text string | source string |
|---|---|
14.24 4.80 25.90 8.81 5.60 2.41 16.44 8.16 2.03 0.98 10.14 5.63 4.74 0.23 17.03 1.01 5.00 1.98 14.55 6.79 15.73 7.97 31.80 18.17 4.30 2.46 14.68 11.25 8.08 2.65 19.87 7.65 gemma-3-4b-it 16.24 22.91 32.96 43.72 17.53 26.98 30.61 42.52 0.52 13.66 1.61 28.41 0.49 5.67 1.90 18.59 6.10 12.65 16.81 33.20 4.41 12.67 11.22 28.... | https://arxiv.org/abs/2505.18383v1 |
24.16 34.26 Table D.2: Detailed translation evaluation results on Flores and In-House datasets. Metrics: spBLEU (sB) and chrF (cF) for 0-shot (0) and 4-shot (4) scenarios. Bold values highlight the top score among models with fewer than 7 billion parameters. Underlined values indicate the highest score overall in each ... | https://arxiv.org/abs/2505.18383v1 |
ﻧظر ﺻﺎﺣﺑك وﺗﻌﺑر ﻋن اﻋﺗذارك ﺑﺻراﺣﺔ. ﻣﻣﻛن ﺗﻘول ﺣﺎﺟﺔ زي: "أﻧﺎ آﺳف إﻧﻲ اﺗﺧﺎﻧﻘت ﻣﻌﺎك اﻣﺑﺎرح. ﻛﻧت ﻏﻠطﺎن وﻣش ﻓﺎھم وﺟﮭﺔ ﻧظرك. ﻛﻧت ﻣﺣﺗﺎج أﺳﻣﻌك وأﻓﮭﻣك أﻛﺗر". ادﯾﻧﻲ ﺧﻣس ﻧﺻﺎﯾﺢ إزاي اﺣﺎﻓظ ﻋﻠﻰ وزﻧﻲ 1. .ﻛل أﻛل ﻣﺗوازن: ﺣﺎول ﺗﺎﻛل أﻛل ﻣﺗﻧوع ﻓﯾﮫ ﺑروﺗﯾن، وﺣﺑوب ﻛﺎﻣﻠﺔ، ودھون ﺻﺣﯾﺔ، وﺧﺿﺎر وﻓﺎﻛﮭﺔ ده ھﯾﺳﺎﻋدك ﺗﺣﺎﻓظ ﻋﻠﻰ وزن ﺻﺣﻲ وﯾوﻓر ﻟﺟﺳﻣك اﻟﻌﻧﺎﺻ... | https://arxiv.org/abs/2505.18383v1 |
arXiv:2505.18405v2 [cs.CL] 27 May 2025RaDeR: Reasoning-aware Dense Retrieval Models Debrup Das1, Sam O’Nuallain1, Razieh Rahimi1 1University of Massachusetts Amherst debrupdas@umass.edu, sonuallain@umass.edu, rahimi@cs.umass.edu Abstract We propose RaDeR, a set of reasoning-based dense retrieval models trained with dat... | https://arxiv.org/abs/2505.18405v2 |
where there are no match- ing terms between the question and the principle. The reasoning steps by GPT-4 also do not simplify the retrieval of the pigeonhole principle, since they outline the steps of applying the pigeonhole prin- ciple to solve the question. On the other hand, reranking models are inherently limited b... | https://arxiv.org/abs/2505.18405v2 |
outperforms BM25 in the zero-shot setting of using reasoning steps as retrieval queries. This achievement provides strong evidence for the ne- cessity of reasoning-based retrievers even when retrievers augment reasoning language models. On the MMTEB (Enevoldsen et al., 2025) reasoning subset, RAR-b (Xiao et al., 2024),... | https://arxiv.org/abs/2505.18405v2 |
3 RaDeR: Reasoning-aware Retrievers We propose a framework that includes a first-stage retriever and a re-ranking model, both performing reasoning to predict relevance. For the first-stage retriever , we adopt a uni-embedding bi-encoder architecture of dense retrieval models (Lei et al., 2023). For re-ranking , we fine... | https://arxiv.org/abs/2505.18405v2 |
Initially, all nodes have Q(s, a) = 0 , resulting in a random tree exploration. As the algo- rithm performs rollouts, the Qvalues of the nodes are updated based on the rewards Rand the search is guided toward better reasoning trajectories. Action space. We extend the rStar frame- work (Qi et al., 2025) by adding two ne... | https://arxiv.org/abs/2505.18405v2 |
generated by the MCTS framework. For a math question M, we extract all solution trajectories that contain at least one retrieval node, denoted as S=M⊕s1⊕s2⊕ ··· ⊕ sQG⊕sRT⊕ ··· ⊕ st. From each selected solution trajectory S, we gen- erate training samples in the form (q, p, N ), where qdenotes a query, pis a positive th... | https://arxiv.org/abs/2505.18405v2 |
are provided in Appendix E. 5 Experimental Settings Base language models. To train dense retrieval models, we primarily utilize the instruction-tuned variants of the Qwen2.5 suite of LLMs (Yang et al., 2024). For ablation studies on model size, we train a series of Qwen-2.5-instruct models with vary- ing parameter size... | https://arxiv.org/abs/2505.18405v2 |
24.9 10.8 16.3 15.4 30.6 1.5 7.5 27.4 11.6 17.9 RaDeR Models (MATH dataset, ( qllmq+qCoT+qlexical ) ) Qwen2.5-7B-instruct 25.4 30.0 16.7 25.3 14.0 21.3 16.3 37.0 8.2 15.7 42.7 44.4 24.6 gte-Qwen2-7B 34.6 38.9 22.1 33.0 14.8 22.5 23.7 37.3 5.0 10.2 28.4 35.1 25.5 Llama3.1-8B-Instruct 29.3 27.3 17.5 28.2 12.1 18.2 16.1 3... | https://arxiv.org/abs/2505.18405v2 |
that training retrieval models for math- ematical reasoning generalizes to other types of reasoning required for different retrieval tasks. Model Math Coding Open-source Baselines Contriever (w/ Inst.) 0.218 0.071 all-mpnet-base-v2 (w/ Inst.) 0.692 0.488 all-MiniLM-L6-v2 (w/ Inst.) 0.624 0.423 Dragon+ (w/ Inst.) 0.362 ... | https://arxiv.org/abs/2505.18405v2 |
14 shows consistent performance improvements with scaling the size of the base LLM. 6.2 Retrieval Performance on RAR-b We evaluate the performance of RaDeR on the reasoning-retrieval tasks of MMTEB (Enevoldsen et al., 2025) which are based on the Math and Cod- ing splits of the RAR-b (Xiao et al., 2024). As shown in Ta... | https://arxiv.org/abs/2505.18405v2 |
6.0 13.0 6.9 14.8 BM25 on GPT-4o CoT 53.6 53.6 24.3 38.6 18.8 22.7 25.9 19.3 17.7 3.9 18.9 20.2 26.5 Reranking on GPT-4o CoT k=100 MonoT5-3B⋆16.0 24.0 17.7 19.5 8.0 10.5 19.5 17.2 29.2 7.1 20.3 12.0 16.8 RankLLaMA-7B⋆17.5 15.5 13.1 13.6 17.9 6.9 16.9 8.4 46.8 2.2 4.5 3.5 13.9 Rank1-7B (Weller et al., 2025) 48.8 36.7 20... | https://arxiv.org/abs/2505.18405v2 |
efficiency compared to existing methods. 8 Limitations While RaDeR achieves strong performance on reasoning-intensive retrieval tasks, it has a few limitations. First, our training approach primar- ily focuses on examples where the retriever reasons over a single document in isolation. A promising direction for future ... | https://arxiv.org/abs/2505.18405v2 |
Dipam Vasani, Pranjal A Chitale, Simone Tedeschi, Nguyen Tai, Artem Snegirev, Mariya Hen- driksen, Michael Günther, Mengzhou Xia, Weijia Shi, Xing Han Lù, Jordan Clive, Gayatri K, Mak- simova Anna, Silvan Wehrli, Maria Tikhonova, He- nil Shalin Panchal, Aleksandr Abramov, Malte Osten- dorff, Zheng Liu, Simon Clematide,... | https://arxiv.org/abs/2505.18405v2 |
. Xueguang Ma, Luyu Gao, Shengyao Zhuang, Ji- aqi Samantha Zhan, Jamie Callan, and Jimmy Lin. 2025. Tevatron 2.0: Unified document retrieval toolkit across scale, language, and modality. arXiv preprint arXiv:2505.02466 .Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023. Fine-tuning llama for multi-stage ... | https://arxiv.org/abs/2505.18405v2 |
arXiv:2503.05592. Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2023. One embedder, any task: Instruction-finetuned text em- beddings. In Findings of the Association for Compu- tational Linguistics: ACL 2023 , pages 1102–1121, Toro... | https://arxiv.org/abs/2505.18405v2 |
Approach zero and anserini at the clef-2021 arqmath track: Applying substructure search and bm25 on operator tree path tokens. Proc. CLEF 2021 (CEUR Working Notes) . A Detailed Discussion of Related Works A.1 Mathematical Information Retrieval Math Information Retrieval (Math IR) has been extensively studied within the... | https://arxiv.org/abs/2505.18405v2 |
applied to trees (UCT) to traverse through child nodes, continuing until a leaf node is reached. For- mally, we select the node with maximum UCT value at each branch of the traversal: UCT( s, a) =Q(s, a) N(s, a)+cs ln(Nparent (s)) N(s, a) where N(s, a)is the number of times node shas been visited till now and Q(s, a)is... | https://arxiv.org/abs/2505.18405v2 |
LLM model (such as Qwen or Llama), which returns the last layer token representations for each input token. We take the representation of the end-of-sequence token as the representation of the input sequence t1, . . . , t k, which can be either a query qor a document d. Relevance of dtoqis computed using the cosine sim... | https://arxiv.org/abs/2505.18405v2 |
We highlight repre- sentative cases where RaDeR successfully retrieves the correct theorem, as well as illustrative failure cases to analyze its limitations. Examples of queries from Retrieval Training Dataset. We present a qualitative example of different query types ( qllmq, qCoTandqlexical ) for a question Mand posi... | https://arxiv.org/abs/2505.18405v2 |
41.6 42.5 26.1 41.2 18.6 32.5 31.7 44.7 3.4 13.4 32.5 47.9 31.3 Llama3.1-8B-Instruct 35.7 33.5 18.4 36.4 14.8 28.5 20.3 45.2 6.4 7.7 37.0 43.4 27.3 RaDeR Models (MATH +NuminaMath datasets, all query types) Qwen2.5-7B-instruct 31.3 34.2 21.5 31.9 18.8 30.5 21.5 45.6 5.1 11.7 44.7 63.8 30.1 Table 12: Recall@10 performanc... | https://arxiv.org/abs/2505.18405v2 |
two of the functions are equal. The graph of j(x)has horizontal segments and linear segments connecting the points where the maximum changes. Lexical query ( qlexical ): What is the definition of a linear real function and what form must it take for all real numbers x? Figure 3: Examples of different query types from o... | https://arxiv.org/abs/2505.18405v2 |
overlap with the original question.) Qwen2 retriever Top Retrieved Theorem :Rooted Tree Corresponds to Arborescence: Let T= (V, E)be a rooted tree with root r. Then there exists a unique orientation of Tthat forms an r-arborescence. (Incorrect theorem retrieved) RaDeR GTE-Qwen2-7B-instruct (all query types) retriever T... | https://arxiv.org/abs/2505.18405v2 |
10. P Prompt for Self Reflection MCTS In this section, we provide the prompts for the self reflection andself-summarization mechanisms used in our MCTS framework. The self-reflection prompt is shown in Figure 11 and the summatiza- tion prompt used is presented in Figure 12. Q Prompt for LLM Reasoning Query generation I... | https://arxiv.org/abs/2505.18405v2 |
points (x1, y1)and(x2, y2)in the complex plane is given byp (x2−x1)2+ (y2−y1)2. Joe’s point is (1,2)and Gracie’s point is (−1,1). Next steps: The distance isp ((−1)−(1))2+ ((1) −(2))2=√4 + 1 =√ 5. Therefore, Gracie and Joe’s points are√ 5units apart. Context: What is the sum of all positive integer values of nfor which... | https://arxiv.org/abs/2505.18405v2 |
mathematical question, an intermediate solution and a mathematical theorem which was retrieved denoted as Retrieved Document. First, please judge whether the mathematical theorem is relevant with the question and the intermediate solution, and put it in the relevant field. If the provided content is irrelevant to the q... | https://arxiv.org/abs/2505.18405v2 |
are put into mcontainers, with n > m , then at least one container must contain more than one item. Preconditions : (1) There are more items than containers. (2) The items are distributed into containers. Why Preconditions are Satisfied: (1) The warehouse has 65 boxes (items) and needs to distribute them among shelves ... | https://arxiv.org/abs/2505.18405v2 |
DanmakuTPPBench: A Multi-modal Benchmark for Temporal Point Process Modeling and Understanding Yue Jiang1, Jichu Li2, Yang Liu3, Dingkang Yang1, Feng Zhou2,4˚, Quyu Kong3˚ 1Fudan University 2Center for Applied Statistics and School of Statistics, Renmin University of China 3Alibaba Cloud 4Beijing Advanced Innovation Ce... | https://arxiv.org/abs/2505.18411v1 |
gap, we propose DanmakuTPPBench , a new benchmark comprising two datasets: DanmakuTPP-Events , for conventional multi-modal TPP modeling, and DanmakuTPP-QA , a novel question-answering dataset designed to assess deeper temporal and cross-modal understanding. We first introduce DanmakuTPP-Events , a novel multi-modal TP... | https://arxiv.org/abs/2505.18411v1 |
categorized into three groups based on follower counts. The StackOverflow dataset [ 10] records user badge-awarding events with 22 distinct badge types. The Taobao dataset [ 26] captures user click behavior across 17 item categories, while the Taxi dataset [ 25] logs pick-up and drop-off events in New York City, catego... | https://arxiv.org/abs/2505.18411v1 |
watch, and annotate videos with Danmaku comments. We manually collect all videos posted by the top 100 most popular creators on Bilibili in 20242, resulting in a total of 7,250 videos. From these videos, we construct a large-scale multi-modal TPP dataset, DanmakuTPP-Events , which includes 10,820,790 Danmaku events. Fo... | https://arxiv.org/abs/2505.18411v1 |
UnRefinedfrom Annotator What a GOAL !!Danmaku Time Annotation Sentiment Visual ElementsEvent Type Code -LLM InterpreterFigure 3: Multi-agent framework for automated construction of DanmakuTPP-QA. The framework consists of five main components: (1) DanmakuTPP-Events (top left) containing synchronized video frames, times... | https://arxiv.org/abs/2505.18411v1 |
Task-2Prediction of the next Danmaku timestampRMSE Closed-ended Task-3Prediction of the next Danmaku burst peak timestampRMSE Closed-ended Task-4Assessment of average sentiment polarityRMSE Closed-ended Task-5Sentiment polarity prediction for the next DanmakuRMSE Closed-ended Task-6Sentiment polarity prediction for the... | https://arxiv.org/abs/2505.18411v1 |
validation performance. We note that the multi-modal information is not used for these models. DanmakuTPP-QA Evaluation. We frist evaluate the TPP understanding capabilities of existing open source pretrained LLMs (Qwen2.5 series [ 28], Qwen3 series [ 29] and Deepseek-V3 [ 15]) and MLLMs (Gemma3 [ 23] and Qwen2.5-VL [ ... | https://arxiv.org/abs/2505.18411v1 |
1.28 123.45 0.30 0.46 0.46 16.00 43.83 Qwen3-8B 6.67 1.80 123.59 0.32 0.41 0.45 19.33 41.50 Qwen3-30B-A3B 0.67 1.33 121.96 0.20 0.33 0.40 23.00 43.67 Qwen3-235B-A22B 8.67 1.39 120.79 0.30 0.31 0.29 10.33 32.50 Llama-3.3-70B-Instruct 1.67 1.11 121.49 0.26 0.27 0.22 17.00 33.33 DeepSeek-V3 25.00 1.30 121.30 0.34 0.26 0.2... | https://arxiv.org/abs/2505.18411v1 |
The finetuned Qwen2.5-VL-3B model, despite its relatively small size (3B parameters), outperforms all larger pretrained models on sentiment-related prediction tasks (Tasks 4-6), achieving RMSE of 0.05, 0.16, and 0.08, respectively. This represents a reduction in error by factors of 4-6 × compared to the best pretrained... | https://arxiv.org/abs/2505.18411v1 |
between specialized TPP models and general-purpose language models demonstrates the need for improved integration of temporal point process understanding into multi-modal systems. Limitations: despite these contributions, our work has several limitations. First, the dataset is primar- ily sourced from Chinese-language ... | https://arxiv.org/abs/2505.18411v1 |
Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. [16] Zefang Liu and Yinzhu Quan. TPP-LLM: Modeling temporal point processes by efficiently fine-tuning large language ... | https://arxiv.org/abs/2505.18411v1 |
Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388 . [30] Chenghao Yang, Hongyuan Mei, and Jason Eisner. Transformer embeddings of irregularly spaced events and their particip... | https://arxiv.org/abs/2505.18411v1 |
serves to identify deficient attributes within the dataset, thereby supporting the resolution of research tasks designed in the preceding module. Within the Annotation Agent group, Qwen2.5 handles textual annotation while Qwen2.5-VL and RAM manage visual annotation. The guiding prompt template and representative annota... | https://arxiv.org/abs/2505.18411v1 |
of Danmaku :0.2 -Visual elements within video frames : [dog, human, frisbee, garden]Quality control Agents{ Output of Annotation Agent Group } Danmaku example : What breed is this dog? Annotation Agent Compare and Analyze Annotations : - Systematically compare the outputs across all Annotation Agents - Identify areas o... | https://arxiv.org/abs/2505.18411v1 |
LatentLLM: Attention-Aware Joint Tensor Compression Toshiaki Koike-Akino, Xiangyu Chen, Jing Liu, Ye Wang, Pu (Perry) Wang, Matthew Brand Mitsubishi Electric Research Laboratories (MERL), Cambridge, MA 02139, USA {koike, xiachen, jiliu, yewang, pwang, brand }@merl.com Abstract Modern foundation models such as large lan... | https://arxiv.org/abs/2505.18413v1 |
four categories: weight quan- tization [11, 23, 40], network pruning [3, 10, 12, 18], knowledge distillation [7, 13, 14], and low-rank factoriza- tion [14, 24, 32, 34, 47]. Among these methods, weight quantization has gained significant traction in the context of large foundation mod- els due to its effectiveness. Howe... | https://arxiv.org/abs/2505.18413v1 |
X∈Rd×l(lis the calibration sam- ple length), ASVD aims to minimize the activation loss: L1=EX WX−ˆWX 2=EX WX−BAX 2,(1) instead of the na ¨ıve weight-based loss: L0= W−ˆW 2= W−BA 2. (2) It is well-known that the optimal solution to minimize L0 can be given by the plain SVD of W. To minimize L1, ASVD introduced a pre-con... | https://arxiv.org/abs/2505.18413v1 |
x ddx dd’x rrx rrx ddx rrx ddx dSVDMap WPUSVBAPd’ddTensor Diagramsdd’rrdd’rdUSVdd’rrJJ+rr=USVd’x rrx rrx d=Jrx rJ+rx rWeightPreConditionLeftSingularSingularValueRightSingularInsert Junction=Block IdentityJsuch that=IA2Figure 2. Activation-aware compression with pre-conditioning and junction matrix. The junction matrix ... | https://arxiv.org/abs/2505.18413v1 |
q,iBk,i|{z} Hi∈Rrq×rkAkC1 2|{z} A′ k 2 =hX i=1 Gi−A′⊤ qHiA′ k 2. (13) This is known as a high-order SVD (HOSVD) problem to decompose for the 3-mode tensor G∈Rh×d×d, whose ith GhG1Wq: (h, dh, d)Wq,1Wq,hWk: (h, dh, d)Wk,1Wk,hG=einsum(“hij,hik->hjk”, Wq, Wk) Dot-ProductG: (h,d, d)HTucker DecompH: (h,rq, rk)AkAqIAq: (rq, d... | https://arxiv.org/abs/2505.18413v1 |
2, (15) for the low-rank compression: ˆWo,i=BoAo,i∈Rd′×dh andˆWv,i=Bv,iAv∈Rdh×dwithBo∈Rd′×ro,Ao,i∈ Rro×dh,Bv,i∈Rdh×rv, and Av∈Rrv×d. The MLA out- put is thus given as ˆY′=hX i=1BoAo,iBv,iAvXsoftmax [Mi]. (16) Interestingly, this is also formulated in a similar manner of Eq. (13), and it can be solved by the joint SVD a... | https://arxiv.org/abs/2505.18413v1 |
635.5 509.8 ASVD (Hessian) 64.0 89.1 50.9 104.6 134.4 80.3 202.1 212.0 145.9 557.3 558.6 371.6 ASVD ( ℓ2-norm) 40.0 59.9 36.6 59.4 78.0 49.8 117.5 134.2 86.9 308.7 283.9 201.1 ASVD (Cov) 78.0 90.6 61.7 100.8 111.0 72.7 311.2 356.8 129.4 1485.3 922.7 548.2 ASVD (RootCov) 30.8 42.2 28.5 39.0 51.4 33.6 71.6 86.1 49.5 118.... | https://arxiv.org/abs/2505.18413v1 |
for LLM and LMM benchmarks to evaluate the effectiveness of our method. Our experiments are based on the same setting of SparseLLM [3] and their code base1. We implemented La- tentLLM in PyTorch [29] and used the HuggingFace Trans- formers library [43] for handling models and datasets. All experiments are conducted on ... | https://arxiv.org/abs/2505.18413v1 |
multiply-accumulation operations (MACs), and parameters are almost linearly reduced with the compression factor. Compression over model size We first look into the compression capabilities of our LatentLLM across various model sizes in comparison to baseline methods. Detailed re- sults are shown for a size reduction ov... | https://arxiv.org/abs/2505.18413v1 |
Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774 , 2023. 1 [2] Guangji Bai, Zheng Chai, Chen Ling, Shiyu Wang, Jiay- ing Lu, Nan Zhang, Tingwei Shi, Ziyang Yu, Mengdan Zhu,Yifei Zhang, et al. Beyond efficiency: A systematic survey of ... | https://arxiv.org/abs/2505.18413v1 |
37.91 41.30 35.15 38.33 38.62 35.27 37.42 ASVD (RootCov) 10% 64.08 56.13 57.36 64.03 60.98 57.35 62.70 57.02 60.67 LatentLLM (RootCov) 10% 68.52 64.23 61.36 69.06 65.20 61.53 68.72 60.45 65.76 Plain SVD (Identity) 20% 0.18 0.00 0.00 0.20 0.20 0.00 0.04 0.20 0.09 ASVD (Hessian) 20% 3.82 2.81 0.00 3.62 5.30 0.14 3.01 1.9... | https://arxiv.org/abs/2505.18413v1 |
Yun- fan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y . X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui ... | https://arxiv.org/abs/2505.18413v1 |
Zhou. Beyond linear approximations: A novel pruning approach for attention matrix. arXiv preprint arXiv:2410.11261 , 2024. 4 [22] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. MoE-LlaVa: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.1... | https://arxiv.org/abs/2505.18413v1 |
A Smith, and Oren Et- zioni. Green AI. Communications of the ACM , 63(12):54– 63, 2020. 1 [37] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing , 568:127063, 2024. 27 [38] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zi... | https://arxiv.org/abs/2505.18413v1 |
The best solution is given by SVD of Was follows: A=J+V, (26) B=USJ, (27) where U∈Rd′×risrmost-principal left-singular vectors, S=diag[σ1, . . . , σ r]∈Rr×ris diagonal singular-values, and V∈Rr×dis the most-principal right-singular vectors for W: USV =svdr[W], (28) where we assume the singular values are sorted in the ... | https://arxiv.org/abs/2505.18413v1 |
Diagonal L1-norm (e.g., AWQ [23], ASVD [47]): P= diag[ ∥[X]1,:∥1, . . . ,∥[X]d,:∥1] • Diagonal Hessian (e.g., OBS [12], GPTQ [11], SparseGPT [10]): P= diag[( XX⊤+λI)−1]−1 2 • Identity (Plain SVD, e.g., [33]): P=I In the context of fine-tuning initialization, CorDA [45] uses covariance matrix Cwithout square root, which... | https://arxiv.org/abs/2505.18413v1 |
approximation. It is shown in Fig. 9. E. Multi-Head Attention (MHA) Typically, the attention projection uses a square weight d′=d, but it is divided into multiple heads such that: Wq= Wq,1 Wq,2 ... Wq,h ∈Rd×d, W k= Wk,1 Wk,2 ... Wk,h ∈Rd×d, (55) 17 where Wq,i∈Rd/h×dandWk,i∈Rd/h×dare the ith head pro... | https://arxiv.org/abs/2505.18413v1 |
optimized to minimize the loss, rather than using the same rank. It can be adaptively adjusted from the eigenvalue distributions. Once we obtained the HOSVD solution for tensor planes AqandAk, the tensor core Hi∈Rrq×rkis generated by (64) as Hi=AqGiA⊤ k=AqW⊤ q,iWk,iA⊤ k. Given optimized Hi, any arbitrary Bq,iandBk,ipro... | https://arxiv.org/abs/2505.18413v1 |
1 | {z } A′⊤q∈R(d+1)×(rq+1)H′ i∈R(rq+1)×(rk+1) z }| {B⊤ q,i ˆb⊤ q,i |{z} B′⊤ q,i∈R(rq+1)×d Bk,iˆbk,i |{z} B′ k,i∈Rd×(rk+1) AkOrk×1 O1×d 1 | {z } A′ k∈R(rk+1)×(d+1). (105) 22 We may use the HOSVD to decompose with one more rank for bias, while the compression matrix A′ qandA′ kneeds to be a particular format. Non... | https://arxiv.org/abs/2505.18413v1 |
0W⊤ q,i(Wk,iµ+bk,i)(Wk,iµ+bk,i)⊤Wq,iC1 2 0 . (140) 24 The first term is the solution if no bias and mean are present. Similarly the solution for Akis given by AkC1 2 0=RightSingularrk C1 2 0Id×(d+1)(X iGk,i)I(d+1)×dC1 2 0 (141) =RightSingularrqX iC1 2 0W⊤ k,iWq,iC0A⊤ qAqC0W⊤ q,iWk,iC1 2 0 +X iC1 2 0W⊤ k,i(Wq,iµ+bq,... | https://arxiv.org/abs/2505.18413v1 |
qEkand there is no impact in loss function with low-rank approximation. F.3. Multiplicative PE Consider a multiplicative PE for token X: X′=X⊙E, (171) where ⊙denotes Hadamard product. We just need to replace the correlation with C′=X′X′⊤to solve in a straightforward manner. However, rotary PE (RoPE) [37] uses multiplic... | https://arxiv.org/abs/2505.18413v1 |
ˆbois given − Wo,i(Wv,iX+bv,i1⊤) +bo,i1⊤−ˆWo,i(ˆWv,iX+ˆbv,i1⊤)−ˆbo,i1⊤ 1. (192) Thus the KKT condition gives: ˆbo,i=bo,i+Wo,i(Wv,iµ+bv,i)−ˆWo,i(ˆWv,iµ+ˆbv,i). (193) Plugging into the loss gives: L=X i∥Wo,iWv,i(X−µ1⊤)−ˆWo,iˆWv,i(X−µ1⊤)∥2(194) =X i∥Wo,iWv,i|{z} Gi∈Rd×dC1 2 0−BoAo,iBv,i|{z} Hi∈Rro×rvAvC1 2 0∥2. (195) He... | https://arxiv.org/abs/2505.18413v1 |
last KKT condition requires solving in vectorization: X i,j(GjG⊤ i⊗H⊤ jHi)vec[Av] =X jvec[HjB⊤ oWoC1 2G⊤ j], (222) where Gi∈Rd×hdhis defined: Gi=W+ v,i[C1 2 0]i. (223) H. MLP-Aware Joint Compression SparseLLM [3] proposed the way to sparsify MLP layer in LLM models as it consumes two thirds of trainable parameters. The... | https://arxiv.org/abs/2505.18413v1 |
Fig. 14. However, unstructured sparse matrix may require index storage to memorize the non-zero entry locations. When we use a mask, it requires d′dbinary memory as well as non-zero values in D. When the sparsity is small, keeping index will be more efficient, i.e., keeping log2(dd′)τ. Fig. 15 shows the case with spars... | https://arxiv.org/abs/2505.18413v1 |
8192 ( 4d) head dim dh=d/h 64 128 64 num attention heads h 12 32 32 num key value heads hkv 12 32 8 num hidden layers L 12 32 16 qkv bias True False False mlp bias True False False rope theta θ — 1e4 5e5 max position embeddings 197 4096 131072 OPT-350M BLOOM-560M Qwen2-0.5B ID facebook/opt-350m bigscience/bloom-560m Qw... | https://arxiv.org/abs/2505.18413v1 |
Retrieval Augmented Generation -based Large Language Models for Bridging Transportation Cybersecurity Legal Knowledge Gaps Khandakar Ashrafi Akbar * Research Assistant Data Mining Lab The University of Texas at Dallas, 800 W Campbell Rd, Richardson, TX 75080 Email: ashrafi@utdallas.edu Md Nahiyan Uddin Research Assista... | https://arxiv.org/abs/2505.18426v1 |
Large Language Model (LLM), Hallucination , Cybersecurity, Transportation, Policy, Legal 3 1. INTRODUCTION The rise of connected and automated transportation systems underscores the pressing need to address legal uncertainties at the federal, state, and international levels. The intricate nature of these systems —parti... | https://arxiv.org/abs/2505.18426v1 |
the effective handling of targeted queries that reveal legislative knowledge gap s—without requiring extensive retraining of the underlying models. The objective of this study is to leverage the RAG framework to answer queries that can reveal these 4 legislative knowledge gaps accurately . This method eliminates the ne... | https://arxiv.org/abs/2505.18426v1 |
robust retrieval with domain grounding. Liu et al. [1 4] introduced HaDeS, targeting hallucination at the token level without needing full ground truth. While this granularity is valuable, it lacks applicability in domains like law or medicine, where reference precision and context are essential. In another study, Gao ... | https://arxiv.org/abs/2505.18426v1 |
marginalizing over the latent documents, producing a distribution over the generated text. The key di fference between these two models lies in handling retrieved documents during generation. In the RAG -sequence model, distinct output sequences are generated for each retrieved document, which are subsequently marginal... | https://arxiv.org/abs/2505.18426v1 |
single document will contain multiple chunks). These chunks serve as the node in the vector database and embeddings for each node are generated and s tored. When a query is presented to the RAG framework, this entire database is searched through to find the top k (k is variable here, e.g., it can take a value of 5, 10,... | https://arxiv.org/abs/2505.18426v1 |
are provided to a generator model, which uses them as context along with the question to generate a response. This entire process is hosted on a web interface, where users can input questions. Base d on the question, the response is generated on the backend, and the processed responses are displayed on the frontend. Ta... | https://arxiv.org/abs/2505.18426v1 |
(UTSA) or analogous statutory framework were included. 37 38 Across all search terms, the following method was consistently applied: results were filtered by 39 jurisdiction, limited to the “Code of [State]” source category, and inclusive of all publication dates 40 available in LexisNexis. Relevant code sections were ... | https://arxiv.org/abs/2505.18426v1 |
for our framework generation. It 3 highlights the range of collected documents, the various sources of these legislations, and the 4 corresponding document counts. This table breaks down the number of statutes and other 5 legislative documents that have been compiled and indexed in the LLM’s dataset. As shown, the 6 da... | https://arxiv.org/abs/2505.18426v1 |
at federal, state, and international levels into indexed embeddings. When queried 6 about specific contexts or laws, these embeddings are utilized to retrieve the most pertinent 7 information along with their corresponding contexts and sources. 8 9 In Figure 9, the sources displayed are references from documents identi... | https://arxiv.org/abs/2505.18426v1 |
aspect, ensuring alignment with base or reference answers. The next 7 phase involves curating a more comprehensive set of questions to identify and highlight the 8 existing loopholes in the legal landscape of autonomous and connected vehicles. 9 Table 5 presents sample responses generated by baseline LLMs (such as Chat... | https://arxiv.org/abs/2505.18426v1 |
public health, safety, or any combination thereof. Critical infrastructure means existing and proposed information technology and operational technology systems and assets, whether physical or virtual, the incapacity or destruction of which would negatively affect security, economic security, public health , or public ... | https://arxiv.org/abs/2505.18426v1 |
document" is defined as: Any document containing data that is issued to an individual and which that individual, and only that individual, uses alone or in conjunction with any other information for the primary purpose of establishing his or her identity or accessing his or her fi nancial information or benefits. The s... | https://arxiv.org/abs/2505.18426v1 |
Nano —Gemini was designed to compete 36 with models like GPT -4, with Ultra reportedly surpassing them on industry benchmarks. By early 37 2024, Gemini expanded into more applications, including Bard, Pixel smartphones, and Google 38 Cloud's Vertex AI, with updates like Gemini 1.5 enhancing its capabilities and context... | https://arxiv.org/abs/2505.18426v1 |
0.29974 Claude 0.66579 0.66195 0.84620 0.89018 0.86737 0.25837 0.326 60 0.25692 RAG -powered GPT 0.73201 0.7043 3 0.85312 0.88861 0.87033 0.32576 0.56292 0.37340 33 The advantages of the 'State -Wise Index' method are explained through experimental analysis, 34 with the results presented in Table 7. The state -wise ind... | https://arxiv.org/abs/2505.18426v1 |
26 27 These findings suggest that even state -of-the-art LLMs face challenges when responding to 28 specialized queries, particularly those involving recently enacted legislation. This limitation is 29 significant, as inaccurate outputs can misidentify legislative gaps and complicate the work of legal 30 and policy sta... | https://arxiv.org/abs/2505.18426v1 |
of launching the tool for public use, with a primary focus on 34 collecting user feedback to improve its performance in real -world settings. As part of the 35 framework’s extension, we aim to enable both legal and technical users to upload relevant 36 documents or information and receive accurate, context -aware answe... | https://arxiv.org/abs/2505.18426v1 |
depending on the size of the dataset and the embedding model used, as it requires 23 re-generating embeddings for all documents and rebuilding the entire vector index from scratch. 24 25 11. ACKNOWLEDGMENTS 26 27 This work is based upon the work supported by the National Center for Transportation 28 Cybersecurity and R... | https://arxiv.org/abs/2505.18426v1 |
Xu, G., Deng, J., Cheng, J., Zheng, C., Zhou, H., Peng, N., Zhu, X., & Huang, M. (2021). On the Safety of Conversational Models: Taxonomy, Dataset, and Benchmark. ArXiv. /abs/2110.08466 10. Zhang, Y ., Li, Y ., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y ., Chen, Y ., Wang, L., Luu, A. T., Bi, W., ... | https://arxiv.org/abs/2505.18426v1 |
. Zhang (Eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (pp. 3178 –3190). Association for Computational Linguistics. https://doi.org/10.18653/v1/2022.emnlp - main.208 28. Zhang, T., Kishore, V ., Wu, F., Weinberger, K. Q., & Artzi, Y . (2020). BERTScore: Evaluating Text Ge... | https://arxiv.org/abs/2505.18426v1 |
arXiv:2505.18436v1 [cs.CL] 24 May 2025 Voice of a Continent: Mapping Africa’s Speech Technology Frontier AbdelRahim ElmadanyξSang Yun KwonξHawau Olamide ToyinΩ Alcides Alcoba InciarteξHanan AldarmakiΩMuhammad Abdul-Mageedξ,Ω,λ ξThe University of British ColumbiaΩMBZUAIλInvertible AI {a.elmadany,muhammad.mageed}@ubc.ca ... | https://arxiv.org/abs/2505.18436v1 |
of these resources and systems. In particu- lar, we offer a number of contributions: (1) New Speech Benchmark: we conduct extensive data collection and aggregate and harmonize all pub- licly available resources covering ASR, TTS, and SLID tasks. This dataset collection spans diverse linguistic families and geographic r... | https://arxiv.org/abs/2505.18436v1 |
tasks (Adebara and Abdul-Mageed, 2022; Adebara et al., 2025). Benchmarks like SA- HARA (Adebara et al., 2025), IrokoBench (Ade- lani et al., 2024), and others (Ojo et al., 2023; Wang et al., 2023; Oladipo et al., 2023; Reid et al., 2021) have advanced NLU and NLG ca- pabilities. In terms of model development, mod- els ... | https://arxiv.org/abs/2505.18436v1 |
4 14.27 TV SPCS (Code-Switched) (Modipa et al., 2015) 1 10.48 R ASR Statistics 42 6,109.09SLIDNicolingua (0003) (Doumbouya et al., 2021) 6 143.75 R OlongoAfrica (Ours) 10 2.40 SS UDHR (Ours) 6 1.05 HR V oice of Africa (VOA) (Ours) 10 865.08 N V oxLingua (Valk and Alumäe, 2021) 9 773.66 V Zambezi V oice (Audio Only) (Si... | https://arxiv.org/abs/2505.18436v1 |
and eliminating excessive noise, and meta- data consolidation ; reformatting datasets into a unified JSON schema with standardized fields. Our preprocessing pipeline enables robust training and evaluation across diverse African speech corpora, establishing a foundation for consistent benchmark- ing and inclusive model ... | https://arxiv.org/abs/2505.18436v1 |
per language. 0 200 400 600 800 1000 1200 1400 Total Audio Hours0.00000.00050.00100.00150.00200.0025Density (b) Distribution density (KDE). (c) Hours vs. number of sources. Figure 3: Speech data distribution across the 61 African languages in collected data, highlighting volume, density, and source diversity. in media,... | https://arxiv.org/abs/2505.18436v1 |
train- ing approaches to assess their performance on African language audio data. We evaluate Whis- per (Radford et al., 2022), Seamless (Anastasopou- Simba Series (Ours) Language Test Set MMS Seamless Whisper WhisperTSimba -H Simba -M Simba -S Simba -X Simba -W Akuapim-twi (aka) FS 85.82/40.14 219.67/190.49 1181.0/113... | https://arxiv.org/abs/2505.18436v1 |
that supports TTS, we finetune the MMS-TTS model (Pratap et al., 2023) extending support to additional African languages. The original MMS- TTS model only supports 4out of the 11African languages included in our collection. As a result, unlike the ASR setup, we do not finetune on the entire multilingual dataset; instea... | https://arxiv.org/abs/2505.18436v1 |
mod- els. Specifically, Susu ,Tigre ,Tigrinya , and Ga consistently yield high error rates, revealing sub- stantial gaps in support for certain under-resourced languages. Our finetuned Simba -ASR models im- prove upon every test sets compared to the baseline systems, with Simba -S achieving the best overall performance... | https://arxiv.org/abs/2505.18436v1 |
does not ex- tend to every high-resource language: for Amharic andAfrikaans , Seamless—with far smaller cover- age—occasionally surpasses MMS, suggesting that focused training and larger model size can over- come limited coverage when sufficient in-domain data exist. Conversely, Whisper, covering only nine African lang... | https://arxiv.org/abs/2505.18436v1 |
self- supervised pretraining, multitask learning, or data augmentation. 4.Implementation Constraints. Despite our advocacy for inclusive data and policy re- form, real-world implementation requires sus- tained institutional commitment. Bridging the gap between research and impact will necessi- tate long-term investment... | https://arxiv.org/abs/2505.18436v1 |
Emirates. Association for Computational Linguistics. Ife Adebara, AbdelRahim Elmadany, Muhammad Abdul-Mageed, and Alcides Alcoba Inciarte. 2022b. Serengeti: Massively multilingual language models for africa. arXiv preprint arXiv:2212.10785 . Ife Adebara, Hawau Olamide Toyin, Nahom Tesfu Ghe- bremichael, AbdelRahim Elma... | https://arxiv.org/abs/2505.18436v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.