text string | source string |
|---|---|
of speaker, hours and sentences represent their combined sum over the lan- guages supported. on automatic speech recognition to illustrate its potential for multilingual speech processing. The NaijaV oices dataset is licensed under the CC BY-NC-SA 4.0 license and is accessible at https:// naijavoices.com/ . 2. How Naij... | https://arxiv.org/abs/2505.20564v1 |
African speech dataset creation efforts [ 7,17,25,26, 27] produce small outputs (tens of hours) due to conditions that, while offering much-needed supervision, limit scala- bility and diversity (such as requiring recorders to travel to a single recording booth for limited sessions one by one [25]). On the other hand, e... | https://arxiv.org/abs/2505.20564v1 |
from approximately 105 voice donors. 3. The NaijaVoices Dataset The NaijaV oices dataset captures the essence of Nigerian culture through authentic, expert-generated, and culturally rich sentences, offering a level of originality rarely found in online texts [ 21]. It features a wide range of speech patterns influenced... | https://arxiv.org/abs/2505.20564v1 |
the majority of audio samples lie within 100. According to [ 29], a value of 100, the highest possible value, means that the audio has a great proportion of silence w.r.t speech. 4. Automatic Speech Recognition We perform automatic speech recognition (ASR) experi- ments to demonstrate the potential of the NaijaV oices ... | https://arxiv.org/abs/2505.20564v1 |
and Hausa, while for Igbo, the English tokenizer was employed, as the Whisper tokenizer was not originally trained for Igbo. The XLSR and MMS model were fine- tuned under similar hardware conditions for 5 epochs, with a gradient accumulation step of 16. Training and validation batch sizes were set to 16, and, where nee... | https://arxiv.org/abs/2505.20564v1 |
as our unique record- ing framework, to speech data cultivation for more African languages. 6. Acknowledgement We express our profound gratitude to the entire NaijaV oices community for making this dataset possible, and the Lacuna Fund for funding the creation of the NaijaV oices dataset. Finally we acknowledge the sup... | https://arxiv.org/abs/2505.20564v1 |
[15] K. Olaleye, D. Onea t ¸˘a, and H. Kamper, “YFACC: A Yor `ub´a speech–image dataset for cross-lingual keyword localisation through visual grounding,” in SLT 2022 . IEEE, 2023, pp. 731–738. [16] T.`Og´unr`em´ı, K. T ´ubos ´un, A. Anuoluwapo, I. Orife, and D. I. Adelani, “ `Ir`oy`ınspeech: A multi-purpose Yor `ub´a s... | https://arxiv.org/abs/2505.20564v1 |
[33] Seamless Communication, L. Barrault, Y .-A. Chung, M. C. Meglioli, D. Dale, N. Dong et al. , “Seamless: Multilingual expressive and streaming speech translation,” ArXiv , vol. abs/2312.05187, 2023. [34] K. Ogueji, Y . Zhu, and J. Lin, “Small Data? No Problem! Ex- ploring the viability of pretrained multilingual la... | https://arxiv.org/abs/2505.20564v1 |
arXiv:2505.20571v1 [cs.CL] 26 May 2025Emotion Classification In-Context in Spanish Bipul Thapa1and Gabriel Cofre2 1Department of Computer and Information Sciences, University of Delaware, Newark, DE, USA bipul@udel.edu 2Department of Physics and Astronomy, University of Delaware, Newark, DE, USA gcofre@udel.edu Abstrac... | https://arxiv.org/abs/2505.20571v1 |
as shown in Table 1. Therefore, utilizing the original Spanish dataset is necessary to achieve higher accuracy in emotion classification for the Spanish language. Furthermore, the selection of appropri- ate models and feature representations is also crucial to ensuring reliable and accurate emotion classification. Tabl... | https://arxiv.org/abs/2505.20571v1 |
the practicality and performance efficiency of the CSE approach. The rest of the paper is organized as follows. In Section 2, we provide a review of related work in this field. Section 3 details the methodology and pro- posed model, including the machine learning approaches and feature represen- tation techniques used ... | https://arxiv.org/abs/2505.20571v1 |
Test-Train Split Feature Extraction TF-IDF, TF-IDF with BERT Classification Process Logistic Regression, KNN, Bagging Classifier, AdaBoost, Build Stacking Classifier Predicting Emotion as 'Positive', 'Neutral' or 'Negative' Fig.1: Overview of the method 3.1 Data Preparation The initial step involves preparing the data ... | https://arxiv.org/abs/2505.20571v1 |
the data. By considering local information in the feature space, KNN offers a different per- spective compared to the gradient boosting and logistic regression models, which is particularly useful for classifying complex patterns in the Spanish dataset. AdaBoost is selected for its capability to improve accuracy by foc... | https://arxiv.org/abs/2505.20571v1 |
and F1 score. Accuracy measures the proportion of correctly classified instances. Precision assesses the accuracy of positive predictions, while recall measures the ability to identify all relevant instances. The F1 score, the harmonic mean of precision and recall, provides a balanced evaluation metric. By comparing th... | https://arxiv.org/abs/2505.20571v1 |
performance by sequentiallyimprovingtheweaklearners.Forthistask,decisiontreesareselected as the base learners, with the number of boosting stages set based on value from grid search (n_estimators). AdaBoost adapts to the misclassified instances by 3Available: https://huggingface.co/Helsinki-NLP/opus-mt-es-en 10 B. Thap... | https://arxiv.org/abs/2505.20571v1 |
indicating the models are stable across validation strategies. Table 5: Cross-Validation and Feature Generation Performance with Spanish Dataset Cross-Validation Accuracy Precision Recall F1-Score Stratified K-fold 0.89 0.89 0.89 0.89 K-fold 0.88 0.89 0.88 0.88 The Stratified K-fold technique yielded an accuracy of 0.8... | https://arxiv.org/abs/2505.20571v1 |
the Spanish dataset generally yielding higher accuracies than the English-translated one. This highlights that the models perform more accurately on the native Spanish dataset than the translation. Fig.3: Model Accuracy Comparison for Spanish and English-translated Datasets The performance results of the state-of-the-a... | https://arxiv.org/abs/2505.20571v1 |
M. & Perea-Ortega, J.:SemanticorientationforpolarityclassificationinSpanishreviews. Expert Systems With Applications .40, 7250-7257 (2013) 5. Bhowmik, S., Prosun, P. & Alam, K.: A novel three-level voting model for detecting misleading information on COVID-19. Proc. Of The Advanced Techniques For IoT Applications: Proc... | https://arxiv.org/abs/2505.20571v1 |
arXiv:2505.20591v1 [cs.CL] 26 May 2025Effectiveness of Prompt Optimization in NL2SQL Systems Sairam Gurajada∗ sairam@megagon.ai Megagon Labs Mountain View, California, USAEser Kandogan eser@megagon.ai Megagon Labs Mountain View, California, USASajjadur Rahman∗ sajjadurr@adobe.com Adobe San Jose, California, USA Abstrac... | https://arxiv.org/abs/2505.20591v1 |
making it a challenging task despite recent breakthroughs [6]. Recent works [ 4,21] emphasize that exemplar selection is crucial for building effective NL2SQL systems. Retrieval-based exemplar selection—i.e., identifying exemplars similar to the user query—has become the de facto method. However, studies [ 4,19] highli... | https://arxiv.org/abs/2505.20591v1 |
retrieval indexes and the complexity of online synthetic generation. Prompt Optimization. Optimizing LLM prompts has been a focus for several years [ 22,26], showing effectiveness across a multitude of applications. More recently, DSPy [ 9] introduced a declarative framework for expressing and optimizing prompts for NL... | https://arxiv.org/abs/2505.20591v1 |
selection is choosing an appropriate value for 𝑘. A small𝑘 may fail to capture the diversity of the NL and SQL constructs, while a large𝑘can lead to lost-in-the-middle issues with LLMs [14] and increase generation costs due to the larger prompt size. A simple yet effective approach is to treat 𝑘as a hyperparameter ... | https://arxiv.org/abs/2505.20591v1 |
selection. Additionally, we observed that IPO often generates more concise NL2SQL prompts by pruning irrelevant schema information from the exemplars. For example, Figure 3 shows an exemplar whose schema includes only the table film and the columns film_id ,title , and rating from the database movie_3 . Although schema... | https://arxiv.org/abs/2505.20591v1 |
optimization techniques described in Section 3, it is possible to jointly optimize for both SQL efficiency and generation accuracy, leading to more practical and performant NL2SQL systems. 5 Preliminary Results Here, we present preliminary results demonstrating the effective- ness of prompt optimization in NL2SQL syste... | https://arxiv.org/abs/2505.20591v1 |
13m16s IPO 6,495 8m53s Table 2: Quantitative analysis of PO on BIRD (dev) dataset 5.2 Multi-Objective Optimization Table 3 demonstrates the effectiveness of multi-objective optimiza- tion using the IPO approach. For this, we consider both accuracy and latency on the BIRD (dev) dataset. When compared to the ground truth... | https://arxiv.org/abs/2505.20591v1 |
Alex Van Grootel, Brandon Chow, Kai Deng, Katherine Lin, Marcos Campos, K. Venkatesh Emani, Vivek Pandit, Victor Shnayder, Wenjing Wang, and Carlo Curino. 2024. NL2SQL is a solved problem... Not!. In CIDR . https://www.cidrdb.org/cidr2024/ papers/p74-floratou.pdf [7]Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen... | https://arxiv.org/abs/2505.20591v1 |
Gan, Amin Saberi, Fatma Ozcan, and Sercan O Arik. 2025. CHASE-SQL: Multi-Path Reasoning and Preference Optimized Candidate Selection in Text-to-SQL. In The Thirteenth International Conference on Learning Representations . https://openreview.net/forum?id=CvGqMD5OtX [20] Mohammadreza Pourreza and Davood Rafiei. 2023. DIN... | https://arxiv.org/abs/2505.20591v1 |
of 3the database, and relevant evidence, generate a 4valid SQLite SQL query that satisfies the NLQ. Use 5the provided schema and evidence to ensure the SQL 6query correctly answers the NLQ. Only utilize 7relevant columns and tables in the query. 8Return only the SQL query without any prefixes 9or block quotes. 10 11 #E... | https://arxiv.org/abs/2505.20591v1 |
arXiv:2505.20606v1 [cs.CL] 27 May 2025Towards Pretraining Robust ASR Foundation Model with Acoustic-Aware Data Augmentation Dancheng Liu1, Amir Nassereldine1, Chenhui Xu1, Jinjun Xiong1 1University at Buffalo, USA {dliu37,amirnass,cxu26,jinjun }@buffalo.edu Abstract Whisper’s robust performance in automatic speech reco... | https://arxiv.org/abs/2505.20606v1 |
beyond its synthetic data dis- tribution. Second, some of the de facto speech augmentation strategies, such as SpecAugment [5], which are used by foun- dation models including Whisper-v2 [1], contribute primarily to linguistic diversity rather than acoustic diversity. As a result, it is less effective than techniques t... | https://arxiv.org/abs/2505.20606v1 |
not to study the best specific data augmentation method, we omit the discussion of those works in later sections. 2.2. Usage of Synthetic Data in ASR Training While augmentation methods focus on manipulating existing real-world recordings, generating entirely synthetic speech data has also gained substantial traction i... | https://arxiv.org/abs/2505.20606v1 |
converts the text corpus into the desired audio and text tuples, and such synthetically generated data will be used in the pre-training stage of the ASR model. Additionally, acoustic features can be augmented by ma- nipulating the waveforms or their respective spectrograms to broaden the acoustic diversity of the train... | https://arxiv.org/abs/2505.20606v1 |
size changed to 64) whenever experiments are conducted. All experiments are conducted on a server with 4 A6000 GPUs. We train our model until convergence, which surprisingly takes less than 2% of the training corpus. We then evaluate the pre-trained ASR model on both synthetic data and real-world data. Our experiments ... | https://arxiv.org/abs/2505.20606v1 |
to other techniques, and we hope that our preliminary results can shed light on the potential future directions of foundation ASR model pre-training. On the prosody side, pitch and amplitude are two straight- forward augmentations. The training set, Librispeech-960h [7], only contains adult speech. Thus, using Liborsa ... | https://arxiv.org/abs/2505.20606v1 |
test dataset and three out-of-distribution datasets: L2-Arctic [30] signify- ing accented speech, My Science Tutor (MyST) [29] signifying school-aged children speech, and ENNI [28], a semi-proprietarydataset containing young children’s speech that did not appear in OpenAI Whisper’s pre-training. As shown in Table 1, th... | https://arxiv.org/abs/2505.20606v1 |
Y . Meng, R. Maas, and J. Droppo, “Synthasr: Unlocking synthetic data for speech recognition,” in Interspeech 2021 , 2021, pp. 896–900. [4] B. Hilmes, N. Rossenbach, and R. Schl ¨uter, “On the effect of purely synthetic training data for different automatic speech recognition architectures,” in Synthetic Data’s Transfo... | https://arxiv.org/abs/2505.20606v1 |
spectrogram predictions,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP) , 2018, pp. 4779–4783. [18] C. Donahue, J. McAuley, and M. Puckette, “Adversarial audio synthesis,” in ICLR , 2019. [19] Hexgrad, “Kokoro-82m (revision d8b4fc7),” 2025. [Online]. Available: https://hugg... | https://arxiv.org/abs/2505.20606v1 |
Compari sons between a Large Language Model -based Real -Time Compound Diagnostic Medical AI Interface and Physician s for Common Internal Medicine Cases using Simulated Patient s Hyungjun Park, M.D., Ph.D.1,2*, Chang -Yun Woo, M.D.3*, Seungjo Lim2, Seunghwan Lim 2, Keunho Kwak2, Ju Young Jeong4, Chong Hyun Suh, M.D., ... | https://arxiv.org/abs/2505.20609v1 |
98.1% compared to the physicians' average ( $4.2). Patient satisfaction scores ranged from 4.2 to 4.3 for care by physicians and were 3.9 for the AI interface Conclusion An LLM based realtime compound diagnostic medical AI interface demonstrated diagnostic accuracy and patient satisfaction comparable to those of a phys... | https://arxiv.org/abs/2505.20609v1 |
beforehand. Each simulated patient engaged in Q&A -style chats regarding the clinical vignettes with three physicians and a real-time compound diagnostic medical AI interface (Figure 1A) . The physicians, who were blinded to the clinical vignettes, interacted with all five simulated patients, experiencing 10 clinical v... | https://arxiv.org/abs/2505.20609v1 |
dyspnea, headache, or dizziness), the patient receives a survey comprising 10 –15 multiple -choice questions. Should the logic connecting to a symptom - specific survey seem inappropriate to the user or if the patient’s symptom is not i ncluded among the predefined questionnaires, the consultation shifts to the Supplem... | https://arxiv.org/abs/2505.20609v1 |
. Results Three physicians and a real-time compound diagnostic medical AI interface completed 10 clinical vignettes . Each clinical vignette was addressed four times, three times by physicians and once by the chatbot. Primary Outcome The accuracy of the physicians ’ first differential diagnosis ranged from 50% (5 out o... | https://arxiv.org/abs/2505.20609v1 |
average (4.2 dollars) . Patient satisfaction scores ranged from 4.2 to 4.3 for care by physicians and were 3.9 for the AI interface . Therefore, in a clinical trial involving first -time patients in primary care consultations for common internal medicine cases , the AI interface achieved results in less time and at a l... | https://arxiv.org/abs/2505.20609v1 |
p articipants in our study knew that they were interacting with an AI ; thus , our findings differ somewhat from those of previous studies. This study had several limitations. First, t he number of clinical vignettes was small, and only a limited number of physicians and simulated patients participated. Second, In LLM ... | https://arxiv.org/abs/2505.20609v1 |
through patient simulation and structured feedback: a randomized controlled trial. BMC medical education. 2024;24(1):1391. doi:10.1186/s12909 -024-06399 -7 12. Tu T, Palepu A, Schaekermann M, et al. Towards conversational diagnostic ai. 2024. 13. Suh CH, Yi J, Shim WH, Heo H. Insufficient Transparency in Stochasticity ... | https://arxiv.org/abs/2505.20609v1 |
and (D) patient satisfaction with care IM = internal medicine (A) (B) (C) (D) Figure 4. Live chat with an LLM -based real-time compound diagnostic medical AI interface for a representative case and its three differential diagnoses. Figure 1. Overview of Clinical Trial Figure 1: Overview of clinical trial. P = physician... | https://arxiv.org/abs/2505.20609v1 |
arXiv:2505.20612v1 [cs.CV] 27 May 2025Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models Peter Robicheaux1,∗, Matvei Popov1,∗, Anish Madan2, Isaac Robinson1, Joseph Nelson1, Deva Ramanan2, Neehar Peri2 1Roboflow,2Carnegie Mellon University rf100-vl.org Abstract Vision-language models (... | https://arxiv.org/abs/2505.20612v1 |
internet pre-training. We demonstrate that such benchmarks artificially inflate model performance and are not representative of many real-world applications (cf. Table 1). To address this limitation, we introduce RF100-VL, a large-scale detection benchmark comprised of 100multi-modal datasets from diverse domains (cf. ... | https://arxiv.org/abs/2505.20612v1 |
query for the entire image. Detic [60] improves long-tail detection performance by utilizing image-level supervision from ImageNet [40]. Notably, recent VLMs achieve remarkable zero-shot performance and are widely used as “black box” models in diverse downstream applications [29, 37, 19, 34, 45]. More recently, multi-m... | https://arxiv.org/abs/2505.20612v1 |
platform that hosts diverse open-source datasets created to solve real-world computer vision tasks. With more than 500,000 public datasets spanning medical imaging, agriculture, robotics, and manufacturing, we focus on selecting high-quality datasets not commonly found in internet-scale pre-training (e.g. COCO [25], Ob... | https://arxiv.org/abs/2505.20612v1 |
highest number of classes (142), followed by “Industrial” (122) and “Flora & Fauna” (70). Despite having fewer classes, the “Flora & Fauna” category has the highest number of images (46,718) and annotations (441,677), indicating a higher density of annotations per image. Figure 5 (right) provides a visual representatio... | https://arxiv.org/abs/2505.20612v1 |
prompting models with class names, few-shot images, and annotator instructions. Lastly, federated fine-tuning modifies the standard cross-entropy classification to only treat exhaustively annotated classes as true negatives for each image. We follow the implementation from Madan et. al. [30] when fine-tuning Detic [60]... | https://arxiv.org/abs/2505.20612v1 |
Models Struggle on Roboflow100-VL. RF100-VL is a much harder dataset than prior open-vocabulary object detection benchmarks. Specifically, GroundingDINO achieves 49.2 mAP on ODinW-13, but only reaches 16 mAP on RF100-VL. Similar trends can be seen with Qwen2.5-VL and Gemini 2.5 Pro (cf. Table 1). Notably, both RF100-VL... | https://arxiv.org/abs/2505.20612v1 |
(Instructions Only) 5.4 4.8 14.5 5.4 1.7 7.5 7.5 7.4 Gemini 2.5 Pro [9] (Class Names Only) 8.4 13.3 22.4 9.7 3.5 11.2 17.1 13.3 Gemini 2.5 Pro [9] (Instructions Only) 4.2 10.6 9.9 3.2 0.9 6.1 7.2 6.1 Few-Shot (10 shots) Detic [60] w/ Federated Loss [30] 19.5 19.6 28.4 25.9 8.5 26.6 25.7 22.8 MQ-GLIP-Image [52] (Images ... | https://arxiv.org/abs/2505.20612v1 |
generated by GPT-4o and are manually verified for correctness. However, they may not fully reflect the nuances of real-world instructions typically developed alongside dataset collection. We encourage the community to release real annotator instructions generated through iterative discussions between annotators and sta... | https://arxiv.org/abs/2505.20612v1 |
Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. “Gpt-4 technical report”. In: arXiv preprint arXiv:2303.08774 (2023). [2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun... | https://arxiv.org/abs/2505.20612v1 |
Raff, Francis Ferraro, and Cynthia Ma- tuszek. “A Spoken Language Dataset of Descriptions for Speech-Based Grounded Language Learning”. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1) . 2021. [18] Rahima Khanam and Muhammad Hussain. “Yolov11: An overview of t... | https://arxiv.org/abs/2505.20612v1 |
Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. “Scaling Open-V ocabulary Object Detection”. In: arXiv preprint arXiv:2306.09683 (2023). [33] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, e... | https://arxiv.org/abs/2505.20612v1 |
Xie. “Eyes wide shut? exploring the visual shortcomings of multimodal llms”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 2024, pp. 9568–9578. [48] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. “A comprehensive survey of continual learning: Theory, method and application”... | https://arxiv.org/abs/2505.20612v1 |
we use the model checkpoint trained on LVIS, COCO and ImageNet-21K. We use class names provided as text prompts for Detic’s CLIP classifier. GroundingDINO. We use GroundingDINO [26] with pretrained weights from mmdetection (MM- GroundingDINO-L*). We prompt the model with all the class names combined into a single promp... | https://arxiv.org/abs/2505.20612v1 |
many few-shot image examples, the API fails to return a valid response for requests of excessive size. In such cases, we simply assign a score of 0 AP for those images. Due to Gemini and Qwen not always predicting a confidence score for their bounding boxes, we set it to 1.0 by default. YOLOv8 and YOLOv11. We train our... | https://arxiv.org/abs/2505.20612v1 |
5.4 D Comparing Different Model Sizes In Figure 6, we evaluate the performance of the Gemini model family over time (e.g. Gemini Flash 2.0 was released before Gemini Flash 2.5). Although Gemini has not been explicitly fine-tuned on RF100-VL, we see a significant increase in performance. This suggests that Gemini is mak... | https://arxiv.org/abs/2505.20612v1 |
While Madan et al.’s method performs the best, it is biased towards Detic and is computationally expensive. Our approach offers a tractable alternative that improves over the random baseline. ApproachAverage Precision (AP) All Many Medium Few Detic (Zero-Shot) [60] 14.40 25.83 16.59 2.32 Detic w/ Federated Fine-Tuning ... | https://arxiv.org/abs/2505.20612v1 |
90% STAC psuedo-labels) achieves better performance overall than YOLOv8m (medium) trained on just 10% labeled data. Table 7: Roboflow100-VL Semi-Supervised and Fully-Supervised Benchmark. We find that semi-supervised learners are able to reach nearly 80% of the performance of fully supervised models using 10% labeled d... | https://arxiv.org/abs/2505.20612v1 |
top 15 datasets where Gemini outperforms Qwen and GroundingDINO, seven overlap. This suggests that Gemini may excel on datasets similar to those found in its pretraining, but struggles to generalize to novel domains. Figure 8: Correlation Between Models Type and Performance. We see stronger linear trends between Gemini... | https://arxiv.org/abs/2505.20612v1 |
the following outline: ‘‘‘markdown # Overview Table of contents # Introduction Introduction to the dataset. Introduce what task the dataset is trying to solve. List all of the classes and provide a brief description of each class. # Object Classes ## Class 1 ### Description Provide a description of the class, paying at... | https://arxiv.org/abs/2505.20612v1 |
or very fine particles that do not form distinct objects. ## Cardboard ### Description Cardboard objects are typically flat and have a layered texture. They may appear as boxes or sheets. ### Instructions Annotate only distinguishable pieces of cardboard, focusing on their flat surfaces and any visible layering. Do not... | https://arxiv.org/abs/2505.20612v1 |
Supervised wine-labels FSOD, Fully Supervised Medical Link canalstenosis FSOD, Fully Supervised crystal-clean-brain-tumors-mri-dataset FSOD, Fully Supervised dentalai FSOD, Fully Supervised inbreast FSOD, Fully Supervised liver-disease FSOD, Fully Supervised nih-xray FSOD, Fully Supervised spinefrxnormalvindr FSOD, Ful... | https://arxiv.org/abs/2505.20612v1 |
arXiv:2505.20613v1 [cs.CL] 27 May 2025REAL-Prover: Retrieval Augmented Lean Prover for Mathematical Reasoning Ziju Shen1∗Naohao Huang2∗Fanyi Yang1∗Yutong Wang3∗Guoxiong Gao1∗ Tianyi Xu1Jiedong Jiang1Wanyi He1Pu Yang1Mengzhou Sun3Haocheng Ju1 Peihao Wu4Bryan Dai4Bin Dong567† 1Peking University2Renmin University of China... | https://arxiv.org/abs/2505.20613v1 |
Specifically, at each tactic step, we extract the current proof state from the Lean compiler and then query LeanSearch-PS to obtain candidate theorems. Second, we introduce REtrieval Augmented Lean Prover, named REAL-Prover , an integrated framework which combines the proof state information and these retrieved premise... | https://arxiv.org/abs/2505.20613v1 |
REAL-Prover model training pipeline. The figure illustrates two components: (a) the HERALD-AF pipeline, which translates informal mathematical statements into formal ones; (b) the expert iteration pipeline, which iteratively refines the prover model. lems into formal Lean 4 statements. HERALD-AF converts informal mathe... | https://arxiv.org/abs/2505.20613v1 |
best-first search strategy, selecting the node with the highest score to expand further by generating new tactics. This iterative process continues until either the proof is completed or a predefined search budget is exhausted. Once completed, the full proof can be reconstructed from the search tree. To enhance our pro... | https://arxiv.org/abs/2505.20613v1 |
3 Experiment Setup 3.1 Implementation Detail REAL-Prover Training We perform supervised fine-tuning on the base models Qwen2.5-Math-7B [23] using a learning rate of 5×10−5with a cosine decay scheduler and a maximum context length of 8192 tokens. Our training data comprises the following sources, totaling 210,420 state-... | https://arxiv.org/abs/2505.20613v1 |
core areas of elementary mathematics, including algebra, number theory, and mathematical induction. FATE-M (Formal Algebra Theorem Evaluation-Medium) FATE-M is a benchmark designed to evaluate theorem-proving capabilities in Lean4 for undergraduate-level abstract algebra. It consists of 141 problems formalized in Lean4... | https://arxiv.org/abs/2505.20613v1 |
benchmark, we compare REAL-Prover with several leading provers, Geodel-Prover [30], DeepSeek-Prover-V1.5 [ 13]. For whole-proof system, the sampling budget in table means the number of generation. And for tree-search systems, the sampling budget is M ×N, where M is the total number of passes and N is the number of ever... | https://arxiv.org/abs/2505.20613v1 |
that REAL-Prover-v1 achieves a performance improvement over REAL-Prover- v1-NoRet on both benchmarks. This demonstrates that the retrieval system enhances the prover’s performance on college-level mathematics problems. We present a comparison between proofs with and without LeanSearch-PS. In Figure 3, the proof assiste... | https://arxiv.org/abs/2505.20613v1 |
4 interactive environment, Jixia- Interactive, which facilitates both the training of our prover model, REAL-Prover-v1, and formal proof generation. To further enhance the prover’s capabilities, we present LeanSearch-PS, a theorem retrieval system that boosts performance on the college-level mathematics problem benchma... | https://arxiv.org/abs/2505.20613v1 |
and Lingpeng Kong. Subgoal-based demonstration learning for formal theorem proving. In Forty-first International Conference on Machine Learning , 2024. [10] Haiming Wang, Huajian Xin, Chuanyang Zheng, Zhengying Liu, Qingxing Cao, Yinya Huang, Jing Xiong, Han Shi, Enze Xie, Jian Yin, Zhenguo Li, and Xiaodan Liang. LEGO-... | https://arxiv.org/abs/2505.20613v1 |
Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, and Stanislas Polu. Numinamath. Hugging Face repository , 2024. [25] Huaiyuan Ying, Zijian Wu, Yihan Geng, JIayu Wang, Dahua Lin, and Kai Chen. Lean workbook: A... | https://arxiv.org/abs/2505.20613v1 |
Wen-tau Yih. Dense passage retrieval for open-domain question answering. In EMNLP (1) , pages 6769–6781, 2020. [38] Adarsh Kumarappan, Mo Tiwari, Peiyang Song, Robert Joseph George, Chaowei Xiao, and Anima Anandkumar. Leanagent: Lifelong learning for formal theorem proving. In The Thirteenth International Conference on... | https://arxiv.org/abs/2505.20613v1 |
+ c + (b + d) 13 Table 5: REAL-Prover training hyperparameters Component Setting Full Fine Tuning Learning rate : 5×10−5; Scheducler: Cosine Decay Backbone Qwen2.5-Math-7B Length Prompt Max Length: 8192 Optimizations bf16, flash_attn Batch Sizes train_batch_size = 2 Training Schedule 3 Epochs Table 6: LeanSearch-PS tra... | https://arxiv.org/abs/2505.20613v1 |
LeanSearch-PS. Figure 5: Compare the proofs with and without LeanSearch-PS. The prover in (a) uses the existing instance ‘IsPGroup.of_surjective’ from Mathlib, resulting in a more readable proof. 1.assisting mathematicians in constructing formal proofs more efficiently by automating tedious or routine steps; 2.facilita... | https://arxiv.org/abs/2505.20613v1 |
SeqPO-SiMT: Sequential Policy Optimization for Simultaneous Machine Translation Ting Xu♠*, Zhichao Huang♣†, Jiankai Sun♢, Shanbo Cheng♣†, Wai Lam♠, ♠The Chinese University of Hong Kong,♣Bytedance,♢Stanford University, xut0092@link.cuhk.edu.hk ,jksun@stanford.edu , {zhichao.huang, chengshanbo}@bytedance.com, wlam@se.cuh... | https://arxiv.org/abs/2505.20622v1 |
the overall translations. We claim that traditional RLHF methods commonly used in single-step reveal deficiencies in modeling the complex dependence relations in the multi-step SiMT setting. To this end, we propose a new policy optimiza- tion method, Sequential Policy Optimization (Se- qPO), and apply it to the SiMT ta... | https://arxiv.org/abs/2505.20622v1 |
of SiMT LLMs as a sequential decision-making process to model the complex dependencies among steps in SiMT. 2. SeqPO-SiMT fuses both translation quality and latency into a reward. With a carefully designed fu-sion function, SeqPO-SiMT successfully improves the two metrics. 3. Extensive experiments demonstrate the super... | https://arxiv.org/abs/2505.20622v1 |
employ an LLM as the policy model πθto generate transla- tions. At time step t, the input to the LLM is based on existing source text chunks x1:t= (x1, x2,···, xt)and previous translation history ˆy1:t−1= (ˆy1,ˆy2,···,ˆyt−1). Concatenating all ex- isting source text chunks and previous translations, the policy model pr... | https://arxiv.org/abs/2505.20622v1 |
pouring ") with the current text chunk ("rain hikers kept ") yields the source text (" De- spite the pouring rain, hikers kept "). This is then concatenated with the previous trans- lation (" 尽管"). We fill in the source texts and translation into the template, construct a prompt for the model to generate translation, a... | https://arxiv.org/abs/2505.20622v1 |
such as DeepSeekMath (Shao et al., 2024) and DeepSeek-R1 (DeepSeek- AI, 2025), is selected as our optimization method. Specifically, we sample Btrajectories for each to calculate the baseline reward. In SeqPO-SiMT, we choose GRPO over the popular PPO (Schulman et al., 2017) because of the following reasons: 1.Resources... | https://arxiv.org/abs/2505.20622v1 |
Figure 3: COMET v.s. AL on Zh →En and En →Zh SiMT tasks. Dataset MethodLow latency High latency BLEURT ↑COMET ↑GPT-4 ↑AL↓LAAL ↓BLEURT ↑COMET ↑GPT-4 ↑AL↓LAAL ↓ REALSISFT 64.14 83.49 83.24 15.1 15.87 64.8 83.77 84.07 18.27 18.94 SFT+wait- k 59.37 79.6 78.9 16.75 16.97 61.2 80.97 79.87 22.17 22.37 SeqPO-SiMT 65.93 84.23 8... | https://arxiv.org/abs/2505.20622v1 |
both quality and latency. The COMET and AL performance for different methods are demonstrated in Figure 3. The results show that SeqPO-SiMT consistently achieves a superior translation quality across all latency levels and all datasets, particularly in the low latency level. Other figures about COMET v.s. LAAL, BLEURT ... | https://arxiv.org/abs/2505.20622v1 |
65.84 86.75 92.27 offline SFT 66.28 86.94 92.61 offline LLaMa3 60.43 83.78 86.98 offline Qwen2.5 65.47 86.49 91.97 SeqPO-SiMT 66.76 87.55 92.7 REALSI Zh →En SFT 64.14 83.49 83.74 offline SFT 65.06 83.92 83.72 offline LLaMa3 63.27 81.19 85.66 offline Qwen2.5 64.08 82.14 85.79 SeqPO-SiMT 65.93 84.23 85.59 Table 4: Compar... | https://arxiv.org/abs/2505.20622v1 |
study. BLEURT COMET GPT-4 MUSTC En →Zh offline SFT 66.28 86.94 92.61 offline LLaMa3 60.43 83.78 86.98 offline Qwen2.5 65.47 86.49 91.97 offline SeqPO-SiMT 67.59 87.74 93.33 REALSI Zh →En offline SFT 65.06 83.92 83.72 offline LLaMa3 63.27 81.19 85.66 offline Qwen2.5 64.08 82.14 85.79 offline SeqPO-SiMT 66.82 84.62 86.79... | https://arxiv.org/abs/2505.20622v1 |
refine the translation quality and reduce latency. We conduct extensive experiments on six datasets from the diverse do- mains for En →Zh and Zh →En SiMT tasks. Ex- perimental results demonstrate that SeqPO-SiMT consistently achieves significantly higher transla- tion quality with lower latency. Moreover, SeqPO- SiMT a... | https://arxiv.org/abs/2505.20622v1 |
Maha Elbayad, Hongyu Gong, Francisco Guzmán, Kevin Heffernan, Somya Jain, Justine Kao, Ann Lee, Xutai Ma, Alex Mourachko, Benjamin Pelo- quin, Juan Pino, Sravya Popuri, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, Anna Sun, Paden Tomasello, Changhan Wang, Jeff Wang, Skyler Wang, and Mary Williamson. 2023. Seamle... | https://arxiv.org/abs/2505.20622v1 |
Jason Phang, Samuel R. Bowman, and Ethan Perez. 2023. Pretraining language models with human preferences. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , volume 202 of Proceedings of Machine Learning Research , pages 17506–17533. PMLR. Roman Koshkin, Katsuhito Sudoh,... | https://arxiv.org/abs/2505.20622v1 |
2nd edn. adaptive computation and machine learning. Changhan Wang, Juan Pino, Anne Wu, and Jiatao Gu. 2020. CoV oST: A diverse multilingual speech-to-text translation corpus. In Proceedings of the Twelfth Lan- guage Resources and Evaluation Conference , pages 4197–4203, Marseille, France. European Language Resources As... | https://arxiv.org/abs/2505.20622v1 |
experiments are shown in Table 6. During multi-step SiMT sampling, we randomly sample five translations with greedy search and tempera- ture= 1.0. Figure 9 illustrates the template used to score translations from different models. Trans- lation results were evaluated using gpt-4o-2024-08- 06. Transformer Hyper-paramete... | https://arxiv.org/abs/2505.20622v1 |
encoder- decoder methods. Figure 7: Human Evaluation between SeqPO-SiMT and the SFT model. demonstrating that SeqPO-SiMT significantly out- performs SFT in translation quality at low latency, highlighting the effectiveness of SeqPO-SiMT. B.5 Human Evaluation To verify that SeqPO-SiMTaligns with human pref- erence, we r... | https://arxiv.org/abs/2505.20622v1 |
AL on Zh →En and En →Zh SiMT tasks. 12 14 16 18 20 22 24 LAAL5860626466BLEURT SFT SFT + wait-k SeqPO-SiMT (a) REALSI Zh →En 13 14 15 16 17 LAAL57585960616263BLEURT (b) COVOST Zh →En 10 12 14 16 18 LAAL50.052.555.057.560.062.565.067.5BLEURT (c) Newstest2021 Zh →En 4 6 8 10 12 LAAL565860626466BLEURT (d) REALSI En →Zh 4 6... | https://arxiv.org/abs/2505.20622v1 |
arXiv:2505.20624v1 [cs.CL] 27 May 2025POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization Usman Naseem1, Juan Ren1, Saba Anwar2, Sarah Kohail6, Rudy Alexandro Garrido Veliz2, Robert Geislinger2,Aisha Jabr6,Idris Abdulmumin5,Laiba Qureshi2,Aarushi Ajay Borkar2, Maryam Ibrahim Mukhtar7... | https://arxiv.org/abs/2505.20624v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.