text string | source string |
|---|---|
GPT-3.5-Turbo to establish a baseline against which our custom-trained models can be compared. Despite its extensive parameter count, Model ParameterNon-Generalization Generalization pass@k↑Rgranularity ↑BLEU ↑pass@k↑Rgranularity ↑BLEU ↑ Phi 1.5 k=1 0.86 0.84 92.1 0 0.7 54.4 GPT-2 k=1 0.90 0.84 94.4 0 0.41 0.3 GPT-3.5 ... | https://arxiv.org/abs/2505.19971v1 |
the relationship be- tween natural language and SPARQL. The model’s strong performance in familiar scenarios coupled with poor generalization indicates effective pattern learning but limited transfer capability. 5.4 Qualitative Analysis Our qualitative analysis reveals distinct patterns across models. Phi-1.5 demonstra... | https://arxiv.org/abs/2505.19971v1 |
IEEE 18th International Conference on Semantic Computing (ICSC) , pages 277–284. IEEE. Debayan Banerjee, Pranav Ajit Nair, Jivat Neet Kaur, Ricardo Usbeck, and Chris Biemann. 2022. Modern baselines for SPARQL Semantic Parsing. In SIGIR . Julia Bosque-Gil, Jorge Gracia, and Elena Montiel- Ponsoda. 2017. Towards a module... | https://arxiv.org/abs/2505.19971v1 |
and Ben Bogin. 2021. Text-to-SQL in the wild: A naturally-occurring dataset based on stack exchange data. In Proceedings of the 1st Workshop on Natural Language Processing for Programming (NLP4Prog 2021) , pages 77–87, Online. Association for Computational Linguistics. Sen Hu, Lei Zou, Jeffrey Xu Yu, Haixun Wang, and D... | https://arxiv.org/abs/2505.19971v1 |
Bangkok, Thailand. Gilles Sérasset. 2012. Dbnary: Wiktionary as a LMF based multilingual RDF network. In Proceedings of the Eighth International Conference on Language Resources and Evaluation, LREC 2012, Istanbul, Turkey, May 23-25, 2012 , pages 2466–2472. Euro- pean Language Resources Association (ELRA). Tommaso Soru... | https://arxiv.org/abs/2505.19971v1 |
2: medailon gender Czech SPARQL 2: SELECT ?lexeme ?qitem ?lemma ?qitemLabel WHERE { VALUES ?lemma { 'medailon '@cs} . ?lexeme wikibase:lemma ?lemma ; wdt:P5185 ?qitem. SERVICE wikibase:label { bd:serviceParam wikibase:language 'en' } } Utterance: What is Probekörpers gender in German? B Lexicographical Data on Wikidata... | https://arxiv.org/abs/2505.19971v1 |
attributed to lexemes and identified by unique IDs (lexeme ID + -S+ decimal number as in L16168-S1 for the act of booking in the “book” lexeme L16168 ). Each sense typically includes a gloss providing a natural language definition and may have statements de- scribing relationships with other senses and items (synonyms,... | https://arxiv.org/abs/2505.19971v1 |
in medical terms 9 what is a pa c in medical terms 10 what does la stand for in medical terms 11 what does ts stand for in medical terms 12 how do you write twice a day in medical terms 13 what does dc stand for in medical terms 14 what does ta stand for in medical terms 15 what does ibm stand for in medical terms 16 w... | https://arxiv.org/abs/2505.19971v1 |
arXiv:2505.19978v1 [cs.CL] 26 May 2025DeepDialogue: A Multi-Turn Emotionally-Rich Spoken Dialogue Dataset Alkis Koudounas Politecnico di Torino Turin, Italy alkis.koudounas@polito.itMoreno La Quatra Kore University of Enna Enna, Italy moreno.laquatra@unikore.itElena Baralis Politecnico di Torino Turin, Italy elena.bara... | https://arxiv.org/abs/2505.19978v1 |
comprehensive multimodal dataset containing 40,150 high-quality multi-turn conversations across 41 diverse domains, each incorporating emotional progression across 20 distinct emotional states. Our methodology employs a systematic approach to dialogue generation by orchestrating interactions between 9 different text-on... | https://arxiv.org/abs/2505.19978v1 |
7 DailyDialog [30] 13,118 103,632 - 10 6 ABCD [12] 8,034 177,407 - 30 Chatbot Arena [14] 33,000 39,600 - 8 LMSYS-Chat-1M [51] 1,000,000 2,000,000 - * IEMOCAP [10] 151 10,039 12 * 4 DSTC2 [23] 1,612 23,354 32 1 DSTC10 [49] 107 2,292 45 3 MELD [38] 1,433 13,000 13.7 * 7 DailyTalk [29] 2,514 23,774 21.7 * Expresso [34] 39... | https://arxiv.org/abs/2505.19978v1 |
presents natural speech conversations, and MultiDialog [ 36], which extends this to multimodal settings with audio-visual 3 content. Still, many spoken datasets remain constrained by scale, domain scope, or lack fine-grained style and emotion information. Speaking Style and Emotion in Dialogues. Understanding speaking ... | https://arxiv.org/abs/2505.19978v1 |
with emotionally realistic dynamics rather than exhibiting implausible emotional shifts. For example, transitions from frustration might lead to anger ,disappointment , oranxiety , but would rarely jump directly to excitement orhappiness without intermediate states. We developed a directed graph of permissible emotion ... | https://arxiv.org/abs/2505.19978v1 |
gradual transitions between affective states. Each dialogue continues for a randomly chosen number of turns, between three and ten, to simulate natural dialogue variation. All responses are designed to resemble friendly, emotionally expressive exchanges between humans, avoiding system-level phrasing or rigid structures... | https://arxiv.org/abs/2505.19978v1 |
model evaluates the human-annotated subset using identical prompts that detail the assessment criteria. Specifically, following human annotations, we ask the model to assess each dialogue across three dimensions: (1) Coherence: logical flow, natural turn transitions, and absence of contradictions, (2) Emotional consist... | https://arxiv.org/abs/2505.19978v1 |
the two standard RA VDESS sentences to create reference audio samples, using only normal intensity recordings rather than strong intensity to avoid overemphasized emotional expressions. This integration of emotional reference samples directly into the synthesis process produces natural vocalizations that convey subtle ... | https://arxiv.org/abs/2505.19978v1 |
3 4 5 6 7 8 9 10 Number of Turns020406080100CountDecision 0 1 (a)Number of valid/invalid dialogues per number of turns. 109 8 7 6 5 4 3 Num Turns020406080Count Celeb-ritiesTradi-tionsWork Relation-shipsBooks Domain05101520Countirr,hall hall irr,hall,dom irr irr,emo,hall irr,dom dom hall,dom emo irr,emo,hall,dom(b)Inval... | https://arxiv.org/abs/2505.19978v1 |
Concrete (e.g., “ Cars ”) and Ab- stract (e.g., “ Philosophy ”). We randomly sample 1,000 dialogue turns from each domain type and compute the average concreteness score per turn by matching words against [ 8] and averaging the associated concreteness ratings for all matched words. The resulting distributions, visualiz... | https://arxiv.org/abs/2505.19978v1 |
where maintaining coherence becomes difficult even for state-of-the-art models. Our emotion taxonomy, though grounded in psychological literature, represents a practical simplification of human emotional expression, which in reality exists on continuous spectra rather than as discrete categories. Regarding speech synth... | https://arxiv.org/abs/2505.19978v1 |
Conference on Empirical Methods in Natural Language Processing , pages 5016–5026, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. [10] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemoc... | https://arxiv.org/abs/2505.19978v1 |
units. IEEE/ACM transactions on audio, speech, and language processing , 29:3451–3460, 2021. [25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276 , 2024. [26] Moreno La Qu... | https://arxiv.org/abs/2505.19978v1 |
challenges, datasets, and recent advances. IEEE access , 7:100943– 100953, 2019. [40] Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. Recipes for building an open- domain chatbot. In Paola Merlo, Jorg Tiedemann, and Re... | https://arxiv.org/abs/2505.19978v1 |
domain-grounded dialogues are generated, evaluated, and converted into speech. Section §A.1 describes the four-stage pipeline, from LLM-based generation and human evaluation to automated filtering and multimodal synthesis. Section §A.2 introduces the 41 domains used, catego- rized by concreteness and balanced across di... | https://arxiv.org/abs/2505.19978v1 |
(as assessed by GPT-4o), and the number of dialogue instances per domain. 14 Table 4: DEEPDIALOGUE ’s 41 domains, detailed descriptions, concrete vs. abstract categories (as assessed by GPT-4o), and overall final count. Domain Description Category Count Art Discussing paintings, sculptures, or artistic movements Abstra... | https://arxiv.org/abs/2505.19978v1 |
a directed graph of plausible emotional transitions. Each emotion is associated with a set of semantically or psychologically plausible next emotions. For instance, Frustrated is associated with Angry, Disappointed, Anxious, Worried, Surprised ,Curious with Surprised, Excited, Confused, Amused, Enthusiastic , and Sadwi... | https://arxiv.org/abs/2505.19978v1 |
domain. At each turn, the next emotion is chosen using the directed transition graph, modulated by the previous emotional state, the current domain, the sentiment cues in the previous message, and some variability. Transitions are not deterministic; instead, weights introduce stochasticity while preventing abrupt, impl... | https://arxiv.org/abs/2505.19978v1 |
our framework. Once the domain and initial emotional state are selected, we initiate a conversation between two LLM agents, which alternate turns in a fully autonomous manner. The agents are sampled from a pool of nine text-only instruction-tuned LLMs, yielding 16 unique pairings, including both same-model and cross-mo... | https://arxiv.org/abs/2505.19978v1 |
to a close friend, showing authentic emotion. NOTE: Do NOT include any special tokens, prefixes, or suffixes in your response. Do NOT include the prompt in your response. Strictly follow the instructions. EXAMPLE: For example, if the domain is CARS, an answer would be: “Have you heard about the new Tesla model?”. INPUT... | https://arxiv.org/abs/2505.19978v1 |
Qwen2.5-32B ⇄Phi4-14B 3627 Qwen2.5-32B ⇄Gemma3-27B 3589 Phi4-14B ⇄Gemma3-27B 3552 Qwen2.5-32B ⟲ 3379 Gemma3-27B ⟲ 3302 LLaMA3-8B ⇄Gemma3-4B 2046 LLaMA3-8B ⇄Command-r7B 1695 Phi4-14B ⟲ 1668 LLaMA3-8B ⟲ 1611 Gemma3-4B ⟲ 1270 Command-r7B ⇄Gemma3-4B 1190 Phi4-Mini ⟲ 1038 Command-r7B ⟲ 694Once assigned, voice identity remai... | https://arxiv.org/abs/2505.19978v1 |
mma n d - r 7 B _ G e mma 3 - 4 B Phi4-Mini C o m m a n d - r 7 BFigure 10: Model pairs distribution in D EEPDIALOGUE . Gemma3-4B, to ensure architectural diversity. The presence of both symmetric (same model) and asymmetric (cross-model) configurations allows for robust comparative analyses. A pie chart summa- rizing ... | https://arxiv.org/abs/2505.19978v1 |
ensuring that emotion-aware models 23 Phi4-Mini vsPhi4-Mini Gemma3-4B vsGemma3-4B Command-r7B vsGemma3-4B Command-r7B vsCommand-r7BLLaMA3-8B vsLLaMA3-8BLLaMA3-8B vsGemma3-4B LLaMA3-8B vs Command-r7BQwen2.5-32B vsQwen2.5-32BGemma3-27B vsGemma3-27BQwen2.5-32B vsGemma3-27BPhi-4 vs Gemma3-27BQwen2.5-32B vsPhi-4 Qwen2.5-72B... | https://arxiv.org/abs/2505.19978v1 |
the top-10 domains with the highest gender-PMI differences, while the second block highlights the top-10 domains with the greatest age-PMI differences. Domain M/F ratio PMI M PMIF∥PMIM−F∥Y/O ratio PMI Y PMIO∥PMIY−O∥ Sports 0.00 0.49 -7.44 7.93 1.88 2.47 -1.35 3.82 Coding 0.00 0.49 -6.64 7.13 0.05 -1.23 0.11 1.34 Philos... | https://arxiv.org/abs/2505.19978v1 |
} Male : {Man, Boy, Gentleman, Father, He, Him, Sir, Son } Young : {Young, Teen, Child, Kid, Baby } Old: {Old, Elder, Senior, Grandpa, Grandma, Aged } 25 dialogues. Annotations were collected on a voluntary basis from researchers familiar with dialogue systems rather than through paid crowdsourcing platforms. For each ... | https://arxiv.org/abs/2505.19978v1 |
scale, and checkboxes for specific quality issues when dialogues are marked negative. This example shows a conversation about classic movies in the Celebrities domain. configurations, finding that our ensemble of three open-source models outperformed individual models in alignment with human judgments. Statistics and p... | https://arxiv.org/abs/2505.19978v1 |
I'm on top of the world, my coding skills are finally paying off! That's amazing! I've been working on some tough bugs too. It feels so rewarding when it all clicks, doesn't it? Haha, I know that feeling! It's like a rush of adrenaline, and suddenly all the hours of debugging make sense! Absolutely! Those moments are pu... | https://arxiv.org/abs/2505.19978v1 |
a lightweight multi-layer perceptron (MLP) classifier. The classifier consists of two hidden layers with 128 units each, LayerNorm, GELU activation, and dropout (dropout probability = 0.1) applied between layers. Training Details. We only trained the classifier head for a maximum of 10 epochs, with an early stopping pa... | https://arxiv.org/abs/2505.19978v1 |
How Well Do Large Reasoning Models Translate? A Comprehensive Evaluation for Multi-Domain Machine Translation Yongshi Ye1,3,*, Biao Fu2,3,*,†, Chongxuan Huang2,3, Yidong Chen2,3, Xiaodong Shi2,3,† 1Institute of Artificial Intelligence, Xiamen University 2School of Informatics, Xiamen University 3Key Laboratory of Digit... | https://arxiv.org/abs/2505.19987v1 |
⇔English setting, covering a wide range of content—from general-purpose domains such as conversation and news to highly specialized ar- eas like law, medicine, and literature. Our evalua- tion combines automatic metrics (BLEU, COMET, CometKiwi) with an LLM-powered scoring proto- col based on the Multidimensional Qualit... | https://arxiv.org/abs/2505.19987v1 |
these methods have mainly focused on general-domain or single- domain settings, and a systematic evaluation of reasoning capabilities in MDMT remains lacking. 2.2 Evaluation for LLM-based MDMT LLMs have shown strong performance in general- domain MT, prompting growing interest in their ap- plicability to domain-specifi... | https://arxiv.org/abs/2505.19987v1 |
stylistically demanding content such as literature. This design ensures high diversity in domain coverage, textual style, and se- mantic complexity, enabling a comprehensive and multidimensional evaluation of model performance. Detailed data statistics are provided in Appendix A. Metrics. In this experiment, we evaluat... | https://arxiv.org/abs/2505.19987v1 |
accurate seman- tic representation. However, in structurally rigid and terminology-intensive domains such as Med- ical, Law, IT, and Koran, LRMs achieve lower COMET scores than traditional LLMs. This may be attributed to their tendency to prioritize seman- tic plausibility over strict adherence to domain- specific term... | https://arxiv.org/abs/2505.19987v1 |
Style, Terminology, Others, Source Error, and Non-translation Error. In MQM score calcula- tion and error category statistics, Source Error are excluded (Freitag et al., 2021). As shown in Table 1, traditional LLMs achieve lower overall MQM scores, showing higher consis- tency and precision. Interestingly, LRMs achieve... | https://arxiv.org/abs/2505.19987v1 |
better resolve semantic ambiguity in Chinese and plan conceptual mappings into English, ensuring the naturalness and accuracy of terminology in the En- glish context. These results suggest that LRMs hold strong potential in terminology translation. In contrast, traditional LLMs outperform LRMs in the De⇒En direction, w... | https://arxiv.org/abs/2505.19987v1 |
IT, Law, and Medical domains. As shown in Table 6, 6 ModelConversation E-commerce News Social Literary BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI Document-Level GPT-4o 40.00 84.71 82.02 35.33 87.35 64.55 36.07 86.60 66.01 33.75 84.36 68.70 17.55 77.57 77.14 DeepSeek-V3 39.28 84.51 8... | https://arxiv.org/abs/2505.19987v1 |
main and adjust their translation accordingly.PromptIT Law Medical BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI P1 36.66 83.49 78.68 39.16 84.97 82.35 40.23 83.82 81.81 P2 34.20 82.94 79.33 41.46 85.01 83.46 38.92 83.38 82.57 P3 36.55 83.57 79.82 40.57 85.32 83.54 41.14 84.06 82.97 Table 6: Translation performance a... | https://arxiv.org/abs/2505.19987v1 |
culty. For example, LRMs should generate accu- rate translations efficiently with minimal reasoning for simple inputs to avoid unnecessary overthink- ing (Chen et al., 2025b), whereas for more complex sentences, they should engage in deeper analysis and structured reasoning to ensure semantic accu- racy and completenes... | https://arxiv.org/abs/2505.19987v1 |
annotations rely on a single model as the scorer, which ensures consistency but may introduce bias aligned with that model’s own translation behavior. References Roee Aharoni and Yoav Goldberg. 2020. Unsupervised domain clusters in pretrained language models. In Proceedings of the 58th Annual Meeting of the Asso- ciati... | https://arxiv.org/abs/2505.19987v1 |
Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxi- ang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y . K. Li, Y . Q. Wang, Y . X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu... | https://arxiv.org/abs/2505.19987v1 |
Yuheng Zou, Yujia He, Yukun Zha, Yunfan Xiong, Yunxian Ma, Yuting Yan, Yux- iang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Z. F. Wu, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhen Huang, Zhen Zhang, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhibin Gou, Zhicheng Ma, Zhigang Yan, Zhihong Shao, Zhipeng Xu, Zhiyu... | https://arxiv.org/abs/2505.19987v1 |
quality of machine translation. In Findings of the Association for Computational Linguistics: ACL 2024 , pages 3546–3562, Bangkok, Thailand. Associ- ation for Computational Linguistics. Zhen Huang, Haoyang Zou, Xuefeng Li, Yixiu Liu, Yuxiang Zheng, Ethan Chern, Shijie Xia, Yiwei Qin, 10 Weizhe Yuan, and Pengfei Liu. 20... | https://arxiv.org/abs/2505.19987v1 |
Codispoti, An- drew Galu, Andrew Kondrich, Andrew Tulloch, An- drey Mishchenko, Angela Baek, Angela Jiang, An- toine Pelisse, Antonia Woodford, Anuj Gosalia, Arka Dhar, Ashley Pantuliano, Avi Nayak, Avital Oliver, Barret Zoph, Behrooz Ghorbani, Ben Leimberger, Ben Rossen, Ben Sokolowsky, Ben Wang, Benjamin Zweig, Beth ... | https://arxiv.org/abs/2505.19987v1 |
de Castro, Mikhail Pavlov, Miles Brundage, Miles Wang, Mi- nal Khan, Mira Murati, Mo Bavarian, Molly Lin, Murat Yesildal, Nacho Soto, Natalia Gimelshein, Na- talie Cone, Natalie Staudacher, Natalie Summers, Natan LaFontaine, Neil Chowdhury, Nick Ryder, Nick Stathas, Nick Turley, Nik Tezak, Niko Felix, Nithanth Kudige, ... | https://arxiv.org/abs/2505.19987v1 |
Hyung Won Chung, Ian Kivlichan, Ian O’Connell, Ian Osband, Ignasi Clavera Gilaberte, Ilge Akkaya, Ilya Kostrikov, Ilya Sutskever, Irina Kofman, Jakub Pachocki, James Lennon, Jason Wei, Jean Harb, Jerry Twore, Jiacheng Feng, Jiahui Yu, Jiayi Weng, Jie Tang, Jieqi Yu, Joaquin Quiñonero Candela, Joe Palermo, Joel Parish, ... | https://arxiv.org/abs/2505.19987v1 |
Li, and Pengfei Liu. 2024. O1 repli- cation journey: A strategic progress report – part 1. Preprint , arXiv:2410.18982. Shanghaoran Quan, Jiaxi Yang, Bowen Yu, Bo Zheng, Dayiheng Liu, An Yang, Xuancheng Ren, Bofei Gao, Yibo Miao, Yunlong Feng, et al. 2025. Codeelo: Benchmarking competition-level code generation of llms... | https://arxiv.org/abs/2505.19987v1 |
Ma, Weiyu Chen, Yvette Graham, Bonnie Webber, Philipp Koehn, Andy Way, Yulin Yuan, and Shuming Shi. 2023b. Findings of the WMT 2023 shared task on discourse-level literary translation: A fresh orb in the cosmos of LLMs. In Proceedings of the Eighth Conference on Machine Translation , pages 55–67, Singapore. Association... | https://arxiv.org/abs/2505.19987v1 |
Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2024. Multilingual machine translation with large language models: Empirical results and anal- ysis. In Findings of the Association for Computa- tional Linguistics: NAACL 2024 , pages 2765–2781, Mexico City, Mexico. Association for Computational Linguistics. 14... | https://arxiv.org/abs/2505.19987v1 |
causing semantic distortion. Addition Adding extra information or emotions not in the source. Under-translation Failure to fully convey cultural or contextual nuances. Omission Unintentional exclusion of content from the source text. Untranslated Retaining source text without translation. Hallucination Generating conte... | https://arxiv.org/abs/2505.19987v1 |
25.21 83.04 81.10 27.29 83.77 80.64 25.93 83.36 79.95 DeepSeek-V3 30.26 84.96 79.67 19.47 81.55 78.32 25.77 83.00 81.18 27.55 83.56 80.64 25.76 83.27 79.95 Gemini-2.0-Flash 32.79 84.46 79.19 19.97 81.54 77.90 26.62 83.04 80.93 29.25 83.24 80.16 27.16 83.07 79.55 OpenAI-o1 29.52 84.95 80.90 17.82 81.81 79.16 22.52 82.96... | https://arxiv.org/abs/2505.19987v1 |
31.07 84.71 81.30 32.02 84.66 81.00 33.36 85.16 81.28 Gemini-2.0-Flash-Thinking 36.35 86.05 81.79 33.05 85.77 81.14 31.59 85.83 82.42 31.48 84.76 81.28 33.12 85.60 81.66 Table 20: Results on the De ⇒En WMT22 General Machine Translation task. 17 ModelWMTBio18 En ⇒Zh WMTBio18 Zh ⇒En WMTBio18 En ⇒De WMTBio18 De ⇒En Averag... | https://arxiv.org/abs/2505.19987v1 |
KIWI BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI BLEU COMET KIWI GPT-4o 20.49 78.44 76.50 16.12 77.55 77.82 21.08 78.91 77.31 12.49 75.38 76.91 17.54 77.57 77.13 DeepSeek-V3 15.66 77.47 77.84 16.25 77.14 77.49 19.57 78.94 77.48 11.87 75.28 77.05 16.81 77.48 77.17 Gemini-2.0-Flash 22.11 78.25 75.94 16.25 77.14 77.49... | https://arxiv.org/abs/2505.19987v1 |
Prediction: {prediction_text} Figure 4: Full prompt used to calculate MQM scores with DeepSeek-V3. 19 Your task is to assess the difficulty of translating a given {src_lang} sentence into {tgt_lang} . Please evaluate the difficulty based on the following criteria and output the result in JSON format, with the key "leve... | https://arxiv.org/abs/2505.19987v1 |
arXiv:2505.19997v1 [cs.LG] 26 May 2025Embracing Imperfection: Simulating Students with Diverse Cognitive Levels Using LLM-based Agents Tao Wu1, Jingyuan Chen2†, Wang Lin1, Mengze Li3, Yumeng Zhu2,Ang Li1,Kun Kuang1,Fei Wu1†, 1College of Computer Science and Technology, Zhejiang University, 2College of Education, Zhejia... | https://arxiv.org/abs/2505.19997v1 |
shown in Figure 1, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 152.53.03.54.04.55.0Cognitive Score Real Student Our Simulated Student Naive Prompt Simulated StudentFigure 2: Cognitive scores of 15 different students. The naive prompt-based simulations tend to produce overly advanced responses. In contrast, our method adaptively s... | https://arxiv.org/abs/2505.19997v1 |
results prove that our method overcomes the bias of behavioral sim- ulations and can produce realistic simulations. 2 Related Work 2.1 LLM-based Education Simulation Recent research explores using large language mod- els (LLMs) to simulate educational roles, support- ing teaching strategy evaluation (Markel et al., 202... | https://arxiv.org/abs/2505.19997v1 |
each student. We choose programming as the task-solving scenario due to its complexity, error diversity and real-world applicability (Dai et al., 2024, 2025), making it a representative setting for student simulation3. To ensure a stable cognitive state, only records com- pleted within a week are included in each seque... | https://arxiv.org/abs/2505.19997v1 |
students complete a task, they engage with knowledge con- cepts at various levels, from foundational concepts (e.g., basic coding grammar) to advanced ones ( e.g., algorithm design). However, relying solely on the task statement tiand student solution sioften lim- its the extraction to foundational concepts, as ad-vanc... | https://arxiv.org/abs/2505.19997v1 |
will effectively support subse- quent behavior prediction and solution simulation.4.3 Behavior Prediction Once the student’s cognitive prototype is con- structed, the next step is predicting their behavior on a new task. As shown in Figure 3, a common approach retrieves the most similar task from past learning records.... | https://arxiv.org/abs/2505.19997v1 |
behavior and the retrieved past learning record to improve alignment. During each refinement step, Bcandidate solutions are sampled to increase the likelihood of achieving a successful refinement. Each candidate is evaluated by model πvalue , which assigns a score between 0 and 1 based on its alignment with the predict... | https://arxiv.org/abs/2505.19997v1 |
Random Similarity Level Level+Random Level+Similarity Prototype Mapping Solution Simulation IO CoT Refine IO CoT Refine IO CoT Refine IO CoT Refine IO CoT Refine IO CoT Refine LLaMA-3.3- 70B-InstructAcc 0.37 0.37 0.37 0.41 0.41 0.41 0.39 0.39 0.39 0.4 0.4 0.4 0.43 0.43 0.43 0.61 0.61 0.61 Con 12.29 2.29 2.29 2.45 2.45 ... | https://arxiv.org/abs/2505.19997v1 |
Please note that since AccandCon 1exclusively assess the accuracy and consistency of behavior prediction, their values are invariant across differ- ent solution simulation methods in Table 2 and 3. This design ensures a clean separation of evaluation for the two stages in our framework. Please refer to Appendix D for e... | https://arxiv.org/abs/2505.19997v1 |
student solutions leads to a performance drop (Row 3), highlighting the importance of clear and accurate behavior descriptions in guiding solution simula- tion effectively. 4) Removing the self-refinement process leads to poor performance in Row 4, indi- cating the importance of iterative refinement . How- ever, Row 5 ... | https://arxiv.org/abs/2505.19997v1 |
student cog- nitive ability scores (described in Appendix A.2), as shown in Figure 5. The figure reveals a clear positive correlation between simulation quality and cognitive ability, indicating that simulating students with higher cognitive abilities is relatively easier. This finding aligns with expectations—students... | https://arxiv.org/abs/2505.19997v1 |
this challenge is not unique to our ap- proach—baseline methods are also susceptible to such noise. While techniques like causal in- ference and debiasing could help mitigate the impact of confounding factors, fully addressing this issue falls beyond the scope of our current work. We leave this as a direction for futur... | https://arxiv.org/abs/2505.19997v1 |
Xiu Tang, Sai Wu, Chang Yao, Zhipeng Gao, and Jingyuan Chen. 2025. Less is more: Adaptive program repair with bug localization and preference learning. In AAAI-25, Sponsored by the Association for the Ad- vancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA , pages 128– 136. AAAI Pre... | https://arxiv.org/abs/2505.19997v1 |
Zhou Zhao, Fei Wu, Chang Yao, and Jingyuan Chen. 2024. E3: Exploring embodied emotion through A large- scale egocentric video dataset. In Advances in Neural Information Processing Systems 38: Annual Confer- ence on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024 .... | https://arxiv.org/abs/2505.19997v1 |
2021, virtual . Alexis Ross and Jacob Andreas. 2024. Toward in- context teaching: Adapting examples to students’ misconceptions. In Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pages 13283–13310. Asso... | https://arxiv.org/abs/2505.19997v1 |
Quoc V . Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Ad- vances in Neural Information Processing Systems 35: Annual Conference on Neural Information Process- ing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 . Hanyi Xu, Wensheng ... | https://arxiv.org/abs/2505.19997v1 |
. . . . . . . 20 E.6 Distinctions from Knowledge Trac- ing . . . . . . . . . . . . . . . . . 22 A Dataset Statistics and Details A.1 Privacy Protection To ensure privacy protection, we excluded all per- sonal and sensitive data, such as student names and email addresses, retaining only anonymized unique student IDs as ... | https://arxiv.org/abs/2505.19997v1 |
programming problem according to this ability level. Output your code directly.Problem:{problem} Figure 9: Details of the naive prompt. closely with the goals of student simulation, where iterative behavior prediction and solu- tion refinement are central. We believe these characteristics make program- ming tasks a str... | https://arxiv.org/abs/2505.19997v1 |
to 5, while 1 indicates the lowest ability and 5 indicates the highest ability. Directly output your prediction. If you think the student will make mistakes, provide possible details about the mistakes.-Output Format-Error Prediction: (Yes/No)Error Description: (Your detailed analysis)-Real Data-Problem:{problem}Figure... | https://arxiv.org/abs/2505.19997v1 |
number of local analyses remains compu- tationally manageable. Upper limit on knowledge concepts . The rela- tionship extraction process (Step 2) focuses on an- alyzing the relationships between knowledge con- cepts within each task to form the edges of the knowledge graph. To avoid redundancy and ensure manageability,... | https://arxiv.org/abs/2505.19997v1 |
Details of the prompt for Con 1metric. for multilingual dialogue use cases and outper- forms many of the available open source and closed chat models on common industry benchmarks. The test model version is LLaMA-3.3-70B-Instruct. Claude is a large language model developed by Anthropic, designed to generate human-like ... | https://arxiv.org/abs/2505.19997v1 |
level. As mentioned in Benedetto et al. (2024), this relative cog- nitive level simulation approach yields bet- ter results. We incorporate this level into the prompt to indicate the student’s proficiency, enabling the model to perform behavior pre- diction accordingly. The prompt is shown in Figure 11. •Level+Random ,... | https://arxiv.org/abs/2505.19997v1 |
2.65 2.17 2.48 2.73 2.27 3.61 3.65 3.83 Table 5: End-to-end comparison across 6 behavior prediction and 3 solution simulation methods on the whole 100 students in Student_100 . Behavior Prediction Random Similarity Level Level+Random Level+Similarity Prototype Mapping Solution Simulation IO CoT Refine IO CoT Refine IO ... | https://arxiv.org/abs/2505.19997v1 |
Metrics We further evaluated the results of our solution simulation using captioning metrics, specifically ROUGE-L (Lin, 2004) and BLEU-4 (Papineni et al., 2002). The results, shown in Tables 6 and 7, indicate that our prototype mapping and self- refinement methods consistently achieve the best performance in most case... | https://arxiv.org/abs/2505.19997v1 |
evaluate behavior descriptions, yielding identical values for the same behavior prediction method. However, we believe that captioning metrics are not a suitable evaluation measure in the context of student simulation. The core focus of our task eval- uation is to assess whether we can accurately and reasonably simulat... | https://arxiv.org/abs/2505.19997v1 |
3.0 1.76 1.76 1.76 2.22 2.22 2.22 2.52 2.52 2.52 3.16 3.16 3.16 Con 22.62 2.26 2.26 2.52 2.36 2.62 1.94 1.76 1.48 1.86 1.88 2.1 2.46 2.18 1.88 2.98 3.1 3.16 GPT-3.5Acc 0.5 0.5 0.5 0.52 0.52 0.52 0.48 0.48 0.48 0.44 0.44 0.44 0.44 0.44 0.44 0.58 0.58 0.58 Con 12.62 2.62 2.62 2.62 2.62 2.62 2.36 2.36 2.36 2.24 2.24 2.24 ... | https://arxiv.org/abs/2505.19997v1 |
recognize that predicting these occasional careless mistakes remains a challeng- ing problem. Our method, however, is primar- ily designed to assess general cognitive profi- ciency—capturing a student’s overall learning pat- terns rather than isolated, occasional errors. Fully addressing this issue would require additi... | https://arxiv.org/abs/2505.19997v1 |
arXiv:2505.20006v1 [cs.CL] 26 May 2025Mixture of LoRA Experts for Low-Resourced Multi-Accent Automatic Speech Recognition Rapha ¨el Bagat1, Irina Illina1, Emmanuel Vincent1 1Universit ´e de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France raphael.bagat@loria.fr, irina.illina@loria.fr, emmanuel.vincent@inria.fr Abstr... | https://arxiv.org/abs/2505.20006v1 |
the model’s input also showed promising improvements [9]. However, these methods considered native accents only. To bridge the gap with non-native accents, [10] used various transfer learning methods to improve non-native multi-accent ASR, exhibiting the importance of a multilingual model to han- dle pronunciation diff... | https://arxiv.org/abs/2505.20006v1 |
the computational cost at inference time. 2.2. Mixture of Accent-Specific LoRAs (MAS-LoRA) We propose MAS-LoRA, a MoE method using LoRA experts trained on single-accent data and combined at inference time to process multi-accent data. In detail, if the training data con- tains naccents, we instantiate nLoRA experts, on... | https://arxiv.org/abs/2505.20006v1 |
and Decoder columns indicate the fine-tuning method used in the encoder and the decoder. The percentage of trained parameters is with respect to the total model size. Bold numbers indicate the best result for each corpus and those results which are statistically equivalent to it. Encoder Decoder Trained params. (%) WER... | https://arxiv.org/abs/2505.20006v1 |
is set to start at 1e-5 for full fine-tuning and 5e-5 for parameter-efficient fine-tuning meth- ods, and decreases linearly to its half throughout the fine-tuning. Fine-tunings have been conducted on NVIDIA A100 GPUs and tests on NVIDIA V100 GPUs. For decoding, greedy search is used for computational reasons. Our code ... | https://arxiv.org/abs/2505.20006v1 |
important to note that when MAS-LoRA is used in the encoder, the decoder has to be fine-tuned. Not fine-tuning the decoder degrades the results, especially when MAS-LoRA-qv is used in the encoder. Performance on native speech — After fine-tuning the models on non-native speech, we tested them on native speech to evalua... | https://arxiv.org/abs/2505.20006v1 |
on the task of improving ASR when facing multiple non-native accents. We introduced Mixture of Accent-Specific LoRAs, a fine-tuning method based on a mix- ture of LoRA experts. Each expert specializes in a specific ac- cent, and their combined knowledge is used at inference. We showed that, when the accent is unknown a... | https://arxiv.org/abs/2505.20006v1 |
Y . Qian, X. Gong, and H. Huang, “Layer-wise fast adaptation for end-to-end multi-accent speech recognition,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 30, pp. 2842–2853, 2022. [14] E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptat... | https://arxiv.org/abs/2505.20006v1 |
arXiv:2505.20013v1 [cs.CL] 26 May 2025WEBCOT: Enhancing Web Agent Reasoning by Reconstructing Chain-of-Thought in Reflection, Branching, and Rollback Minda Hu♣♠*, Tianqing Fang♠∗, Jianshu Zhang♥, Junyu Ma♠, Zhisong Zhang♠, Jingyan Zhou♣, Hongming Zhang♠, Haitao Mi♠, Dong Yu♠, Irwin King♣ ♣Chinese University of Hong Kon... | https://arxiv.org/abs/2505.20013v1 |
patterns into agents (Chen et al., 2024; Zhao et al., 2024; Hu et al., 2025) combines the adaptability of learned policies with the interpretability and task- aware heuristics of curated reasoning, mitigating both the overthinking problem and the exploration burden of pure RL. In this paper, we carefully examine and de... | https://arxiv.org/abs/2505.20013v1 |
the performance of open-source models in web-based tasks. Addi- tionally, several studies focus on equipping agents with self-improvement mechanisms (Fang et al., 2025; Aksitov et al., 2023; Patel et al., 2024; Zhang et al., 2025c), enabling models to iteratively refine their strategies through bootstrapped learning. T... | https://arxiv.org/abs/2505.20013v1 |
ment evolves based on: st+1=T(st, at), o t+1= Ω(st+1),(2) producing a trajectory τ={(oi, hi, ai)}T i=1, where Tis the total number of steps.3.2 Optimization We adopt a self-improvement optimization frame- work as in OpenWebV oyager (He et al., 2024b). We introduce the backbone agent foundation model, de- noted as M, al... | https://arxiv.org/abs/2505.20013v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.