text string | source string |
|---|---|
Simeng Han, Hailey Schoelkopf, Yilun Zhao, Zhenting Qi, Martin Riddell, Wenfei Zhou, James Coady, David Peng, Yujie Qiao, Luke Benson, Lucy Sun, Alex Wardle-Solano, Hannah Szabo, Ekaterina Zubova, Matthew Burtell, Jonathan Fan, Yixin Liu, Brian Wong, Malcolm Sailor, Ansong Ni, Linyong Nan, Jungo Kasai, Tao Yu, Rui Zhan... | https://arxiv.org/abs/2505.18658v1 |
URL https://aclanthology.org/2024.findings-emnlp.785/ . Nandan Kumar Jha and Brandon Reagen. Aero: Softmax-only llms for efficient private inference, 2024. URL https://arxiv.org/abs/2410.13060 . Chengze Jiang, Zhuangzhuang Wang, Minjing Dong, and Jie Gui. Survey of adversarial robustness in multimodal large language mo... | https://arxiv.org/abs/2505.18658v1 |
2025a. URL https://arxiv.org/abs/2504.12314 . Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. URL https://arxiv.org/abs/ 2305.11747 . LinLi, YifeiWang, ChawinSitawarin, andMichaelW.Spratling. OODRobustben... | https://arxiv.org/abs/2505.18658v1 |
Ruocheng Guo, Hao Cheng, Yegor Klochkov, Muhammad Faaiz Taufiq, and Hang Li. Trustworthy llms: a survey and guideline for evaluating large language models’ alignment, 2024b. URL https://arxiv.org/abs/2308.05374 . Shuo Lu, Yingsheng Wang, Lijun Sheng, Aihua Zheng, Lingxiao He, and Jian Liang. Recent advances in ood dete... | https://arxiv.org/abs/2505.18658v1 |
Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Cha... | https://arxiv.org/abs/2505.18658v1 |
Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report, 2024. URL https://arxiv.org/abs/2303.08774 . Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex ... | https://arxiv.org/abs/2505.18658v1 |
Processing, pp. 8988– 9003, Miami, Florida, USA, November 2024. Association for Computational Linguistics. doi: 10.18653/ v1/2024.emnlp-main.507. URL https://aclanthology.org/2024.emnlp-main.507/ . Chufan Shi, Haoran Yang, Deng Cai, Zhisong Zhang, Yifan Wang, Yujiu Yang, and Wai Lam. A thorough examination of decoding ... | https://arxiv.org/abs/2505.18658v1 |
Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subra- manian, Xiaoqing Ellen Ta... | https://arxiv.org/abs/2505.18658v1 |
Chen, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. Towards robust alignment of language models: Distributionally robustifying direct preference optimization. In TheThirteenth International Conference onLearning Representations, 2025. URL https://openreview.net/forum?id=CbfsKHiWEn . 46 Shijie Wu, Ozan Irsoy, St... | https://arxiv.org/abs/2505.18658v1 |
An, and Qingsong Wen. A survey on trustworthy llm agents: Threats and countermeasures, 2025. URL https://arxiv.org/abs/2503.09648 . Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. Rankrag: Unifying context ranking with retrieval-augmented generation in llms, 2024... | https://arxiv.org/abs/2505.18658v1 |
Haohan Wang. Robust prompt optimization for defending language models against jailbreaking attacks, 2024a. URL https://arxiv.org/abs/2401.17263 . Yujia Zhou, Yan Liu, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Zheng Liu, Chaozhuo Li, Zhicheng Dou, Tsung-Yi Ho, and Philip S. Yu. Trustworthiness in retrieval-augmented generati... | https://arxiv.org/abs/2505.18658v1 |
arXiv:2505.18668v1 [cs.CV] 24 May 2025ChartGalaxy: A Dataset for Infographic Chart Understanding and Generation Zhen Li1∗Yukai Guo1∗Duan Li1∗Xinyuan Guo1∗Bowen Li1∗Lanxi Xiao1Shenyu Qiao1 Jiashu Chen1Zijian Wu1Hui Zhang1Xinhuan Shu2Shixia Liu1† 1Tsinghua University2Newcastle University Ref er ence GPT -Image-1 Fidelity... | https://arxiv.org/abs/2505.18668v1 |
to generalize across different real-world applications where infographic charts are commonly used. To address this limitation, we build ChartGalaxy, a million-scale dataset of high-quality real and synthetic infographic charts to facilitate automated understanding and generation. As shown in Fig. 2, we build ChartGalax... | https://arxiv.org/abs/2505.18668v1 |
t types & 330 v ariationsT extLif est y le C hang esRemote W orkandImage 68 la y out templatesTEXTTEXTSynthetic infogr aphic char t B et t er lif es t y le40%W ors e lif es t y le22%38%N o chang e Lif est y le C hang esRemote W orkandFinal Dataset104,519 r eal infogr aphic char ts + tables 1,151,0 87 synthetic infogr a... | https://arxiv.org/abs/2505.18668v1 |
creation stage follows an inductive structuring process that extracts design patterns, such as layout templates and chart variations, from real infographic charts and then uses these patterns to programmatically create high-quality synthetic charts. It includes three steps: 1) identifying chart types and their variatio... | https://arxiv.org/abs/2505.18668v1 |
[ 52], Our World in Data [ 53], and Papers with Code [ 54]. For synthetic data, we generate 98,483 tables with Gemini-2.0-Flash following Han et al.’s method [ 20]. To facilitate downstream processing, we also complement each table with a topic ( e.g., “US election,” “NBA play-offs”) extracted by Gemini-2.0-Flash and s... | https://arxiv.org/abs/2505.18668v1 |
( e.g., scales). For example, a data table with one categorical column and two numerical columns may be mapped to a scatter plot. When multiple chart types are suitable, we prompt Gemini-2.0-Flash to select the optimal one based on data-chart compatibility [ 60]. A full list of mapping rules and prompts used for chart ... | https://arxiv.org/abs/2505.18668v1 |
training and evaluation tasks. 4 Experiments 4.1 Instruction Dataset for Infographic Chart Understanding In this experiment, we construct an instruction dataset with ChartGalaxy to enhance model capabilities on infographic chart understanding. We validate its usefulness by fine-tuning two open-source LVLMs, demonstrati... | https://arxiv.org/abs/2505.18668v1 |
ANLS for textual answers, and exact matching for multiple-choice questions. Results and analysis Tables 1 and 2 show the evaluation results on the public benchmarks and our evaluation set. After fine-tuning with ChartGalaxy, both models demonstrate improved performance gains across all question types. On the public ben... | https://arxiv.org/abs/2505.18668v1 |
overall score as the average of the high-level and low-level scores, ranging from 0 to 100. Notably, if the code fails to render the chart, both scores are set to 0. Experimental setup We benchmark 17 widely used LVLMs, including 12 proprietary ones and 5 open-source ones, as shown in Table 3. Model configurations and ... | https://arxiv.org/abs/2505.18668v1 |
its column descriptions. The key feature of this method is its ability to generate visually coherent infographic charts by reusing the layout templates of well-designed examples and leveraging powerful detection and vision-language models. To enable this capability, we first apply the detection model described in Sec. ... | https://arxiv.org/abs/2505.18668v1 |
pipeline enables the scalable creation of diverse infographic charts. By providing aligned data-chart pairs, extracted layout templates, and three representative applications, we aim to advance the development of foundation models capable of interpreting, reasoning, and generating complex infographic charts. 9 At the s... | https://arxiv.org/abs/2505.18668v1 |
, 2020, pp. 3512–3521. [16] Z. Xu, S. Du, Y . Qi, C. Xu, C. Yuan, and J. Guo, “ChartBench: A benchmark for complex visual reasoning in charts,” arXiv preprint arXiv:2312.15915 , 2023. [17] B. Tang, A. Boggust, and A. Satyanarayan, “VisText: A benchmark for semantically rich chart captioning,” in Proceedings of the Annu... | https://arxiv.org/abs/2505.18668v1 |
Liang, Z. Lu, Y . Shan, and P. Luo, “Plot2Code: A comprehensive benchmark for evaluating multi-modal large language models in code generation from scientific plots,” arXiv preprint arXiv:2405.07990 , 2024. [31] Z. Yang, Z. Zhou, S. Wang, X. Cong, X. Han, Y . Yan, Z. Liu, Z. Tan, P. Liu, D. Yu, Z. Liu, X. Shi, and M. Su... | https://arxiv.org/abs/2505.18668v1 |
Conference on Applications of Computer Vision , 2022, pp. 1697–1706. [42] “Image deduplicator (imagededup),” 2019, accessed: 2025-05-07. [Online]. Available: https://github.com/idealo/imagededup [43] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger... | https://arxiv.org/abs/2505.18668v1 |
7, pp. 3089–3103, 2024. [62] J. Chen, W. Yang, Z. Jia, L. Xiao, and S. Liu, “Dynamic color assignment for hierarchical data,” IEEE Transactions on Visualization and Computer Graphics , vol. 31, no. 1, pp. 338–348, 2025. [63] C. K. Coursaris and K. Kripintris, “Web aesthetics and usability: An empirical study of the eff... | https://arxiv.org/abs/2505.18668v1 |
no. 5, pp. 340–350, 2001. 14 NeurIPS Paper Checklist 1.Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: In this paper, we build a million-scale dataset consisting of 1,151,087 program- matically created infogra... | https://arxiv.org/abs/2505.18668v1 |
3.Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? 15 Answer: [NA] Justification: This paper does not include any theoretical results. Guidelines: • The answer NA means that the paper does not include theoretical ... | https://arxiv.org/abs/2505.18668v1 |
results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d)We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source ... | https://arxiv.org/abs/2505.18668v1 |
answer NA means that the paper does not include experiments. •The authors should answer "Yes" if the results are accompanied by error bars, confi- dence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. •The factors of variability that the error bars a... | https://arxiv.org/abs/2505.18668v1 |
•If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. 18 •Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (... | https://arxiv.org/abs/2505.18668v1 |
version of the asset is used and, if possible, include a URL. • The name of the license (e.g., CC-BY 4.0) should be included for each asset. •For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. 19 •If assets are released, the license, copyrigh... | https://arxiv.org/abs/2505.18668v1 |
you should clearly state this in the paper. •We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. •For initial submissions, do not include any information that would br... | https://arxiv.org/abs/2505.18668v1 |
The notation X×k indicates kdistinct attributes of type X. When a symbol such as ∗is specified (e.g., for the Diverging Bar Chart), it indicates that the first categorical attribute must contain exactly two distinct values. Chart Type Attribute Combinations Vertical Bar Chart C ×1 + N×1 Vertical Stacked Bar Chart C ×2 ... | https://arxiv.org/abs/2505.18668v1 |
C ×1 + N×1 Small Multiples of Pie Charts C ×2 + N×1 Small Multiples of Donut Charts C ×2 + N×1 Small Multiples of Semicircle Pie Charts C ×2 + N×1 Small Multiples of Semicircle Donut Charts C ×2 + N×1 Small Multiples of Rose Charts C ×2 + N×1 Radar Line Chart C ×1 + N×1 Radar Spline Chart C ×1 + N×1 Small Multiples of ... | https://arxiv.org/abs/2505.18668v1 |
(4 v ariations)12 01234 ABCDA dd grid lines and Y -axis (4 v ariations)T ype 2: V er tical Stack ed Bar Char t (wit h 10 v ariations)1 ABCD0.511.52.50.511.51.5Plain char t #12 ABCD11.52.50.511.51.50.5Plain char t #23 AABCD1.520.511.520.511 2 3 4 3D style4 ABCD0.511.52.50.511.51.5A dd icons ne xt t o t he labels (3 v ar... | https://arxiv.org/abs/2505.18668v1 |
ne xt t o t he labels30ABCD A dd icons t o t he end of t he barsT ype 9: Radial Gr ouped Bar Char t (wit h 6 v ariations)10AB Plain char t #120AB Plain char t #23 ABCDABCDPlain char t #340AB A dd icons ne xt t o t he labels (3 v ariations)T ype 10: Cir cular Bar Char t (wit h 4 v ariations)1 Plain char t #1212345678910... | https://arxiv.org/abs/2505.18668v1 |
2 Plain char t #22A134B 3 Plain char t #3 2A134B4 A dd icons ne xt t o t he labels (3 v ariations)2AB134 A dd icons ne xt t o t he labels4321DCBA Plain char t #24321ABCD Plain char t #14321DCBAT ype 20: V er tical Dot Bar Char t (wit h 3 v ariations)123 T ype 21: Horiz ontal Dot Bar Char t (wit h 3 v ariations) T ype 2... | https://arxiv.org/abs/2505.18668v1 |
Range Ar ea Char t (wit h 2 v ariations)1 00’01’02’03’14 Plain char t #12 00’01’02’03’14 A dd icons on t he endpoint s of t he linesT ype 37 : Stack ed Ar ea Char t (wit h 3 v ariations)1 00’01’02’03’Plain char t #12 00’01’02’03’Plain char t #23 00’01’02’03’ABA dd icons on t he ar easT ype 35: La y er ed Spline Ar ea C... | https://arxiv.org/abs/2505.18668v1 |
ariations)140.0% 60.0% Plain char t #1240.0% 60.0% A dd icons on t he edge of pies3 A dd icons in t he cent er of pies40.0%60.0%4 Hand-dr awn style (3 v ariations) 53 . 6%46 .4%T ype 51: Semicir cle Donut Char t (wit h 8 v ariations)2 A dd icons on t he edge of sect ors #120.0%30.0%40.0%10.0%4 A dd icons on t he edge o... | https://arxiv.org/abs/2505.18668v1 |
h 1 v ariation)1 Plain char t #1T ype 63: Scatt erplot (wit h 4 v ariations)1 0123456BACDE43210F Plain char t #12 -100 11-1Plain char t #23 0123456ACDE43210BF R eplace scatt er point s wit h icons (2 v ariations) T ype 67 : W affle Char t (wit h 4 v ariations)1AA1A1B2 Plain char t #12AA1A1B2 Plain char t #23A1B2C1Plain... | https://arxiv.org/abs/2505.18668v1 |
layout templates (Part 2). 43 General AdmissionVIP 040801201602002402802024-07-15 10:0085.5250.02024-07-15 12:0092.8275.32024-07-15 14:001 10.0310.52024-07-15 16:00105.3295.82024-07-15 18:0098.0280.02024-07-15 20:0090.5265.52024-07-15 22:0080.0240.02024-07-16 00:0075.3225.8VIP T ickets Command HigherResale Prices Throu... | https://arxiv.org/abs/2505.18668v1 |
chart examples (Part 1). 44 22.345.868.184.518.538.259.478.215.232.151.772.975.192.897.598.921.436.554.2 20242028203220362040GermanyFranceUKNorwaySpainCountryElectric V ehicle AdoptionRates on the RiseElectric vehicle adoption rates are projected torise significantly across European countries by2040. Soviet Union270000... | https://arxiv.org/abs/2505.18668v1 |
Years? Answer with exactly ’Yes’ or ’No’. Visual-Element-Based Reasoning Data Identification (DI)Identify data values associated with specific visual elements (e.g., icons, symbols) in the chart.What is the WaterConsumption for in the Industrial group? Please provide a numerical answer. Data Comparison (DC)Compare data... | https://arxiv.org/abs/2505.18668v1 |
complete outputs under our task setting—for example, Phi-4—due to their limited capacity or inability to handle long sequences. We use greedy decoding (temperature τ= 0) across all models to ensure deterministic outputs. To maximize the chance of obtaining complete and executable code, we configure each model to genera... | https://arxiv.org/abs/2505.18668v1 |
n−1do 5: CostMatrix [i][j]←LeafCost (gt_leafs [i], pr_leafs [j]) 6: end for 7:end for 8:(rows, cols )←HungarianAlgorithm (CostMatrix ) ▷Returns optimal row-column pairs 9:foreach pair (i, j)in(rows, cols )do 10: ifCostMatrix [i][j]≤1AND gt_matched [i] =−1AND pr_matched [j] =−1then 11: gt_matched [i]←j 12: pr_matched [j... | https://arxiv.org/abs/2505.18668v1 |
solutions. Among the five open-source models, Qwen2.5-VL-32B has the longest average code length despite achieving the lowest overall score, while the remaining models exhibit comparable average lengths. These findings highlight the distinct coding styles of different LVLMs when generating extended code sequences. Tabl... | https://arxiv.org/abs/2505.18668v1 |
the model is provided the ground-truth chart, previously generated code from direct prompting, and the rendered chart, and is instructed to revise the given code. Detailed prompts are available in our code repository1. The results, summarized in Table 10, show that the SelfReflection method consistently achieves the be... | https://arxiv.org/abs/2505.18668v1 |
compensated with 30 USD for their participation. Fig. 28 illustrates the user study interface using a specific example. Figs. 29-30 present all 30 triplets of infographic charts: one reference, two infographic charts to be rated that are generated by GPT-Image- 1 and our method, respectively. Please rate the infographi... | https://arxiv.org/abs/2505.18668v1 |
embedding meaningful icons into the titles, and by exploring less conventional chart types beyond basic bar and line charts. In contrast, GPT-Image-1 tends to generate conventional titles and favors basic chart types, leading to lower perceived creativity. E Prompts for Data Processing E.1 Instruction Dataset for Infog... | https://arxiv.org/abs/2505.18668v1 |
the headline inflation and core inflation?” “In the years in which the red line was higher than the blue line, which year had the smallest difference between the red and green lines?” “Which country had the highest increase in the number of cases between Jun and Jul?” “Which country had the most significant drop in its... | https://arxiv.org/abs/2505.18668v1 |
and direct; avoid contradict- ing the table data. # INSTRUCTIONS You are an AI that generates concise and specific hypothetical questions based on chart images. Your task is to analyze the chart and generate a short, data-driven hypothetical question that explores future trends, impacts, or extrapolations based on the ... | https://arxiv.org/abs/2505.18668v1 |
# Technical Requirements •Charting library : Use D3.js to implement the chart. Write the code to be clean, modular, and easy to understand and modify. •Single file output : Provide one standalone HTML file that includes everything needed to render the chart. •Chart fidelity : Replicate all visual elements—shapes, color... | https://arxiv.org/abs/2505.18668v1 |
text colors, etc.)? Minor differences due to rendering or anti-aliasing can be tolerated if the overall color scheme is preserved. 6.Validity (20 points): •Is the AI-generated chart clear, readable, and free of overlapping or occluded elements? •Are fundamental charting conventions followed? For example: Are axis ticks... | https://arxiv.org/abs/2505.18668v1 |
the rights of original creators. Our project has been approved by our institution’s internal ethics review. All human subjects involved in our user study have signed the user consent forms and have been provided with fair compensation in accordance with the local minimum wage standards. Societal impacts Infographic cha... | https://arxiv.org/abs/2505.18668v1 |
arXiv:2505.18673v1 [cs.CL] 24 May 2025Cross-Lingual Pitfalls: Automatic Probing Cross-Lingual Weakness of Multilingual Large Language Models Zixiang Xu1*, Yanbo Wang1*, Yue Huang2*, Xiuying Chen1†,Jieyu Zhao3,Meng Jiang2,Xiangliang Zhang2† 1MBZUAI,2University of Notre Dame,3University of Southern California Corresponde... | https://arxiv.org/abs/2505.18673v1 |
English but incorrectly in at least one other language. This definition requires the model to provide the correct answer in English, as failure across all languages English Questions Preprocessing 0.70.70.7 Bilingual Pairs 0.70.7 LLM-Based Simulation 0.7 Bilingual Accuracy Problem0.7 Simulation Score Rank byTarget Lang... | https://arxiv.org/abs/2505.18673v1 |
weaknesses; and 2) fine- tuning LLMs on one language improves perfor- mance more significantly in linguistically similar languages. These results highlight that linguistic relationships strongly influence cross-lingual per- formance. In summary, our contributions are: 1) We present an efficient, precise methodology for... | https://arxiv.org/abs/2505.18673v1 |
a translation module that strictly translates the inserted perturbation without modifying other parts of the question. This results in the perturbed target- language question: qT′=⊕(qT, δqT). We optimize the perturbation to minimize the model’s accuracy in the target language while main- taining near-perfect performanc... | https://arxiv.org/abs/2505.18673v1 |
pairs where the model maintains strong perfor- mance in English ( ¯βE′≈1) but exhibits significant degradation in the target language ( ¯βT′). 2.4 Beam Search with Optimization Strategies Since beam search is an effective heuristic for ex- ploring a constrained search space, we employ it to solve the objective function... | https://arxiv.org/abs/2505.18673v1 |
to higher costs. 3 Experiment 3.1 Experiment Overview In this section, we conduct a series of experi- ments to evaluate the effectiveness of our proposed method as well as to explore the cross-lingual weak- nesses of multilingual models. Overall, we mainly aim to address the following questions: •RQ1: How effective and... | https://arxiv.org/abs/2505.18673v1 |
ure 5 and Figure 10 reveals that: Qwen2.5-72B and Gemma-2-27B show minor accuracy improve- ments after being removed from the simulation models, GPT-4o—despite being a top-tier multilin- gual model—suffers a sharp 58% accuracy drop. Our method enables the cost-effective identifi- cation of cross-lingual weaknesses. We ... | https://arxiv.org/abs/2505.18673v1 |
stantial and relatively consistent accuracy declines. In contrast, when these Asian seed pairs are ex- panded into European languages, the accuracy drops are considerably smaller and more variable. A similar trend is observed within the European language family: pairs expanded from French, Ger- man, or Spanish into oth... | https://arxiv.org/abs/2505.18673v1 |
a closer linguistic relationship between the two languages. As shown in Figure 7, it reveals a clear pat- tern: lower RAS values Dx,yare predominantly ob- served for language pairs ( x, y) with linguistic and cultural proximities. This observation strongly sup- ports our hypothesis that languages with closer lin- guist... | https://arxiv.org/abs/2505.18673v1 |
embeddings generated by Llama- 3.1-8B for seven English–target language question pairs. Yoruba exhibited significantly more errors in the Science & Technology domain compared to most languages. Conversely, higher-resource languages like Chinese, Spanish, and German demonstrated stronger performance in this area. Intere... | https://arxiv.org/abs/2505.18673v1 |
cross-lingual pitfalls encoun- tered by various models across different languages. They showcase specific failure modes, such as mis- interpretation of nuanced phrasing, incorrect entity mapping, or breakdowns in reasoning when faced with linguistic structures that differ significantly from English. These case studies ... | https://arxiv.org/abs/2505.18673v1 |
potential for bias within LLMs, particularly across different lan- guages and cultural contexts. Our language se- lection was carefully considered to ensure diver- sity, encompassing both high-resource and lower- resource languages. All generated content and model outputs were scrutinized for potential bi- ases. No per... | https://arxiv.org/abs/2505.18673v1 |
values. Pro- ceedings of the International Conference on Learning Representations (ICLR) . Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Stein- hardt. 2021b. Measuring massive multitask language understanding. Proceedings of the International Con- ference on Learning Represe... | https://arxiv.org/abs/2505.18673v1 |
benchmark. arXiv preprint arXiv:2405.12209 . Wei Liu, Zhongyu Niu, Lang Gao, Zhiying Deng, Jun Wang, Haozhao Wang, and Ruixuan Li. 2025. Ad- versarial cooperative rationalization: The risk of spu- rious correlations in even clean datasets. Preprint , arXiv:2505.02118. Wei Liu, Chenxi Wang, YiFei Wang, Zihao Xie, Rennai... | https://arxiv.org/abs/2505.18673v1 |
natural language in- ference via ensemble adversarial training. arXiv preprint arXiv:2004.07790 . Lichao Sun, Yue Huang, Haoran Wang, Siyuan Wu, Qihui Zhang, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, et al. 2024. Trustllm: Trustworthiness in large language models. arXiv preprint arXiv:2401.05561 , 3.... | https://arxiv.org/abs/2505.18673v1 |
Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, and Others. 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 . Binwei Yao, Ming Jiang, Diyi Yang, and Junjie Hu. 2023. Benchmarking llm-based machine translation on cultural awareness. arXiv preprint arXiv:2305.14328 . Jiayi Ye, Yanbo Wang, Yue Huang, Do... | https://arxiv.org/abs/2505.18673v1 |
own dataset. A.2 Cross-lingual Capablity of LLMs. The cross-lingual capabilities of LLMs have be- come a central focus in NLP research. Multi- task finetuning (MTF) has proven effective for en- hancing cross-lingual generalization, as shown by Muennighoff et al. (2022), where finetuning mul- tilingual models like BLOOM... | https://arxiv.org/abs/2505.18673v1 |
of 0.7 to en- courage more diverse and creative responses. In the translation, semantic checking, and simulation tasks, the temperature was reduced to 0.001 to ensure stability in the responses. The maximum output length for these tasks was capped at 1,024 tokens. During beam search, we initialized the pro- cess with W... | https://arxiv.org/abs/2505.18673v1 |
Leveraging the English-Chinese and English-French question pairs generated in our dataset, we performed SFT and DPO on several Large Language Models: Llama- 3.1-8B, Qwen2.5-7B, Gemma-2-9B, and Phi-3.5- Mini. For each model, we conducted separate fine- tuning runs using both the Chinese and French datasets. To ensure co... | https://arxiv.org/abs/2505.18673v1 |
4.55% 19.01% 21.90% 12.40% Swahili 47.96% 8.87% 3.12% 30.94% 4.32% 4.80% Ukrainian 39.01% 10.53% 4.33% 34.98% 6.81% 4.33% Yoruba 53.01% 6.52% 7.87% 21.03% 4.67% 6.89% Zulu 47.40% 13.36% 0.98% 25.60% 8.44% 4.22% Overall Average 45.36% 12.20% 6.63% 23.40% 7.15% 5.26% Table 5: Performance comparison of Gemma-2-9B and Qwen... | https://arxiv.org/abs/2505.18673v1 |
on English-Yoruba pairs in our candidate list. Figure 25: Performance of LLMs on English-Zulu pairs in our candidate list. Target Language: Korean English Question: Many cities around the world, like London, are known for their rich cultural scenes and historic landmarks that attract millions of visitors each year.Kiwi... | https://arxiv.org/abs/2505.18673v1 |
German Question: Veranschaulichend, wie die Fo rm der Funktion folgt, sind lange, schlanke Proteinstränge, die welches Gewebe bilden, essentiell für das Zusammenziehen und Entspannen?Das Design verschiedener biologischer Strukturen spie gelt o ft ihre spezifischen Rollen wider; zum Beispiel kann die Art und Weise, wie ... | https://arxiv.org/abs/2505.18673v1 |
response:La respuesta correcta es: viernes Figure 31: Case study: Llama-3.1-8B’s responses to English-Spanish question pairs. Target Language: Japanese English Question: Prophase is preceded by a preprophase stage in what type of cells? In many organisms, including certain plants, the development of reproductive struct... | https://arxiv.org/abs/2505.18673v1 |
Veuillez respecter strictement les règles suivantes : - La traduction des réponses et des choix doit refléter fidèlement le sens original, sans aucune altération, omission ou ajout. - Toutes les phrases comportant un point d’interrogation doivent rester sous forme de question après traduction, sans changer le ton ou la... | https://arxiv.org/abs/2505.18673v1 |
completamente el significado original, sin desviaciones ni adiciones. - Todas las oraciones que contengan un signo de interrogación deben mantener la forma de pregunta en la traducción, sin cambiar el tono ni la estructura de la oración. - El contenido traducido debe ajustarse a las costumbres del idioma español, expre... | https://arxiv.org/abs/2505.18673v1 |
arXiv:2505.18675v1 [cs.CV] 24 May 2025Can MLLMs Guide Me Home? A Benchmark Study on Fine-Grained Visual Reasoning from Transit Maps Sicheng Feng1,2,†, Song Wang3,2,†, Shuyi Ouyang3,2, Lingdong Kong2, Zikai Song4,2, Jianke Zhu3, Huan Wang1,∗, Xinchao Wang2 1Westlake University2National University of Singapore3Zhejiang U... | https://arxiv.org/abs/2505.18675v1 |
of the route is separated by two '-'. ReasonMap Google Map Gaode MapDifficulty Label Map Difficulty Refer ence Route- Route Name - Departure Stop - Arrival Stop - Via Stops - Number of Via Stops Route Info Section Hard - 30 Cities - 1000+ Questions Resolution: 9921 9908 Question Difficulty Medium Figure 1: Overview of ... | https://arxiv.org/abs/2505.18675v1 |
1,003 601 ×331 ✓ ✗ ✗ VisualPuzzles [32] 2025 1,168 767 ×464 ✗ ✗ ✗ VGRP-Bench [33] 2025 20(×5) 790×790 ✗ ✓ ✗ R-Bench [34] 2025 665 629 ×348 ✗ ✗ ✓ (2) V∗Bench [37] 2023 191 2 ,246×1,582 ✗ ✗ ✗ REASON MAP 2025 1,008(×2) 5,839×5,449 ✓ ✓ ✓ (4) Our main contributions are summarized as follows: •We develop an extensible, semi-... | https://arxiv.org/abs/2505.18675v1 |
a structured evaluation suite for map- related reasoning, and GeoNav [ 52] investigates geospatial navigation using LLMs. Most existing methods [ 26,35,55] depend on external tools (e.g., map services or APIs) to complete spatial tasks, which often bypasses the need for genuine visual reasoning. However, spatial reason... | https://arxiv.org/abs/2505.18675v1 |
long question based on predefined question templates and two stops (Figure 2). The short question has only one fixed template, while the long question is randomly assigned one of two available templates during generation. Additionally, the two long question templates differ in focus: one asks for the number of via stop... | https://arxiv.org/abs/2505.18675v1 |
for Evaluation. We first parse the model-generated answers according to the required format. Answers that do not comply with the specified format or cannot be parsed due to model hallucination [ 57] are marked as invalid. Invalid responses are excluded from subsequent evaluations, with accuracy and map score set to zer... | https://arxiv.org/abs/2505.18675v1 |
at 10for the full route. Specific Via Stop Evaluation. For long questions that require explicit enumeration of intermediate stops, we compute via_stop_score using a combination of two factors: the number of correctly matched via stops, and the intersection-over-union (IoU) between via stop sets of the answer and refere... | https://arxiv.org/abs/2505.18675v1 |
strategies when handling high-resolution visual inputs in Appendix D. Difficulty-Aware Weighting. To better reflect the varying complexity of different samples, we adopt a difficulty-aware weighting strategy based on the combination of question difficulty and map difficulty. Specifically, each difficulty pair is assign... | https://arxiv.org/abs/2505.18675v1 |
closed-source reasoning models outperform their base variants. One possible explanation lies in the broader knowledge coverage and better visual integration observed in these models [19, 47, 49]. We further analyze the effect of model size by examining performance within the same architecture series. Qwen2.5-VL and Int... | https://arxiv.org/abs/2505.18675v1 |
case analysis of various MLLMs using REASON MAP. For reasoning models, the reasoning process is explicitly marked with <think> and</think> tags. We highlight error contents in the answers with red and categorize them accordingly. 5.2.2 Performance of MLLMs without Visual Input To further investigate the reliance of MLL... | https://arxiv.org/abs/2505.18675v1 |
et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479 , 2025. [4]Yangliu Hu, Zikai Song, Na Feng, Yawei Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Sf2t: Self-supervised fragment finetuning of video-llms for fine-grained understand... | https://arxiv.org/abs/2505.18675v1 |
Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. In NeurIPS , 2024. [23] Fatemeh Shiri, Xiao-Yu Guo, Mona Golestan Far, Xin Yu, Gholamreza Haffari, and Yuan-Fang Li. An empirical analysis on spatial re... | https://arxiv.org/abs/2505.18675v1 |
study from the reliability, data, and metric perspectives. arXiv preprint arXiv:2501.04003 , 2025. [37] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In CVPR , 2024. [38] Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. Efficient reasoning models: A survey. arXiv p... | https://arxiv.org/abs/2505.18675v1 |
Xu, and Yong Li. Perceive, reflect, and plan: Designing llm agent for goal-directed city navigation without instructions. arXiv preprint arXiv:2408.04168 , 2024. 12 [57] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A sur... | https://arxiv.org/abs/2505.18675v1 |
15 B Evaluation Details 15 B.1 Correctness and Quality Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 15 B.2 Details about Difficulty-Aware Weighting. . . . . . . . . . . . . . . . . . . . . . . 16 C Case Analysis 16 D Further Discussions 17 D.1 High-Resolution Image Preprocessing. . . . . . . . . . . . .... | https://arxiv.org/abs/2505.18675v1 |
Stops: x Long Question Template 2 According to the subway map, how do I get from [Stop 1] to[Stop 2] ? Provide only one optimal route, including all the via stops. The format should be strictly followed: Route Name: Line x Departure Stop: xx Station Arrival Stop: xx Station Via Stops: xx Station, xx Station, xx Station... | https://arxiv.org/abs/2505.18675v1 |
the text-only variant makes an early misclassification, placing both stations on the Blue Line (Azul) and constructing a plausible but entirely incorrect sequence of intermediate stops. Although the final answer format appears coherent, the underlying logic is flawed due to the initial error in line recognition. This f... | https://arxiv.org/abs/2505.18675v1 |
is likely needed.So the route could be: Augustins(Line 12) → Transfer to Line 18 → Poterie. Or maybe there's a direct connection?No, Augustinsis on Line 12, and Poterieis on Line 18. So we must go from Augustinsto a transfer station—maybe Place de Neuve? ...Steps: Augustins(Line 12) → Plainpalais(transfer) → Poterie(Li... | https://arxiv.org/abs/2505.18675v1 |
goes from ... Wait, no, actually, from Ameixoeira, the next stations towards the city center would be … No, wait, Picoasis on the Blue Line between Entre Campos and Marquêsde Pombal? Wait, maybe I should list the stations in order. Let's see: ... ... (continuous self-correction)Wait, maybe I got the direction wrong … L... | https://arxiv.org/abs/2505.18675v1 |
and generalizable multimodal systems. 19 A: <think> ...First, let's locate both stations on the map. York University Station is on Line 1 (the yellow line), which runs north-south. Chester Station is also on Line 1, south of Castle Frank and north of Pape.... (continuous self-correction)After re-evaluating, the correct... | https://arxiv.org/abs/2505.18675v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.