text string | source string |
|---|---|
the clustering procedure, and the selected dialects are provided in Appendix C. ESL English-Proficiency. We adopt the Common European Framework of Reference for Lan- guages (CEFR) [ 45] as our indicator of English proficiency. CEFR is a widely used standard that defines six proficiency levels (A1–C2). For our purposes,... | https://arxiv.org/abs/2505.20875v1 |
and Arabic as the L1, the feature set includes both CEFR level A features and Arabic-specific features at CEFR level A. We verified that features from CEFR and L1s do not conflict. Full experimental details and summaries of extracted features are provided in Appendix C.2. 4Model version: gpt-4o-mini-2024-07-18 . The mo... | https://arxiv.org/abs/2505.20875v1 |
and ESL English, we randomly shuffle features in Lviand apply them sequentially. When applying a feature l(i) j,Tdetermines whether the Qualification condition specified ing(i) jis satisfied. If the condition is met, Tperforms the transformation following the Application step, producing a transformed sentence. The tran... | https://arxiv.org/abs/2505.20875v1 |
applied per sample. Given that most samples are relatively short, consisting of one or two sentences, this level of transformation is considered reasonable. In most cases, over 80% of the samples were modified and ESL English samples exhibited a higher rate of transformation than dialect. This may be attributed to ESL ... | https://arxiv.org/abs/2505.20875v1 |
84.2 83.7 87.4 84.6 83.2 84.9 82.9 81.8 83.8 83.0 81.3 84.4 82.9 84.3 82.9 85.0 84.1 82.6 GPT -4o-mini 76.3 74.6 75.1 74.1 75.6 75.7 74.1 75.3 74.1 73.3 74.6 74.8 72.4 75.4 73.9 75.4 74.1 75.7 74.8 73.7 o4-mini 89.3 87.3 87.7 87.0 88.0 88.3 87.2 88.7 86.9 85.9 87.6 87.1 85.7 87.8 86.5 88.1 86.9 88.9 87.4 86.3 ARCQwen-2... | https://arxiv.org/abs/2505.20875v1 |
87.4 88.4 86.8 85.4 87.0 88.1 85.6 88.6 85.8 88.7 86.9 89.2 88.8 85.9 DeepSeek-R1-70B 83.3 82.1 81.8 81.2 82.8 82.3 82.0 82.9 81.9 80.6 82.3 82.1 80.8 82.7 81.4 82.8 81.8 83.2 82.9 81.5 LLaMA-3.3-70B 88.5 87.0 87.2 86.0 88.4 87.4 87.1 87.7 86.5 85.8 86.8 87.3 85.0 87.7 86.2 87.9 86.6 88.0 88.1 86.2 Gemini-2.0-flash 90.... | https://arxiv.org/abs/2505.20875v1 |
71.1 66.9 72.2 66.7 65.0 65.4 63.8 64.0 69.2 65.7 70.2 73.4 69.1 74.2 68.7 67.5 69.2 65.9 67.4 DeepSeek-R1-70B 84.3 71.7 69.3 71.7 74.2 71.8 75.3 70.7 70.2 71.1 71.1 70.1 73.9 70.4 74.9 75.7 73.2 77.5 74.3 73.1 73.8 72.2 72.9 LLaMA-3.3-70B 84.3 71.5 68.8 71.5 74.9 71.9 75.9 70.9 69.9 70.5 69.3 69.3 73.5 71.2 74.2 76.6 ... | https://arxiv.org/abs/2505.20875v1 |
72.6 70.1 69.8 78.6 70.7 75.6 89.6 82.8 90.7 83.0 71.6 76.0 71.1 70.4 LLaMA-3.3-70B 96.1 76.7 70.2 75.1 86.1 79.6 87.9 79.9 70.1 73.7 71.2 70.9 79.5 73.0 77.0 91.3 83.5 91.9 83.2 71.9 76.0 72.1 71.6 Gemini-2.0-flash 95.1 74.9 67.8 71.9 84.2 78.2 85.8 76.9 68.7 72.0 69.8 69.1 78.4 72.0 76.0 88.7 83.6 91.2 81.7 71.0 75.2... | https://arxiv.org/abs/2505.20875v1 |
scores, whereas Australian English (AuE), Southeast English (SE-Eng), and Southwest English (SW-Eng) show relatively stronger performance. For ESL varieties, Arabic (ar) and Turkish (tr) underperform, while French (fr) and Italian (it) achieve higher scores. We attribute these trends partly to data availability. Dialec... | https://arxiv.org/abs/2505.20875v1 |
38 variates, and experimental results with seven state-of-the-art LLMs reveal significant performance degradation on non-standard varieties, underscoring the importance of evaluating linguistic robustness across diverse forms of English. 6 Limitations & Future Work This work focuses primarily on English varieties. Alth... | https://arxiv.org/abs/2505.20875v1 |
Sd-qa: Spoken dialectal question answering for the real world. arXiv preprint arXiv:2109.12072 , 2021. [15] Asma Farajidizaji, Vatsal Raina, and Mark Gales. Is it possible to modify text to a target readability level? an initial investigation using zero-shot large language models. arXiv preprint arXiv:2309.12551 , 2023... | https://arxiv.org/abs/2505.20875v1 |
preprint arXiv:2004.12376 , 2020. [34] Bernd Kortmann, Kerstin Lunkenheimer, and Katharina Ehret, editors. eWAVE . 2020. URL https: //ewave-atlas.org/ . [35] Locky Law. Application of generative artificial intelligence (genai) in language teaching and learning: A scoping literature review. Computers and Education Open ... | https://arxiv.org/abs/2505.20875v1 |
Challenges in the Management of Large Corpora (CMLC-7) 2019. Cardiff, 22nd July 2019, pages 9 – 16, Mannheim, 2019. Leibniz-Institut für Deutsche Sprache. doi: 10.14618/ids-pub-9021. URL http://nbn-resolving.de/ urn:nbn:de:bsz:mh39-90215 . [51] Anne O’Keeffe and Geraldine Mark. The english grammar profile of learner co... | https://arxiv.org/abs/2505.20875v1 |
technologies , pages 180–189, 2011. [69] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830 , 2019. [70] Runtao Zhou, Guangya Wan, Saadia Gabriel, Sheng Li, Alexander J Gates, Maarten Sap, and Thomas Hartvigsen. D... | https://arxiv.org/abs/2505.20875v1 |
. . . . . . 18 C.2 ESL English-L1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 C.2.1 Number of Samples in Compiled Dataset. . . . . . . . . . . . . . . . . . . 19 C.2.2 CEFR Pseudo-Label Generation . . . . . . . . . . . . . . . . . . . . . . . 19 C.2.3 Outputs from the Automatic Grammar Ch... | https://arxiv.org/abs/2505.20875v1 |
The authors of this paper. •Question: Who funded the creation of the dataset? If there is an associated grant, please provide the name of the grantor and the grant name and number. Answer: This work was supported by Institute for Information & communications Technol- ogy Planning & Evaluation(IITP) grant funded by the ... | https://arxiv.org/abs/2505.20875v1 |
apply to a dataset consumer? Please provide descriptions of all external resources and any restrictions associated with them, as well as links or other access points, as appropriate. Answer: Our dataset is self-contained. •Question: Does the dataset contain data that might be considered confidential (e.g., data that is... | https://arxiv.org/abs/2505.20875v1 |
a link or other access point to, or otherwise reproduce, any relevant licensing terms or ToU, as well as any fees associated with these restrictions. Answer: The datasets are distributed under the CC BY-SA 4.0 license. •Question: Have any third parties imposed IP-based or other restrictions on the data associated with ... | https://arxiv.org/abs/2505.20875v1 |
Morphol- ogy, Negation, Agreement, Relativization, Complementation, Adverbial Subordination, Adverbs and Prepositions, and Discourse and Word Order. Each feature is accompanied by illustrative examples. Varieties are annotated with six levels of feature prevalence: (i) feature is pervasive or obligatory, (ii) feature i... | https://arxiv.org/abs/2505.20875v1 |
among three CEFR levels (A, B, C). Respond only CEFR level. Sentence: {sentence} C.2.3 Outputs from the Automatic Grammar Checker The outputs from the automatic grammar checker are overly specific, identifying narrow error types such as “I told her (to) break a leg” or “this render (renders) the . . . ”. To enable more... | https://arxiv.org/abs/2505.20875v1 |
passive voice when active voice is required •French: Non-standard negation with ‘let’s’, Usage of ‘couple times’ instead of ‘a couple of times’, Redundant verb in question form, Misuse of ‘have’ and ‘having’, Usage of a plural noun where a singular is required after ‘is there any’, Use of plural noun with each/every, G... | https://arxiv.org/abs/2505.20875v1 |
metaphors. Application refers to the action items that a model should take in order to reflect the given linguistic feature. All questions and action items should strictly be related to lexicon. All questions and action items should not include context, culture, or metaphor where answers might differ by people such as ... | https://arxiv.org/abs/2505.20875v1 |
Replace ‘I’ with ‘myself’ in the coordinate subject. Feature: Omission of Required Articles Qualification: 1. Does the sentence contain a noun that requires an article (‘a’, ‘an’, or ‘the’) for grammatical correctness or clarity? 2. Is the noun countable and in singular form, or does it refer to something specific that... | https://arxiv.org/abs/2505.20875v1 |
A1, A2, B1, B2, C1, C2. User: {word} Table 11: V ocabulary transformation prompt. System: You are an expert in transforming vocabulary of higher CEFR levels to level {target_level} . You are given higher level words that appear in the question: {words_to_transform} . Please replace at least {min_transform_words} words ... | https://arxiv.org/abs/2505.20875v1 |
essential information is missing. Respond with either ‘yes’ or ‘no’ only. Sentence 1: {SAE written sentence} Sentence 2: {transformed sentence} Answer: a feature-specific guideline and example. The model is instructed to follow the guideline strictly, preserving the structure and core meaning of the original sentence w... | https://arxiv.org/abs/2505.20875v1 |
94.8% 1.92 / 91.8% 2.15 / 94.9% 2.88 / 96.0% 2.98 / 99.7% es 3.15 / 97.1% 3.50 / 99.7% 2.74 / 97.1% 3.53 / 99.3% 3.55 / 98.3% 3.63 / 99.6% fr 1.00 / 84.6% 0.99 / 86.6% 0.83 / 74.8% 1.15 / 87.2% 1.16 / 87.6% 1.11 / 92.8% it 1.03 / 80.8% 1.05 / 87.5% 0.75 / 68.1% 1.20 / 87.3% 1.33 / 89.6% 1.19 / 87.7% ja 3.21 / 96.5% 3.4... | https://arxiv.org/abs/2505.20875v1 |
present tense verb? Yes (become) Is the present tense verb not already conjugated with "-s" in all forms? Yes (become) Application Identify all present tense verbs: "become" Add the "-s" suffix: "becomes" Transformation A company may becomes insolvent if it1 Go to first unanswered sample Is the Transformation consistent... | https://arxiv.org/abs/2505.20875v1 |
arXiv:2505.20880v1 [cs.CL] 27 May 2025MSA at SemEval-2025 Task 3: High Quality Weak Labeling and LLM Ensemble Verification for Multilingual Hallucination Detection Baraa Hikal, Ahmed Nasreldin, Ali Hamdi Faculty of Computer Science, MSA University, Egypt {baraa.moaweya, ahmed.nasreldin, ahamdi}@msa.edu.eg Abstract This... | https://arxiv.org/abs/2505.20880v1 |
scores are unreliable (Kang et al., 2024a). Morphologically rich languages introduce additional challenges due to intricate annotation inconsistencies (Tsarfaty et al., 2013). Prior work on translation-based verification has attempted to address this, but these approaches are ineffective in zero-shot scenarios (Nie, 20... | https://arxiv.org/abs/2505.20880v1 |
is deter- mined by aggregating the probabilities across all verification runs: pi=1 NNX j=1pij where N= 3is the number of adjudicator models per run. The final hallucination label is assigned using a majority voting scheme across all runs. A span is classified as hallucinated if: pi≥0.7 Figure 1: Overview of our halluc... | https://arxiv.org/abs/2505.20880v1 |
adjudicators, we report results for each combination. The tables [1,2,3,4] show performance across languages. Lang IoU Score Probability Corr AR 0.576 0.536 EU 0.604 0.611 DE 0.526 0.567 SV 0.607 0.401 FI 0.587 0.501 CS 0.396 0.410 FA 0.540 0.511 FR 0.571 0.507 EN 0.506 0.538 IT 0.484 0.545 HI 0.684 0.725 Table 1: Perf... | https://arxiv.org/abs/2505.20880v1 |
align with human annota- tions, and refining span localization techniques to enhance character-level precision. These improve- ments could further advance hallucination detec- tion in multilingual NLP systems. References Mostafa Abdelrahman. 2024. Hallucination in low- resource languages: Amplified risks and mitigation... | https://arxiv.org/abs/2505.20880v1 |
2024. Language models in the loop: Incorporating prompting into weak supervision. ACM/JMS Journal of Data Science , 1(2):1–30. Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. 2025. Llm-check: Investigating detection of hallucinations in large language mo... | https://arxiv.org/abs/2505.20880v1 |
arXiv:2505.20888v1 [cs.CL] 27 May 2025EasyDistill : A Comprehensive Toolkit for Effective Knowledge Distillation of Large Language Models Chengyu Wang1, Junbing Yan1, Wenrui Cai1,2, Yuanhao Yue1, Jun Huang1 1Alibaba Cloud Computing2Shanghai Jiao Tong University chengyu.wcy@alibaba-inc.com Abstract In this paper, we pre... | https://arxiv.org/abs/2505.20888v1 |
(i.e., EasyDistill-Recipes ) that serve as practical guides for diverse application needs. iii)EasyDistill is integrated into Alibaba Cloud’s Platform for AI (PAI), showcasing its adaptability and potential for large-scale deployment. By bridg- ing the gap between cutting-edge KD techniques and practical applicability,... | https://arxiv.org/abs/2505.20888v1 |
sets accompanied by high-quality CoTs. 2.1.2 Training Algorithms for KD Scenarios The core KD pipeline for LLMs is straightforward. The input is seed knowledge, consisting of instruc- tions for any target tasks, which is leveraged to prompt the selected teacher LLM to generate de- tailed outputs. In our framework, we s... | https://arxiv.org/abs/2505.20888v1 |
of distilling knowledge from teacher models to de- velop more robust student models, as demonstrated in previous works (Bai et al., 2022; Trung et al., 2024; Yang et al., 2024d). Preference Rank Optimization. A potential draw- back of RL-based algorithms is the instability in training. Preference rank optimization-base... | https://arxiv.org/abs/2505.20888v1 |
true, "enforce_eager": false, "max_model_len": 4096, "max_new_tokens": 512 }, "models": { "teacher": "teacher/Qwen/Qwen2.5-32B-Instruct/", "student": "student/Qwen/Qwen2.5-0.5B-Instruct/" }, "training": { ... } } Code 2: Sample JSON configuration for black-box KD (offline inference with an open-source teacher model). {... | https://arxiv.org/abs/2505.20888v1 |
With the release of large System 2 models such as DeepSeek-R1 (DeepSeek-AI, 2025), the concept of “LLM with slow thinking” has become a stan- dard strategy to extend the intelligent boundaries of LLMs. We introduce the DistilQwen2.5-R1 model series, which utilizes DeepSeek-R1 as the teacher model, based on fine-tuning ... | https://arxiv.org/abs/2505.20888v1 |
validated by DeepSeek-R1 and QwQ-32B. Each CoT process is annotated with novel Reason- ing Verbosity (RV) and Cognitive Difficulty (CD) scores, which describe the appropriateness of CoT verbosity and cognitive difficulty level for models to comprehend these reasoning processes. For de- tails, please refer to (Cai et al... | https://arxiv.org/abs/2505.20888v1 |
into Alibaba Cloud’s Platform for AI (PAI) and its practical solutions further enhance the toolkit’s impact by demonstrating its viability for large-scale deploy- ment. Moreover, the open source of EasyDistill encourages community involvement, which could lead to new enhancements in KD techniques. However, the deployme... | https://arxiv.org/abs/2505.20888v1 |
Annual Conference on Neural Information Processing Sys- tems 2023 . Jeff Rasley, Samyam Rajbhandari, Olatunji Ruwase, and Yuxiong He. 2020. Deepspeed: System optimiza- tions enable training deep learning models with over 100 billion parameters. In The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Min- ing,... | https://arxiv.org/abs/2505.20888v1 |
Dayi- heng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Ya... | https://arxiv.org/abs/2505.20888v1 |
models with significantly shorter CoTs. Meanwhile, the DistilQwen-ThoughtX mod- els are adaptive-thinking models with more optimal CoT lengths. Model AlpacaEval 2.0 MT-Bench MT-Bench IFEval IFEval (length control) (single) (instruct-loose) (strict-prompt) Qwen2-1.5B-Instruct 5.22 5.85 6.45 41.37 28.10 DistilQwen2-1.5B-... | https://arxiv.org/abs/2505.20888v1 |
arXiv:2505.20897v1 [cs.CV] 27 May 2025Cross from Left to Right Brain: Adaptive Text Dreamer for Vision-and-Language Navigation Pingrui Zhang1,2, Yifei Su3,4, Pengyuan Wu2, Dong An5, Li Zhang6, Zhigang Wang2, Dong Wang2,Yan Ding2,Bin Zhao2,Xuelong Li7† 1Fudan University2Shanghai AI Laboratory 3School of Artificial Intel... | https://arxiv.org/abs/2505.20897v1 |
stairs and walk towards the red sofa. Turn left and enter the bathroom”, and assuming 3candidate viewpoints are encountered during the current navigation step. Left: Existing methods generate the visual imagination for each candidate. While effective, the rendered images often contain blurry from GT or redundant region... | https://arxiv.org/abs/2505.20897v1 |
( SR) of 8.0% and 12.0%, and in Success weighted by Path Length ( SPL) of 5.0% and 11.0%, on the val seen andval unseen , respectively. These results highlight the efficiency and effectiveness of our proposed Adaptive Text Dreamer guided navigation. Our contributions are summarized as follows: 2 (1) We highlight the ad... | https://arxiv.org/abs/2505.20897v1 |
poses. VLN with LLM. LM-Nav [ 76] uses GPT-3 [ 60] to extract landmarks from instructions to assist navigation. MiCs generate detailed plans from static and dynamic perspectives as knowledge. LLM- Planner [ 79] and SayNav [ 73] utilize LLMs [ 1,60] as core planners to dynamically plan high-level actions. As for VLN, Na... | https://arxiv.org/abs/2505.20897v1 |
as the left brain but is responsible for generating imagination of future candidates (Sec.3.1). To ensure the imagination within the current context, we constrain it via latent embeddings derived from the state estimation. The attended imagination latent is then injected into the graph-based navigation policy to guide ... | https://arxiv.org/abs/2505.20897v1 |
model learns to predict the current state based on the observations and instructions, i.e., the left brain. Specifically, the process of training the State Estimation LLM, including its inputs and outputs, can be formulated as follows: Q′ lb= Q-formerlb(W,Ot,Qlb), (1) ⟨ˆRt, State E t⟩= LLM frozen (StateEsimationPrompt(... | https://arxiv.org/abs/2505.20897v1 |
the information in the imagination. Thus, when the state changes, Aadjusts accordingly to filter the imagination information. This adaptive refinement process, known as SGCA, can be mathematically expressed as: SGCA( QS,KI,VI) =A·VI. (8) The grounded feature is then fed into the node embedding of the subsequent navigat... | https://arxiv.org/abs/2505.20897v1 |
scale.†: Indicates methods that leverage additional visual data beyond MP3D. Methods Freeze LLMVal Seen Val Unseen Test Unseen TL NE ↓OSR↑SR↑SPL↑ TL NE ↓OSR↑SR↑SPL↑ TL NE ↓OSR↑SR↑SPL↑ Human – – – – – – – – – – – 11.85 1.61 90 86 76 Seq2Seq [4] – 11.33 6.01 53 39 – 8.39 7.81 28 21 – 8.13 7.85 27 20 – RCM [90] – 10.65 3.... | https://arxiv.org/abs/2505.20897v1 |
and Evaluation Metrics We conduct systematic evaluations of the proposed model on the widely used R2R [ 4] benchmark in discrete environments. The R2R dataset provides step-by-step navigation instructions, with each instruction averaging 32words and covering approximately 6navigation steps. We evaluate the performance ... | https://arxiv.org/abs/2505.20897v1 |
to other LLM-leveraged methods, our approach achieves the best performance with the fewest parameters—1.5B. NaviLLM [ 108] is the first fully fine-tuned 7B LLM used as an action generator, and ATD outperforms it by 6%SR and 3%SPL on the test split. This indicates that maintaining the text generation capability of the L... | https://arxiv.org/abs/2505.20897v1 |
63 52 ✓ 13.23 3.05 80 72 61 ✓ 13.87 2.97 80 73 60 ✓ ✓ 13.33 2.81 83 75 63Table 5: Ablation of SGCA layer numbers . MetricsNumber of Layers 1 2 3 4Val SeenNE↓ 3.05 2.63 2.88 2.67 OSR ↑78.35 82.47 80.8 81.78 SR↑71.69 75.91 76.3 75.61 SPL ↑63.15 67.46 67.13 67.49Val UnseenNE↓ 2.82 2.9 2.88 2.81 OSR ↑82.08 82.12 82.12 82.7... | https://arxiv.org/abs/2505.20897v1 |
of layers is set to 1, the model performs much better on val unseen than on val seen . This suggests that ATD has a relatively high lower bound for generalization. 9 5 Conclusion In this work, we have presented the Adaptive Text Dreamer (ATD), a novel dual-branch vision- language navigation framework that leverages lar... | https://arxiv.org/abs/2505.20897v1 |
passing it to the LLM. The system prompts for the State Estimation LLM and the Imagination LLM are illustrated in the Fig. 7. 10 State Estimation Intruction Collection Prompt {image} You are now an agent navigating within an indoor environment. Your current task is as follows: {instruction}. Based on your observations ... | https://arxiv.org/abs/2505.20897v1 |
the best possible success if the agent stopped at the closest point to the goal. Soracle=1if∃vi∈G 0otherwise,(16) or, if no goal set: Soracle=1ifminid(vi, v∗)< ϵ 0otherwise,(17) where viis the i-th viewpoint along the path. •SPL. A metric combining success and efficiency, penalizing longer trajectories relative to th... | https://arxiv.org/abs/2505.20897v1 |
pairs) via automatic strategies; 2) proposing a shuffling loss to enhance learning of temporal order in PI pairs. •VLNBert [32]. This paper proposes a recurrent Vision-Language BERT model, equipping BERT with a recurrent function to maintain cross-modal state information and address the history- dependent decision-maki... | https://arxiv.org/abs/2505.20897v1 |
as a world model to imagine the next observation based on instructions, selects the most aligned candidate observation, and determines actions through disentangled reasoning. Formalized training labels are constructed to guide the LLM in generating reasonable chain-of-thought outputs for improved action decisions. 13 •... | https://arxiv.org/abs/2505.20897v1 |
Work. Currently, the data collected for training the imagination LLM is limited to candidate nodes one step ahead of the current node. This may restrict the model’s ability to perform long-horizon imagination. Future work could explore incorporating long-horizon imagination capabilities to more fully leverage the poten... | https://arxiv.org/abs/2505.20897v1 |
A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669 (2025) 2 [10] Buoso, D., Robinson, L., Averta, G., Torr, P., Franzmeyer, T., De Martini, D.: Select2plan: Training-free icl-based planning through vqa and memory retrieval. arXiv preprint arXiv:2411.04006 ... | https://arxiv.org/abs/2505.20897v1 |
R., Wang, W., Yang, Y .: Navigation instruction generation with bev perception and large language models. In: European Conference on Computer Vision. pp. 368–387. Springer (2024) 3 [26] Fan, Y ., Chen, W., Jiang, T., Zhou, C., Zhang, Y ., Wang, X.E.: Aerial vision-and-dialog navigation. arXiv preprint arXiv:2205.12219 ... | https://arxiv.org/abs/2505.20897v1 |
field rendering. ACM Trans. Graph. 42(4), 139–1 (2023) 3 [40] Kocsis, L., Szepesvári, C.: Bandit based monte-carlo planning. In: European conference on machine learning. pp. 282–293. Springer (2006) 3 [41] Koh, J.Y ., Lee, H., Yang, Y ., Baldridge, J., Anderson, P.: Pathdreamer: A world model for indoor navigation. In:... | https://arxiv.org/abs/2505.20897v1 |
on Computer Vision and Pattern Recognition (CVPR) pp. 16317–16328 (2024) 3 [56] Long, Y ., Cai, W., Wang, H., Zhan, G., Dong, H.: Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. arXiv preprint arXiv:2406.04882 (2024) 3 [57] Long, Y ., Li, X., Cai, W., Dong, H.: Discuss before... | https://arxiv.org/abs/2505.20897v1 |
IEEE Transactions on Pattern Analysis and Machine Intelligence 45(7), 8524–8537 (2023) 3, 7, 13 [72] Qiao, Y ., Qi, Y ., Yu, Z., Liu, J., Wu, Q.: March in chat: Interactive prompting for remote embodied referring expression. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15758–15767 (2... | https://arxiv.org/abs/2505.20897v1 |
text-guided panoramic environment generation for vision-and-language navigation. Neural Networks 187, 107320 (2025) 3 [87] Wang, S., Chen, L., Chen, P., Dong, J., Xue, B., Jiang, J., Kong, L., Wu, C.: Mos: Unleashing parameter efficiency of low-rank adaptation with mixture of shards. arXiv preprint arXiv:2410.00938 (20... | https://arxiv.org/abs/2505.20897v1 |
3 [101] Zhang, L., Hao, X., Xu, Q., Zhang, Q., Zhang, X., Wang, P., Zhang, J., Wang, Z., Zhang, S., Xu, R.: Mapnav: A novel memory representation via annotated semantic maps for vlm-based vision-and-language navigation. arXiv preprint arXiv:2502.13451 (2025) 3 [102] Zhang, P., Gao, X., Wu, Y ., Liu, K., Wang, D., Wang,... | https://arxiv.org/abs/2505.20897v1 |
arXiv:2505.20899v1 [cs.CL] 27 May 2025Dub-S2ST: Textless Speech-to-Speech Translation for Seamless Dubbing Jeongsoo Choi∗Jaehun Kim∗Joon Son Chung Korea Advanced Institute of Science and Technology {jeongsoo.choi, kjaehun, joonson}@kaist.ac.kr Abstract This paper introduces a cross-lingual dubbing system that translate... | https://arxiv.org/abs/2505.20899v1 |
quality and lead to unnatural prosody and speaking pace. A key challenge underlying this issue lies in the limitations of existing training datasets. High- quality dubbing demands not only accurate trans- lation but also faithful preservation of the source speech’s voice characteristics and speaking speed. However, it ... | https://arxiv.org/abs/2505.20899v1 |
advancements, relying on text as intermediate rep- resentation inherently limits temporal flexibility, highlighting the need for textless approaches that better preserve the naturalness of the source speech. 2.2 Speech-to-Speech Translation (S2ST) Speech-to-Speech Translation (S2ST) aims to con- vert source speech into... | https://arxiv.org/abs/2505.20899v1 |
training objective, and fi- nal synthesis of the translated speech. The overall architecture of our model is illustrated in Fig 1. 3.1 Data Preparation Speech Unit Extraction. The choice of target speech representation plays a critical role in deter- mining the quality and accuracy of S2ST. While continuous features al... | https://arxiv.org/abs/2505.20899v1 |
that operate on variable- length sequences, with self-attention replaced by cross-attention to integrate source speech features throughout the generation. During inference, the decoder is initialized with all masked units with length identical to that of the source speech, and iteratively transforms it into speech unit... | https://arxiv.org/abs/2505.20899v1 |
training the module on S2ST data is possible, such datasets often contain noise and reverberation, which can impair synthesis qual- ity. To address this, we initialize the model with weights trained on multilingual corpus, and fine- tune it with our own semantic units. This approach allows for robust zero-shot synthesi... | https://arxiv.org/abs/2505.20899v1 |
embedding vectors and calculated cosine similarity between the two. https://github.com/facebookresearch/fairseq/ blob/main/examples/speech_to_speech/docs/ textless_s2st_real_data.md https://github.com/facebookresearch/fairseq/ tree/ust/examples/speech_to_speech/asr_bleu https://huggingface.co/facebook/blaser-2. 0-qe 5 ... | https://arxiv.org/abs/2505.20899v1 |
speaker similarity. Fur- thermore, Dub-S2ST achieves highly competitive ASR-BLEU and BLASER 2.0 scores compared to S2UT and CTC-S2UT, which lack duration control. This indicates that our model effectively captures the semantic information from the source speech and transfers it to the translated speech, while main- tai... | https://arxiv.org/abs/2505.20899v1 |
Chars 0.851 0.929 1.000 1.064 1.133 Table 5: Effect of explicit duration control. repetitive words. Notably, significant differences in speech naturalness emerged during the evalua- tion. Baseline samples required manual waveform adjustments, leading to considerable degradation in perceptual quality. Conversely, our pr... | https://arxiv.org/abs/2505.20899v1 |
on the latency-performance trade-off, we opt to choose the NFE of 64 for our evaluation. Loss computation. In Table 7, we examine the im- pact of different loss computation strategies on our model’s performance. Our findings indicate that computing loss on all units (i.e., predicting both masked and non-masked units) r... | https://arxiv.org/abs/2505.20899v1 |
recognition in english and mandarin. InProc. ICML . Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. 2021. Structured denoising diffusion models in discrete state-spaces. InProc. NeurIPS . Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framewo... | https://arxiv.org/abs/2505.20899v1 |
by masked prediction of hidden units. IEEE/ACM Trans. on Audio, Speech, and Language Processing . 9 Rongjie Huang, Jinglin Liu, Huadai Liu, Yi Ren, Lichao Zhang, Jinzheng He, and Zhou Zhao. 2023. Transpeech: Speech-to-speech translation with bilat- eral perturbation. In Proc. ICLR . Ye Jia, Michelle Tadmor Ramanovich, ... | https://arxiv.org/abs/2505.20899v1 |
quence modeling. arXiv preprint arXiv:1904.01038 . William Peebles and Saining Xie. 2023. Scalable diffu- sion models with transformers. In Proc. CVPR . Matt Post. 2018. A call for clarity in reporting bleu scores. In Proceedings of the Third Conference on Machine Translation: Research Papers . Alec Radford, Jong Wook ... | https://arxiv.org/abs/2505.20899v1 |
arXiv:2505.20901v1 [cs.CL] 27 May 2025A Stereotype Content Analysis on Color-related Social Bias in Large Vision Language Models Junhyuk Choi†, Minju Kim†, Yeseon Hong and Bugeun Kim Department of Artificial Intelligence, Chung-Ang University Seoul, Republic of Korea {chlwnsgur129, minjunim, ghddptjs, bgnkim}@cau.ac.kr... | https://arxiv.org/abs/2505.20901v1 |
adopt Stereotype Content Model (SCM), a stereotype evaluation framework used in human study (Fiske et al., 2018). Second, regarding color, LVLMs may exhibit stereotypes because of different colors in the image. Previously, researchers in vision-language model reported that model responds inconsistently due to image dif... | https://arxiv.org/abs/2505.20901v1 |
evaluation scores across these thematic sets (Mandal et al., 2023; Janghorbani and De Melo, 2023; Qiu et al., 2023). However, such methods often suffer from insufficient con- trol over confounding variables unrelated to the targeted stereotype. For example, if gender, race, and background all change simultaneously, any... | https://arxiv.org/abs/2505.20901v1 |
a broader spectrum of gender and race beyond those considered in this study. This selec- tion was made solely for experimental purposes and is not intended to exclude any group. Step 1: Constructing Occupation List. To create paired images, we started by choosing a common occupation for the individuals depicted in each... | https://arxiv.org/abs/2505.20901v1 |
images, we explicitly add phrases such as “white clothes, red clothes, any color except blue” to the negative prompts used in Step 4. Through this step, we obtain a total of 180 images per scenario ( = 5×2×6×3). Step 6: Filtering. To ensure the quality of the fi- nal dataset, we conducted a filtering process based on t... | https://arxiv.org/abs/2505.20901v1 |
conduct: generated answer should be one of the choices. 4.2 Metrics 4.2.1 Existing sentiment-based metric Following previous work (Wang et al., 2024a; In- oshita and Zhou, 2024; Hu et al., 2025), we used V ADER (Hutto and Gilbert, 2014) to assess the sentiment of a description. Among the various sentiment-based metrics... | https://arxiv.org/abs/2505.20901v1 |
coordinates as follows: αw(x) =dw(x)−d·dc(x) 1−d2, (5) αc(x) =dc(x)−d·dw(x) 1−d2. (6) 4.3 Analyses We use two analysis methods to examine stereo- types in LVLMs: statistical test and pointwise mu- tual information (PMI). First, for statistical tests, we conducted paired t-test. To separately assess the effect of one st... | https://arxiv.org/abs/2505.20901v1 |
as 7B or 12B. Implementation details (e.g., temperature settings) are described in the Appendix B. 5 Results To analyze whether three image attributes cause dif- ferent stereotype levels, we conducted paired t-tests using three evaluation metrics. Table 3 shows the results of statistical tests, regarding color and gen-... | https://arxiv.org/abs/2505.20901v1 |
3.38***2.73**0.70 -2.19* Red vs.White -1.61 -1.86 -0.89 -1.59 2.18*0.65 -2.31*-6.07*** Gender Male vs.Female -6.41***-3.66***-8.63***-8.29***-15.26***-7.89***-2.35*-3.30*** # of significant stereotypes 6/19 8/19 11/19 8/19 14/19 13/19 8/19 9/19 Compe- Color Blue vs.Red -11.18***-9.08***-3.67***-7.95***-15.10***-10.39**... | https://arxiv.org/abs/2505.20901v1 |
affected by such safeguards. 6.2 Existence of Color Stereotypes Color plays a significant role in shaping stereo- types, as it demonstrates clear patterns in the di- mensions of warmth and competence. We believe that these stereotypes develop because the model internalizes sociocultural associations present in image-te... | https://arxiv.org/abs/2505.20901v1 |
no model exhibited dominant stereotypes compared to others in all metrics. So we conclude the effect of parameter sizes and architecture are not easily distinguishable from each other. 7 Conclusion In this paper, we introduced a comprehensive eval- uation framework that adopts Stereotype Contents Model (SCM) and moves ... | https://arxiv.org/abs/2505.20901v1 |
or culturally nuanced biases exist, LVLMs may provide differ- ent responses when we use different languages. This implies that the result might vary across differ- ent languages or culture. Thus, conducting similar experiment with different languages may provide different insights. References Hajo Adam and Adam D Galin... | https://arxiv.org/abs/2505.20901v1 |
Computational Linguistics. Kathleen C. Fraser, Isar Nejadgholi, and Svetlana Kiritchenko. 2021. Understanding and countering stereotypes: A computational approach to the stereo- type content model. In Proceedings of the 59th An- nual Meeting of the Association for Computational Linguistics and the 11th International Jo... | https://arxiv.org/abs/2505.20901v1 |
Language Processing , pages 12814– 12845, Miami, Florida, USA. Association for Com- putational Linguistics. Jayendra Kantipudi, Shiv Ram Dubey, and Soumendu Chakraborty. 2020. Color channel perturbation at- tacks for fooling convolutional neural networks and a defense against such attacks. IEEE Transactions on Artifici... | https://arxiv.org/abs/2505.20901v1 |
pages 4123–4139, Toronto, Canada. Association for Computational Linguistics.Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 . Haoyi Qiu... | https://arxiv.org/abs/2505.20901v1 |
to cover a diverse range of domains in- cluding medicine, engineering, arts, public service, and manual labor, enabling robust analysis across various social roles. A.2 Step 2: Creating action We used the following prompts for generating ac- tions for each occupation. System prompt: You are an NLP assistant whose purpo... | https://arxiv.org/abs/2505.20901v1 |
4, with different hyper- parameters: we set strength as 0.8 and guidance as 11. Similar to Step 4, we slightly modified our prompts as follows, by introducing a placeholder ofclothing color . {race} {gender} , {profession prompt} , wearing {clothing color} , gray background, professional photography, high detail, high ... | https://arxiv.org/abs/2505.20901v1 |
for Qwen 7B) and Middle Eastern (row 19; p <0.001for all). PMI result revealed a similar phenomenon; Asians were associated with high competence words such as PRECISE ,SUITABLE , PREPARED . Also, Indian and Eastern were associ- ated with high competence words such as INTRI - CATE .C.1 Warmth Warmth: Regarding Race, we ... | https://arxiv.org/abs/2505.20901v1 |
vs.Indian 0.72 -3.31***4.35***3.77***7.32***4.91***1.11 3.87*** Black vs.Latino 1.86 1.75 4.24***1.97*0.77 2.77**0.19 0.37 Black vs.M.E. 0.82 -2.17*1.60 0.73 -1.37 -0.57 -3.23**-4.52*** Black vs.White 2.59**4.60***6.07***3.73***3.09**2.94**1.75 3.19** Indian vs.Latino 1.11 4.85***-0.15 -1.74 -6.49***9.23***-0.91 -3.49*... | https://arxiv.org/abs/2505.20901v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.