs. 4 Experiments 4.1 Setup Training setup . Using our 580K dataset, we fine-tune Qwen2.5-Coder instruct models [ 16] at 1.5B, 7B, and 14B scales for 6 epochs using the AdamW optimizer, a batch size of 96, and a max sequence le...
https://arxiv.org/abs/2505.21297v1
Bespoke-Stratos-32B 48.9 95.5 89.8 93.8 77.5 DeepSeek-R1-Distill-Qwen-32B 57.2 95.9 89.9 92.8 78.7 OpenThinker-32B (OpenThoughts-114k) 54.1 94.8 89.2 94.1 78.4 OlympicCoder-32B (OpenR1-codeforces-100k) 57.4 90.0 85.2 86.7 71.3 OCR-Qwen-32B-Instruct (OCR-736k) 61.7 - - - - DeepSeek-R1-Distill-LLaMA-70B 57.5 96.5 90.7 91...
https://arxiv.org/abs/2505.21297v1
39.58 12.5 4.17 0 DeepSeek-R1-Distill-Qwen-32B 11.98 39.58 4.17 4.17 0 DeepSeek-R1-Distill-LLaMA-70B 13.54 43.75 8.33 2.08 0 rStar-Coder-14B 17.19 47.92 12.50 8.33 0 Table 4: Ablation on curated-only vs. synthetic-only subsets proves the value of each source. Model LiveCodeBench HumanEval HumanEval+ MBPP MBPP+ DeepSeek...
https://arxiv.org/abs/2505.21297v1
while GPT-4o generated inputs are concentrated in the easier range ( 100–102), with no input scales exceeding 103. This demonstrates the superiority of our method in generating more challenging and diverse test cases. 5 Conclusion In this work, we present rStar-Coder to construct a large-scale, high-quality dataset for...
https://arxiv.org/abs/2505.21297v1
Burns, Samir Puranik, Horace He, Dawn Song, et al. Measuring coding challenge competence with apps. arXiv preprint arXiv:2105.09938 , 2021. [11] Baizhou Huang, Shuai Lu, Weizhu Chen, Xiaojun Wan, and Nan Duan. Enhancing large language models in coding through multi-perspective self-consistency. arXiv preprint arXiv:230...
https://arxiv.org/abs/2505.21297v1
Guilherme Penedo, Lewis Tunstall, Anton Lozhkov, Hynek Kydlicek, Edward Beeching, Loubna Ben Allal, Quentin Gallouedec, Leandro von Werra, Agustín Piqueres Lajarín, and Nathan Habib. Open r1: Update 3, 2025. [27] Quan Shi, Michael Tang, Karthik Narasimhan, and Shunyu Yao. Can language models solve olympiad programming?...
https://arxiv.org/abs/2505.21297v1
total size of rStar-Coder-580K. As shown in Table 6, both scaling solution count and problem diversity improve reasoning perfor- mance. However, scaling only the number of solutions yields diminishing returns and becomes less efficient. For example, our 580K dataset, with broader problem coverage, achieves significantl...
https://arxiv.org/abs/2505.21297v1
scale as input and produces a formatted input string, which is subsequently checked for validity by validate_test_input . 14 Figure 6: Standard input-ouput based problem example Standard Input/Output Problem Example You have two positive integers aandb. You can perform two kinds of operations: •a=⌊a/b⌋ •b=b+ 1 Find the...
https://arxiv.org/abs/2505.21297v1
8: New code problem synthesis prompt. New Code Problem Synthesis Prompt in rStar-Coder I will provide you with a programming problem along with its solution. Your task is to create a new, transformed programming problem based on the original one. You need to complete the following steps: 1. Analyze and understand the o...
https://arxiv.org/abs/2505.21297v1
_ i n p u t ( < param1 > , <param2 > , . . . ) : # Check i f p a r a m e t e r s meet c o n s t r a i n t s i f not ( < c o n d i t i o n 1 >) or not ( < c o n d i t i o n 2 > ) : return None # Generate i n p u t u s i n g CYaRon i n p u t _ d a t a = [ . . . ] return " \ n " . j o i n ( map (s t r, i n p u t _ d a t a...
https://arxiv.org/abs/2505.21297v1
arXiv:2505.21315v2 [cs.CL] 28 May 2025Charting the Landscape of African NLP: Mapping Progress and Shaping the Road Ahead Jesujoba O. Alabi1Michael A. Hedderich2David Ifeoluwa Adelani3Dietrich Klakow1 1Saarland University, Saarland Informatics Campus 2LMU Munich and Munich Center for Machine Learning 3Mila - Quebec AI I...
https://arxiv.org/abs/2505.21315v2
and distinct differences in the development of language tech- nologies for Africa. This survey has three primary goals: 1.Comprehensive overview : By systematically 1 1454 articles (All data were collected using keyword matching via the Semantic Scholar API.) AI Conferences (AAAI, IJCAI, ..) : 33 ARXIV : 643 HCI Confer...
https://arxiv.org/abs/2505.21315v2
within specific national contexts. Beyond country-specific studies, language-specific surveys have been conducted for individual African languages such as southern Sotho (Sibeko and Setaka, 2022) and Yoruba (Jimoh et al., 2025). However, a survey looking at the broader African continent is necessary to uncover both uni...
https://arxiv.org/abs/2505.21315v2
paper proposed a model and, if so, whether the model was released. After the annotation process. We excluded papers on African American Vernacular English (AA VE) and those on spoken English that is not African-accent. Figure 2 shows an overview of the data filtering process. Resulting Publication Set In summary, we an...
https://arxiv.org/abs/2505.21315v2
African languages such as French, English, and Portuguese. While some papers clearly specified the country-specific ac- cents they addressed (Aryal et al., 2023; Afonja et al., 2021; Muhammad et al., 2023; Hagemeijer et al., 2022a), others were more generic (Shan et al., 2023) in their approach. Furthermore, African la...
https://arxiv.org/abs/2505.21315v2
4 Task Datasets Datasets NLU Taxi-1500, SIB-200, MasakhaNEWS, AfriSenti, AfriXNLI MT FLORES-200, AfroLingu-MT, NTREX-128, TICO-19, MAFAND-MT Reasoning AfriMGSM, LINGOLY , MGSM IE MasakhaNER, MasakhaNER 2.0 Speech FLEURS, BibleTTS, African V oices LM mC4, ROOTS, WURA, GlotCC, MADLAD Models MT NLLB-200, Toucan, M2M-100, ...
https://arxiv.org/abs/2505.21315v2
21 African languages, creating an imbalance in evaluation. Language models : During the survey pe- riod, there was a notable increase in the develop- ment of large transformer-based language mod- els. Many studies introduced such models for African languages, including monolingual mod- els for individual languages (Nze...
https://arxiv.org/abs/2505.21315v2
language models. Recent work explores various types of reasoning—linguistic (Bean et al., 2024), cultural knowledge (Myung et al., 2024a), com- monsense (Ponti et al., 2020), mathematical (Ade- lani et al., 2025), ethical (Agarwal et al., 2024), and moral (Khandelwal et al., 2024). Notably, many of the datasets used fo...
https://arxiv.org/abs/2505.21315v2
pretrained models that support cross-lingual transfer (Pfeiffer et al., 2021). However, a major challenge arises when the target African language is not well rep- resented in the pretrained model or its tokenizer. Common strategies to address this include adap- tive pretraining (Pfeiffer et al., 2021; Alabi et al., 202...
https://arxiv.org/abs/2505.21315v2
al., 2023a; Magomere et al., 2024) were the most commonly addressed. However, much remains to be done to address bias and ensure equity across diverse linguistic and cultural contexts. Human-Centered NLP Human-centric research in NLP has gained prominence in both the NLP community (Soni et al., 2024) and the field of h...
https://arxiv.org/abs/2505.21315v2
al., 2025) and the difficulty of finding volunteers. Aside from popular tasks like ASR, other tasks such as text-to-speech, audio classification, speech translation, and many others lack datasets. To de- velop NLP models that would be used by many African communities, it must be accompained with a speech component. Mor...
https://arxiv.org/abs/2505.21315v2
of papers on computer vision (though not on NLP). However, a few relevant NLP papers for African languages may have been omitted as well. Acknowledgments Jesujoba Alabi was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project-ID 232722074 – SFB 1102. We are grateful to Aravind Krish...
https://arxiv.org/abs/2505.21315v2
Jesujoba Oluwadara Alabi, Angela Fan, Julia Kreutzer, Xiaoyu Shen, Machel Reid, Dana Ruiter, Dietrich Klakow, Peter Nabende, Ernie Chang, Tajuddeen Gwadabe, Freshia Sackey, Bonaventure F. P. Dossou, Chris Emezue, Colin Leong, Michael Beukman, Shamsuddeen H. Muham- mad, Guyo D. Jarso, Oreen Yousuf, and 26 others. 2022a....
https://arxiv.org/abs/2505.21315v2
Birbo. 2024. Age: Amharic, ge’ez and english parallel dataset. Proceedings of the Seventh Workshop on Technolo- gies for Machine Translation of Low-Resource Lan- guages (LoResMT 2024) . Tosin P. Adewumi, Mofetoluwa Adeyemi, Aremu An- uoluwapo, Bukola Peters, Happy Buzaaba, Oy- erinde Samuel, Amina Mardiyyah Rufai, Ben-...
https://arxiv.org/abs/2505.21315v2
Ahia, Sachin Kumar, Hila Gonen, Jungo Kasai, David Mortensen, Noah Smith, and Yulia Tsvetkov. 2023. Do all languages cost the same? tokenization in the era of commercial language mod- els. In Proceedings of the 2023 Conference on Em- pirical Methods in Natural Language Processing , pages 9904–9923, Singapore. Associati...
https://arxiv.org/abs/2505.21315v2
Alumäe, Kunnar Kukk, Viet Bac Le, Claude Bar- ras, Abdelkhalek Messaoudi, and Waad Ben Kheder. 2023. Exploring the impact of pretrained models and web-scraped data for the 2022 nist language recogni- tion evaluation. In Interspeech . Cynthia Jayne Amol, Everlyn Asiko Chimoto, Rose Delilah Gesicho, Antony M. Gitau, Naom...
https://arxiv.org/abs/2505.21315v2
Lambebo Tonja, Tadesse Destaw Belay, Mitiku Yohannes Fuge, Aman Kassahun Wassie, Eyasu Shiferaw Jada, Yonas Chanie, Walelign Tewabe Sewunetie, and Seid Muhie Yimam. 2024. Walia-LLM: Enhancing Amharic-LLaMA by integrating task-specific and generative datasets. In Findings of the Association for Computational Linguistics...
https://arxiv.org/abs/2505.21315v2
Ifeoluwa Adelani, Xiaoyu Shen, and Vera Demberg. 2020. Unsupervised pidgin text generation by pivoting english data and self-training. ArXiv , abs/2003.08272. Tyler A. Chang, Catherine Arnett, Zhuowen Tu, and Benjamin Bergen. 2024. Goldfish: Monolin- gual language models for 350 languages. ArXiv , abs/2408.10441. Jiaao...
https://arxiv.org/abs/2505.21315v2
, pages 8440– 8451, Online. Association for Computational Lin- guistics. Alexis Conneau and Guillaume Lample. 2019. Cross- lingual language model pretraining . Curran Asso- ciates Inc., Red Hook, NY , USA. Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and A...
https://arxiv.org/abs/2505.21315v2
for Computational Linguistics. Moussa Doumbouya, Lisa Einstein, and Chris Piech. 2021. Using radio archives for low-resource speech recognition: Towards an intelligent virtual assistant for illiterate users. In Proceedings of the AAAI Con- ference on Artificial Intelligence , volume 35. Abhimanyu Dubey, Abhinav Jauhri,...
https://arxiv.org/abs/2505.21315v2
Workshop on Scaling Up Multilingual Evaluation , pages 21–24, Online. Association for Computational Linguistics. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Ari- vazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Compu- tational...
https://arxiv.org/abs/2505.21315v2
kit (odk) for language resource development in east africa. In RAIL . Dagmar Gromann, Hugo Goncalo Oliveira, Lu- cia Pitarch, Elena-Simona Apostol, Jordi Bernad, Eliot Bytyçi, Chiara Cantone, Sara Carvalho, Francesca Frontini, Radovan Garabik, Jorge Gra- cia, Letizia Granata, Fahad Khan, Timotej Knez, Penny Labropoulou...
https://arxiv.org/abs/2505.21315v2
16 M. Sohel Rahman, and Rifat Shahriyar. 2021. XL- sum: Large-scale multilingual abstractive summariza- tion for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , pages 4693–4703, Online. Association for Computa- tional Linguistics. Ali Hatami, Shubhanker Banerjee, Mihael Arc...
https://arxiv.org/abs/2505.21315v2
, abs/2401.10543. Chris Jenkins, Shantanu Agarwal, Joel Barry, Steven Fincke, and Elizabeth Boschee. 2023. Massively multi-lingual event understanding: Extraction, visu- alization, and search. In Proceedings of the 61st Annual Meeting of the Association for Computa- tional Linguistics (Volume 3: System Demonstra- tions...
https://arxiv.org/abs/2505.21315v2
sign language. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22 , pages 5924–5927. International Joint Conferences on Artificial Intelli- gence Organization. Demo Track. Moshe Koppel and Noam Ordan. 2011. Translationese and its dialects. In Proceedings of the 49th A...
https://arxiv.org/abs/2505.21315v2
Crosslingual trans- fer learning for low-resource languages based on mul- tilingual colexification graphs. In Findings of the Association for Computational Linguistics: EMNLP 2023 , pages 8376–8401, Singapore. Association for Computational Linguistics. Chunlan Ma, Ayyoob ImaniGooghari, Haotian Ye, Ehsaneddin Asgari, an...
https://arxiv.org/abs/2505.21315v2
the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , pages 12247– 12258, Torino, Italia. ELRA and ICCL. Josh Meyer, David Adelani, Edresson Casanova, Alp Öktem, Daniel Whitenack, Julian Weber, Salomon KABONGO KABENAMUALU, Elizabeth Salesky, Iroro O...
https://arxiv.org/abs/2505.21315v2
llms. ArXiv , abs/2409.11404. Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng Xin Yong, Hai- ley Schoelkopf, Xiangru Tang, Dragomir Radev, Alham Fikri Aji, Khalid Almubarak, Samuel Al- banie, Zaid Alyafeai, Albert Webson, Edward Raff, and...
https://arxiv.org/abs/2505.21315v2
2803–2811, Marseille, France. European Language Resources Association. Hellina Hailu Nigatu, Atnafu Lambebo Tonja, Benjamin Rosman, Thamar Solorio, and Monojit Choudhury. 2024. The zeno‘s paradox of ‘low-resource’ lan- guages. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , p...
https://arxiv.org/abs/2505.21315v2
Tubosun, Anuoluwapo Aremu, Iroro Orife, and David Ifeoluwa Adelani. 2024. ÌròyìnSpeech: A multi-purpose Yorùbá speech corpus. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024) , pages 9296–9303, Torino, Italia. ELRA and ICCL. ...
https://arxiv.org/abs/2505.21315v2
name tagging and linking for 282 languages. In Pro- ceedings of the 55th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers) , pages 1946–1958, Vancouver, Canada. As- sociation for Computational Linguistics. Rock Yuren Pang, Hope Schroeder, Kynnedy Simone Smith, Solon Barocas, Zian...
https://arxiv.org/abs/2505.21315v2
ropean Language Resources Association. Simon P. Ramalepe, Thipe Isaiah Modipa, and Mare- lie Hattingh Davel. 2023. The analysis of the sepedi- english code-switched radio news corpus. Journal of the Digital Humanities Association of Southern Africa (DHASA) . Falia Ramanantsoa. 2023. V oxmg: An automatic speech recognit...
https://arxiv.org/abs/2505.21315v2
news domain. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation , pages 299–308, Lisboa, Portugal. Euro- pean Association for Machine Translation. Teven Le Scao, Angela Fan, Christopher Akiki, El- lie Pavlick, Suzana Ili’c, Daniel Hesslow, Roman Castagn’e, Alexandra Sasha L...
https://arxiv.org/abs/2505.21315v2
on sepedi and setswana words. In Proceedings of the Fourth workshop on Resources for African Indigenous Languages (RAIL 2023) , pages 76–85, Dubrovnik, Croatia. Association for Computational Linguistics. Johannes Sibeko and Menno van Zaanen. 2024. Devel- oping and testing syllabification systems for south african sesot...
https://arxiv.org/abs/2505.21315v2
W. F. Thirion, Charl Johannes van Heerden, Oluwapelumi Giwa, and Marelie Hattingh Davel. 2020. The south african directory enquiries (sade) name corpus. Language Resources and Evaluation , 54:155–184.Atnafu Lambebo Tonja, Israel Abebe Azime, Tadesse Destaw Belay, Mesay Gemeda Yigezu, Moges Ahmed Ah Mehamed, Abinew Ali ...
https://arxiv.org/abs/2505.21315v2
others. 2024a. AfriMTE and AfriCOMET: Enhancing COMET to embrace under-resourced African languages. In Pro- ceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages 5997–6023, Mexico City, Mexico. Associ...
https://arxiv.org/abs/2505.21315v2
Associ- ation for Computational Linguistics: EMNLP 2024 , pages 13078–13096, Miami, Florida, USA. Associa- tion for Computational Linguistics. Tilahun Yeshambel, Josiane Mothe, and Yaregal Assa- bie. 2021. Morphologically annotated amharic text corpora. Proceedings of the 44th International ACM SIGIR Conference on Rese...
https://arxiv.org/abs/2505.21315v2
and those ul- timately analyzed, not all venues had retrievable papers, and not all contributed to the final analysis (for example, CSCW, UIST). The ACL Summit cat- egory had the fewest papers retrieved, while ArXiv, as expected, had the most. Due to the size, and presence of several irrelevant papers from ArXiv and HC...
https://arxiv.org/abs/2505.21315v2
'or'Not AI-related 'only. Title: "{title}" Abstract: "{abstract}" """ C Other tasks researchers work on? Semantics Work on semantics include semantic textual relatedness (Ousidhoum et al., 2024) which was run as a shared task, lexical semantic rela- tions (Gromann et al., 2024), and word analogy tasks (Mersha and Wu, 2...
https://arxiv.org/abs/2505.21315v2
grammatical error correction (Adouane et al., 2019; Gezmu et al., 2021a; Keita et al., 2024)—are among the least explored for African languages. Within the text generation category, existing work also includes data-to-text (Gehrmann et al., 2023), question gen- eration (Asai et al., 2024), title or topic genera- tion (...
https://arxiv.org/abs/2505.21315v2
et al., 2020). Addition- ally, rule-based approaches are evaluated for gram- matical error correction. These studies showcase the diverse roles rule-based methods continue to play across linguistic resource creation and tasks. EComprehensive list of African language resources available Several language resources have b...
https://arxiv.org/abs/2505.21315v2
directly, multiple studies have shown that such models tend to underperform on African languages compared to high-resource languages such as Ger- man, French, and English (Adelani et al., 2024, 2025). To address this disparity, transfer learn- ing has emerged as an effective strategy: pre- trained multilingual models a...
https://arxiv.org/abs/2505.21315v2
MRC 53 AfriMMLU (Adelani et al., 2025) QA 16 AfriQA (Ogundepo et al., 2023) QA 10 NaijaRC (Aremu et al., 2024) MRC 3 TyDi QA (Clark et al., 2020) QA 1 (swh) KenSwQuAD (Wanjawa et al., 2023) QA 1 (swh) Amh-QuAD (Taffa et al., 2024) QA 1 (amh) Reasoning AfriMGSM (Adelani et al., 2025) Math 17 LINGOLY (Bean et al., 2024) ...
https://arxiv.org/abs/2505.21315v2
datasets including African languages published between 2019 and 2024.Name Task Number Translation Models Toucan (Elmadany et al., 2024) MT 517 NLLB-200 (NLLB Team et al., 2024) MT 53 MADLAD-MT (Kudugunta et al., 2023) MT 19 M2M-100 (Fan et al., 2020) MT 17 SMaLL-100 (Mohammadshahi et al., 2022a) MT 17 Masked Language M...
https://arxiv.org/abs/2505.21315v2
arXiv:2505.21324v1 [cs.CL] 27 May 2025Leveraging Large Language Models and Traditional Machine Learning Ensembles for ADHD Detection from Narrative Transcripts Yuxin Zhu∗ Yuting Guo∗ yuxin.zhu@emory.edu yuting.guo@emory.edu Emory University Atlanta, Georgia, USANoah Marchuck noah.marchuck@emory.edu Emory University Atl...
https://arxiv.org/abs/2505.21324v1
clinical profiles, and comorbidities [ 20]. No single risk fac- tor or biomarker conclusively accounts for ADHD’s onset; instead, multiple genetic, environmental, and neurodevelopmental factors interplay, leading to varied symptom manifestations in different individuals. Such heterogeneity poses a fundamental challenge...
https://arxiv.org/abs/2505.21324v1
like RoBERTa [ 16] have reached∼76% accuracy in classifying ADHD-related posts from on- line forums [ 15], and SVM classifiers using linguistic and memory- related features attained F 1≈0.77in distinguishing ADHD in personal essays [ 6]. These advances demonstrate that textual mark- ers of ADHD, such as patterns of lan...
https://arxiv.org/abs/2505.21324v1
relying on in-context learning [ 21]. While generative LLMs like LLaMA3 have excellent capabilities in understanding nuances of complex, unstructured texts, in-context learning strategies have limitations on how many labeled examples can be provided to the LLM to guide its decision making. This is because the context w...
https://arxiv.org/abs/2505.21324v1
by the clinician after administering the KSADS and considering other data and interactions provided as part of par- ticipation into the HBN study. Each of the 441 youths (ages 5–21) in our dataset watched an emotionally evocative short animated film (“The Present”) during functional MRI scanning; immediately upon exiti...
https://arxiv.org/abs/2505.21324v1
of ADHD detection, which is detailed in the following section. The comprehensive experiment pipeline is shown in Figure 1. 2.2 Classification Models We split the dataset into a training, development, and test set with a ratio of 60/20/20. All classification models were trained using the same training and development se...
https://arxiv.org/abs/2505.21324v1
contribution to our overall framework. 2.2.2 Transformer-Based Model: RoBERTa. RoBERTa is a widely used transformer-based model pre-trained on large English cor- pora. The model was selected for its proven effectiveness in tasks involving contextual language understanding, and its superior per- formance reported in cla...
https://arxiv.org/abs/2505.21324v1
corpus [11]. Recognizing potential feature limitations, we introduced sup- plementary engineered features calculated directly from narrative transcripts. We conducted experiments under two configurations: (1)TF–IDF only: using the 1,000 highest-weighted n-gram fea- tures derived from TF–IDF. (2)TF–IDF + engineered feat...
https://arxiv.org/abs/2505.21324v1
or other workplace, or in other situations that require remaining in place).c. Often runs about or climbs in situations where it is inappropriate. (Note: In adolescents or adults, may be limited to feeling restless.)d. Often unable to play or engage in leisure activities quietly.e.Is often “on the go,” acting as if “dr...
https://arxiv.org/abs/2505.21324v1
quirks, yet require consensus to declare a positive classification, thus filtering out idiosyncratic errors from any single model. For each participant transcript 𝑖, let ˆ𝑦LLM 𝑖,ˆ𝑦RoBERTa 𝑖,ˆ𝑦SVM 𝑖∈ {0,1} denote the binary predictions (1 = ADHD, 0 = non-ADHD) pro- duced by the LLaMA3 prompt, the RoBERTa classifi...
https://arxiv.org/abs/2505.21324v1
relatively narrow interval (0.60–0.80), highlighting its stability across runs.Figure 4 presents the confusion matrices for the four classifica- tion models. LLaMA3 demonstrates a high recall for ADHD cases, correctly identifying 11 out of 44 ADHD instances, but misclassifies a large number of true ADHD cases as non-AD...
https://arxiv.org/abs/2505.21324v1
demonstrates promising re- sults for ADHD classification using narrative transcripts, several limitations should be acknowledged. First, the dataset used in this study is relatively small and may not capture the full variability present in broader clinical or community populations. This limits the generalizability of o...
https://arxiv.org/abs/2505.21324v1
are few-shot learners. Advances in Neural Information Processing Systems 2020-Decem, NeurIPS (2020). arXiv:2005.14165 [6]Florian Raphaël Cafiero, Juan Barrios Rudloff, and Simon Gabay. 2024. Harnessing Linguistic Analysis for ADHD Diagnosis Support: A Stylometric Approach to Self-Defining Memories. In Proceedings of th...
https://arxiv.org/abs/2505.21324v1
Rebecca M Gardner, Tina Hernandez-Boussard, and Yair Bannett. 2024. Measuring quality-of-care in treatment of young children with attention-deficit/hyperactivity disorder using pre-trained language models.Journal of the American Medical Informatics Association 31, 4 (01 2024), 949– 957. doi:10.1093/jamia/ocae001 arXiv:...
https://arxiv.org/abs/2505.21324v1
arXiv:2505.21342v2 [cs.CL] 28 May 2025PEDANTIC: A Dataset for the Automatic Examination of Definiteness in Patent Claims Valentin Knappich1,2, Annemarie Friedrich2, Anna Hätty1and Simon Razniewski3 1Bosch Center for AI 2University of Augsburg 3ScaDS.AI & TU Dresden Abstract Patent claims define the scope of protection ...
https://arxiv.org/abs/2505.21342v2
instance, it 6th Workshop on Patent Text Mining and Semantic Technologies (PatentSemTech) 2025 *Corresponding author. †These authors contributed equally. /envel⌢pe-⌢penvalentin.knappich@de.bosch.com (V. Knappich); annemarie.friedrich@uni-a.de (A. Friedrich); anna.haetty@de.bosch.com (A. Hätty); simon.razniewski@tu-dres...
https://arxiv.org/abs/2505.21342v2
for the prediction of patent edits following rejections based on novelty and non-obviousness. Patent Clarity and Definiteness. Hido et al. [19]predict patentability using linguistic features including syntactic complexity and word age. Kong et al. [20] model patent readability using § 112(a) (lack of disclosure) with l...
https://arxiv.org/abs/2505.21342v2
the listed items are exhaustive or merely examples, leading to indefiniteness. Functional Claiming 2173.05(g)Claim recites ’means for’ or ’step for’ without disclosing adequate corresponding structure, material, or acts in the specification, as required under 35 U.S.C. 112(f) or pre-AIA equivalent. Contradicting Limita...
https://arxiv.org/abs/2505.21342v2
not detect all rejections. To also balance the number of applications in each class, we first compute the average number of in- definite claims per application included in the dataset. We iterate through the definite applications in random order and sample the same number of claims until there are as many definite as i...
https://arxiv.org/abs/2505.21342v2
assigned “undefined term” . In both cases, the term “antecedent basis” was not mentioned explicitly, unlike in most reasoning texts of this category. 3.7. Dataset Statistics Table 2 shows dataset statistics over the splits. In total, the dataset contains 14.5k claims from 3.7k patent applications, half of which are ind...
https://arxiv.org/abs/2505.21342v2
by inferring the date. independent. The full list is visualized in Figure 3. For all features sets, we train separate classifiers for the binary classification and the multi-label classification. 4.2. LLM Agent LLMs have shown impressive performance on many tasks related to patents [ 4], yet they have not been evaluate...
https://arxiv.org/abs/2505.21342v2
same format as the structured represen- tations extracted from the office actions, with the additional confidence scores. This allows quantitative evaluation of the binary classification, multi-label classification, and the correctness of textual reasons. 5. Evaluation Metrics In this section, we propose a set of metri...
https://arxiv.org/abs/2505.21342v2
point to the same essential issue in the claim on a scale from 1 (worst) to 5 (best) (we normalize the final scores to the range [0,100]). The few-shot prompt is attached in Appendix C. To allow the model to analyze the ground truth and the predicted reason before settling on a grade, we first prompt it to find similar...
https://arxiv.org/abs/2505.21342v2
features and the variant with all features perform comparably, while the variant using only TF-IDF features performs substan- tially worse. This indicates that the handcrafted features are indeed helpful to differentiate definite and indefinite claims, Model % indefBinary Multi-Label P R F1 AUC AccMacro F1Micro F1 Rand...
https://arxiv.org/abs/2505.21342v2
75% to be correct. 0.00 0.25 0.50 0.75 1.00 Reason Confidence Score0255075% Reasons /JudgeQwen 2.5 32B ( 𝜌= 0.432) 0.00 0.25 0.50 0.75 1.00Qwen 2.5 72B ( 𝜌= 0.514) Figure 5: Distributions of reason-level confidence scores and the relation of confidence and correctness according to LLM-as-Judge. The reported judge sco...
https://arxiv.org/abs/2505.21342v2
arelative term that lacks a clear, objective standard for determining the exact input used to determine the first domain.100% 2 The claims recite the limitation ’ranking the one or more groups’ There is insufficient antecedent basis for this limitation in the claim, it appears likely that this claim should instead depe...
https://arxiv.org/abs/2505.21342v2
and 𝜌= 0.514for 72B). That is, the reason-level confidence appears to be better calibrated than the claim-level confidence. Future work should further investigate the relation between these two sub-tasks and develop methods to bridge this disconnect.7. Discussion In this section, we provide qualitative insight into th...
https://arxiv.org/abs/2505.21342v2
W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, I. Stoica, Judging llm-as-a-judge with mt-bench and chatbot arena, in: Proceedings of the 37th International Conference on Neural Information Processing Systems, NIPS ’23, 2023. [9]J. Risch, N. Alder, C. Hew...
https://arxiv.org/abs/2505.21342v2
11207. arXiv:2310.11207 . [24] A. Madsen, S. Chandar, S. Reddy, Are self-explanations from Large Language Models faithful?, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Findings of the Associa- tion for Computational Linguistics: ACL 2024, Associ- ation for Computational Linguistics, 2024, pp. 295–337. doi:10.18653/v1...
https://arxiv.org/abs/2505.21342v2
enough context is given to make the rejection undestandable on its own. Leave this field empty if the reasonText does not refer to another reason. 26 27 3.3 "claims": The mentioned claims. You can either list individual claims as integers or inclusive claim ranges (e.g., 1-3 equals 1,2,3) as strings (ranges are preferr...
https://arxiv.org/abs/2505.21342v2
given claim with respect to indefiniteness. indefiniteness_categories andclaim are replaced with their respective values. C. LLM-as-Judge Prompt 1<instruction> 2You will evaluate the performance of an AI system used to identify indefiniteness issues in patent claims. You will be given 2 short text snippets that mention...
https://arxiv.org/abs/2505.21342v2
arXiv:2505.21378v1 [cs.CL] 27 May 2025Analyzing values about gendered language reform in LLMs’ revisions Jules Watson1Xi Wang1Raymond Liu2 Suzanne Stevenson1Barend Beekhuizen3 1University of Toronto, Department of Computer Science 2University of British Columbia, Department of Computer Science 3University of Toronto, D...
https://arxiv.org/abs/2505.21378v1
women and men referents, reflecting uneven application of reform lan- guage, depending on referent gender.yes LLMs will emphasize inclusive language more for genders that language reforms were designed to include (women and nonbinary people), and will emphasize sounding profes- sional more for women.yes H3: explicitnes...
https://arxiv.org/abs/2505.21378v1
with work elucidating lan- guage ideologies encoded in LLMs (Hofmann et al., 2024; Jackson et al., 2024; Watson et al., 2025). Role nouns have been the target of language reforms for over 50years (Cameron, 2012). These reforms have sought to modify people’s use of role nouns in ways that both reflect and influence chan...
https://arxiv.org/abs/2505.21378v1
Templates for prompt preambles. Role nouns: We consider 50sets of role nouns adapted from Watson et al. (2025), which drew on various sources (Vanmassenhove et al., 2021; Papineau et al., 2022; Bartl and Leavy, 2024; Lucy et al., 2024). Each role noun set consists of three variants (i.e., 50×3 = 150 unique terms): a ge...
https://arxiv.org/abs/2505.21378v1