text
string
source
string
thresholds. Numbers along the red line indicate the average number of tokens decoded at each step. The three dashed lines represent the accuracy of the baseline method when selecting the top 2, 4, or 8 tokens per step. (b) The number of inference steps required under varying confidence thresholds. (c) A comparison betw...
https://arxiv.org/abs/2505.22618v1
cache locality for both efficiency and accuracy. Effect of Cache Block Size Figure 4 analyzes the influence of the cache block size hyperparameter. We observe that smaller block sizes tend to maximize accuracy but incur overhead due to frequent cache updates. In contrast, larger block sizes may diminish accuracy owing ...
https://arxiv.org/abs/2505.22618v1
using masked denoising, while LLaDA and Dream [ 36] demonstrating competitive performance with autoregressive baselines like LLaMA3 [9] through recursive token prediction across diffusion timesteps. 5.2. LLM Acceleration Key-Value Cache. Key-Value (KV) Cache is a fundamental optimization technique in modern large langu...
https://arxiv.org/abs/2505.22618v1
๐‘—=1๐‘(๐‘‹๐‘–๐‘—ฬธ=๐‘ฅ๐‘–๐‘—|๐ธ) = 1โˆ’๐‘›โˆ‘๏ธ ๐‘—=1๐œ–โ€ฒ ๐‘—. Since ๐œ–โ€ฒ ๐‘—< ๐œ–for all ๐‘—, we haveโˆ‘๏ธ€๐‘› ๐‘—=1๐œ–โ€ฒ ๐‘—< ๐‘›๐œ– . So, ๐‘(๐‘ฅ*|๐ธ)>1โˆ’๐‘›๐œ–. Now consider any ๐‘ง= (๐‘ง1, . . . , ๐‘ง ๐‘›)such that ๐‘งฬธ=๐‘ฅ*. This means there is at least one index ๐‘˜such that ๐‘ง๐‘˜ฬธ=๐‘ฅ๐‘–๐‘˜. The event{๐‘‹=๐‘ง}is a sub-event of{๐‘‹๐‘–๐‘˜=๐‘ง๐‘˜}. So, ๐‘(๐‘ง|๐ธ)โ‰ค๐‘๐‘˜...
https://arxiv.org/abs/2505.22618v1
๐‘๐œƒ(๐‘‹๐‘–1, . . . , ๐‘‹ ๐‘–๐‘›|๐ธ):The theorem and proof rely on ๐‘๐œƒ(๐‘‹๐‘–1, . . . , ๐‘‹ ๐‘–๐‘›|๐ธ) being a well-defined joint probability mass function from which the marginals ๐‘๐œƒ(๐‘‹๐‘–๐‘—|๐ธ)are consistently derived. This implies that the joint PMF is coherent and its definition does not depend on a specific factorization ...
https://arxiv.org/abs/2505.22618v1
So, the value is 3The robe takes 2 bolts of blue fiber. It also takes half that much white fiber, so it takes 2/2 = 1 bolt of white fiber. In total, the robe takes 2 + 1 = 3 bolts of fiber. So, the value is 3The robe takes 2 bolts of blue fiber. It also takes half that much white fiber, so it takes 2/2 = 1 bolt of whit...
https://arxiv.org/abs/2505.22618v1
Thomas Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models. Advances in Neural Information Processing Systems , 35:28266โ€“28279, 2022. [4]Zixiang Chen, Huizhuo Yuan, Yongqian Li, Yiwen Kou, Junkai Zhang, and Quanquan Gu. Fast sampling via de-randomization for dis...
https://arxiv.org/abs/2505.22618v1
distributions of clean data. arXiv preprint arXiv:2406.03736 , 2024. [23] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation, 2021. [24] Machel Reid, Vincent J. Hellendoorn, and Graham Neubig. Diffuser: Discrete diffusio...
https://arxiv.org/abs/2505.22618v1
arXiv:2505.22627v1 [cs.CL] 28 May 2025CHAIN -OF-TALKERS (COTALK ): Fast Human Annotation of Dense Image Captions Yijun Shen1โˆ—Delong Chen2โˆ—Fan Liu1,โ€ Xingyu Wang1 Chuanyi Zhang1Liang Yao1Yuhui Zheng3,โ€  1Hohai University2HKUST 3Nanjing University of Information Science and Technology fanliu@hhu.edu.cn ,zhengyh@vip.126.com...
https://arxiv.org/abs/2505.22627v1
al., 2024) 1 cross-annotation overlap is 71.36% as measured by Sentence-BERT (Reimers and Gurevych, 2019)). Our second insight is that speech-based annota- tion significantly surpasses typing in speed and effi- ciency. This is supported by prior research demon- strating that the average throughput for spoken words (161...
https://arxiv.org/abs/2505.22627v1
captioning, system- atically defining key criteria for high-quality an- notations (Figure 2). The framework introduces a semantic space Sโˆˆ[0,1]n, where each dimension corresponds to a semantic unit ฯ‰iinโ„ฆ ={ฯ‰i}n iโˆ’1, representing the probability of that unit appearing in the image. The caption generation process is mode...
https://arxiv.org/abs/2505.22627v1
denoted as eYk. After all n annotators have completed their anno- tations, the LLM merges all the annotations from the different annotators together. eYPar=eYn ฯƒ=ฯƒ(eY1, . . . ,eYn), (3) then the final semantic unit of the parallel process can be determined as YPar=Yn ฯƒ=h(eYn ฯƒ) The total time cost for the parallel anno...
https://arxiv.org/abs/2505.22627v1
to the as- sumptions: Theorem 1 (CoTalk Enhances Annotation Quality) : As defined in Equation 1, high quality annotation is characterized by high information suf- ficiency, minimal redundancy, and strong human comprehensibility. Information Sufficiency : Information suffi- ciency is defined as the completeness of seman...
https://arxiv.org/abs/2505.22627v1
ing on Theorems 1, we establish that CoTalk yields the highest annotation quality. We now compare the efficiency of CoTalk, parallel annotation, and single-round annotation by evaluating the time re- quired to achieve equivalent semantic unit cover- age. Since single-round annotation provides signif- icantly lower info...
https://arxiv.org/abs/2505.22627v1
which means thatvreading > v listening . Therefore, we can achieve thatvreading in> vlistening in. From the above results, we conclude that in CoTalk, using multimodal input and output can save more time compared to a single modality ap- proach. Specifically, for annotatorsโ€™ output, talking should be adopted due to its...
https://arxiv.org/abs/2505.22627v1
As shown in Table 3, the model trained with CoTalk annotations achieves the highest average retrieval score (41.13%), outper- forming the model trained with parallel annotations (40.52%). More details are in H.1. With intrinsic and extrinsic metrics, CoTalk an- notation consistently outperforms parallel annota- tion in...
https://arxiv.org/abs/2505.22627v1
( e.g., urban or industrial). These findings suggest that relying solely on speech (talking) or text for both input and output is suboptimal. Instead, a hybrid approach using text for input and speech for output can achieve both higher annotation quality and greater efficiency. 4.4 Further Analysis 4.4.1 Consistency be...
https://arxiv.org/abs/2505.22627v1
proven effective in enhanc- ing image caption quality and efficiency, several limitations remain worth discussing. Assumptions on Human Ability. Our compari- son of information sufficiency, human comprehen- sibility, and efficiency between sequential annota- tion and CoTalk assumes consistent annotator abil- ities. How...
https://arxiv.org/abs/2505.22627v1
understanding using captions with grounded segmentation. ArXiv , abs/2412.09754. Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi, Jathushan Rajasegaran, Hanoona Abdul Rasheed, Junke Wang, Marco Monteiro, Hu Xu, Shiyu Dong, Nikhila Ravi, Daniel Li, Piotr Dollโ€™ar, and ...
https://arxiv.org/abs/2505.22627v1
Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. 2023. Monkey: Image resolution and text label are important things for large multi-modal models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 26753โ€“ 26763. Zhenshi Li, Dilxat Muhtar, Feng Gu, Xueliang Zhang, Pengfeng Xiao, G...
https://arxiv.org/abs/2505.22627v1
Unlocking the long-text capability of clip. In European Confer- ence on Computer Vision . Appendix A Related Work A growing trend in image captioning research was the emphasis on generating more comprehensive captions (Cho et al., 2025; Bolya et al., 2025; Shab- bir et al., 2025; Hua et al., 2024; Athar et al., 2024; S...
https://arxiv.org/abs/2505.22627v1
reading comprehension speed vtext in(Ruan et al., 2016; Brysbaert, 2019). Given n= 2, the CoTalk annotation time is: T2 CoTalk = 2ยทTeX in+TeY1 CoTalk in+2X i=1TeYi CoTalkout (11) 11 Table 5: The simulation results of sequence and parallel on large model ModelRSICD RSITMD UCM-Captions Image to Text AverageText to Image ...
https://arxiv.org/abs/2505.22627v1
dataset benchmark evaluation. ModelRSICD RSITMD UCM-Captions Image to Text Text to Image AverageImage to Text Text to Image AverageImage to Text Text to Image Average R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 R@1 R@5 R@10 Zero-shot 9.70 23.88 38.15 5.13 19.11 29.85 20.97 13.50 33.63 45.58 10.52 3...
https://arxiv.org/abs/2505.22627v1
tal facts. Participants answer these questions under both reading and listening conditions. The prompts, detailed in Table 10, are designed to generate high- quality questions with varying levels of difficulty, gradually incorporating finer-grained image con- tent to ensure scientific rigor and experimental va- lidity....
https://arxiv.org/abs/2505.22627v1
annotation text, please design questions according to the following rules: Guidelines : โ€ขRule 1: Generate a total of 5 questions. โ€ขRule 2: The five questions should cover the beginning, middle, and later parts of the annotation text. โ€ขRule 3: Design the questions in the order of the text and number them sequentially (Q...
https://arxiv.org/abs/2505.22627v1
Semantic Units: Table 12: Prompt for Deriving the minimal Semantic Units. 16 Figure 8: The first-person annotation interface. Figure 9: The Suquential Subsequent Annotation Interface. 17 Guidelines for Image Annotation First-Person: System Message: Input : You will be provided with an image. Your task is to generate a ...
https://arxiv.org/abs/2505.22627v1
cor- rect errors. They contribute additional information through voice input, guided by a comprehensive understanding of both the image and the existing annotations.Details are shown in Table 13 (down). H Detailed Experimental Process H.1 Extrinsic Metric We evaluate the practicality of each annotation method using ext...
https://arxiv.org/abs/2505.22627v1
Stochastic Chameleons : Irrelevant Context Hallucinations Reveal Class-Based (Mis)Generalization in LLMs Ziling Cheng1,2*Meng Cao1,2* Marc-Antoine Rondeau1Jackie Chi Kit Cheung1,2,3 1Mila โ€“ Quebec Artificial Intelligence Institute 2McGill University3Canada CIFAR AI Chair {ziling.cheng, meng.cao}@mail.mcgill.ca, {ma.ron...
https://arxiv.org/abs/2505.22630v1
matching from pre-training data, allowing us to better isolate prediction shifts driven by added context. Through qualitative analysis of irrelevant con- text hallucinations, as demonstrated in Figure 1, we hypothesize that these errors exhibit struc- tured regularities. We posit that LLMs exhibit a structured but flaw...
https://arxiv.org/abs/2505.22630v1
irrelevant contexts, distin- guishing generalization from memorization. โ€ขWe provide empirical evidence that LLMs ex- hibit class-based (mis)generalization, demon- strating sensitivity to abstract class structures beyond statistical co-occurrences. โ€ขWe uncover the internal computational mecha- nisms of class-based gener...
https://arxiv.org/abs/2505.22630v1
other (Jin et al., 2024; Xu et al., 2024a; Su et al., 2024; Yuan et al., 2024; Marjanovic et al., 2024; Neeman et al., 2023; Chen et al., 2022; Longpre et al., 2021). In contrast, we study irrelevant context hallucinations, where un- related and non-contradicting context still shapes predictions, competing with paramet...
https://arxiv.org/abs/2505.22630v1
4 Dataset and Experimental Design Models & Datasets We evaluate three pretrained LM families โ€” Llama-3 (8B, 70B) (AI@Meta, 2024), Mistral v0.3 (7B) (Jiang et al., 2023), and Pythia (6.9B-deduped, 12B-deduped) (Biderman et al., 2023) โ€” using their base versions to assess raw model behavior. We use the ParaRel dataset (E...
https://arxiv.org/abs/2505.22630v1
=|AC+Qฬธ=AQ| #datapoints. For Llama-3, 38.3% of re- sponses changed after adding irrelevant context, while for Mistral, nearly half of the datapoints (48.0%) experience a shift in predictions (Table 8 in Appendix). We further examine the cases under the C+ Qcondition based on the composition of ( Atop-3 C+Q) (Table 2). ...
https://arxiv.org/abs/2505.22630v1
expected classes, the mean PMI is ap- proximately 4, suggesting a strong association be- tween contexts and their corresponding candidates. To formally assess statistical dependence, we per- form a one-sample t-test against the null hypothesis Query TypeCtx. TypeContext Demonstration + Query and Answer LanguagePerson/ ...
https://arxiv.org/abs/2505.22630v1
tional results are provided in Appendix F. Figure 3 reveals a hierarchical class-to-instance process in answer generation. Early layers prioritize class to- ken logits (solid) like โ€œlanguagesโ€, suggesting that the model first constructs abstract class representa- tions. Around the middle layers, candidate answer logits...
https://arxiv.org/abs/2505.22630v1
prediction is query-dominant, context-based candidates remain actively computed across layers. (ii) critical tran- sition (Layers 17โ€“24) : the decisive competition between query- and context-based candidates oc- curs primarily in this range, determining which candidate is promoted. 6.2 Activation Patching Method To und...
https://arxiv.org/abs/2505.22630v1
0.78 0.81 0.78 0.49 0.44 0.47 0.25 0.25 0.21 0.06 0.00 0.00 0.000.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00(a) Context circuit in context-dominant case. L0L1L2L3L4L5L6L7L8L9L10L11L12L13L14L15L16...
https://arxiv.org/abs/2505.22630v1
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 (c) Context circuit in query-dominant case. L0L1L2L3L4L5L6L7L8L9L10L11L12L13L14L15L16L17L18L19L20L21L22L23L24L25L26L27L28L29L30L31FLNE_outQ_REL 6*Q_SUBJ 5Q_REL 4C_OBJ 3C_REL 2C_SUBJ 1C_REL 0 0.00...
https://arxiv.org/abs/2505.22630v1
increases after layer 24 in context circuit and after layer 16 in query circuit. Both circuits exist across context- and query- dominant cases, but their relative strength deter- mines the final prediction. In context-dominant cases, the context circuit wins, with a larger log- probability difference (max 2.28) compare...
https://arxiv.org/abs/2505.22630v1
Pythia 12B and LLaMA-3 70B. Experimental details are provided in Appendix I.1. Contrary to the hypothesis that scale might resolve this issue, results in Table 7 show that class-based generalization persists with similar frequency as in 7B/8B models (Sec. 5.1). Table 6 provides quali- tative examples. The statistical t...
https://arxiv.org/abs/2505.22630v1
Canada CIFAR AI Chair program. We acknowledge the material support of NVIDIA for providing computational resources. References Vaibhav Adlakha, Parishad BehnamGhader, Xing Han Lu, Nicholas Meade, and Siva Reddy. 2024. Eval- uating correctness and faithfulness of instruction- following models for question answering. Tra...
https://arxiv.org/abs/2505.22630v1
and James R. Glass. 2024. Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps. In Proceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing , pages 1419โ€“1436, Miami, Florida, USA. Association for Computational Linguistics. F...
https://arxiv.org/abs/2505.22630v1
Clara Fan- njiang, and David Sussillo. 2018. Hallucinations in neural machine translation. Daliang Li, Ankit Singh Rawat, Manzil Zaheer, Xin Wang, Michal Lukasik, Andreas Veit, Felix Yu, and Sanjiv Kumar. 2023. Large language models with controllable working memory. In Findings of the As- sociation for Computational Li...
https://arxiv.org/abs/2505.22630v1
Alexander H. Miller, and Sebastian Riedel. 2020. How context affects lan- guage modelsโ€™ factual predictions. In Automated Knowledge Base Construction . Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Schรคrli, and Denny Zhou. 2023. Large language models can be easily distracted by...
https://arxiv.org/abs/2505.22630v1
and resolving knowledge conflicts through adaptive decoding with contextual information-entropy con- straint. In Findings of the Association for Compu- tational Linguistics: ACL 2024 , pages 3903โ€“3922, Bangkok, Thailand. Association for Computational Linguistics. Yuhang Zhou, Paiheng Xu, Xiaoyu Liu, Bang An, Wei Ai, an...
https://arxiv.org/abs/2505.22630v1
datasets are reported. In the โ€˜Totalโ€™ row, under โ€˜Prop.โ€™ column, it indicates the global accuracy across different cases, while under โ€˜ โˆ†Rateโ€™ column, it underlies the global answer change rate. Table 8 shows that models are not robust against irrelevant context. Even when a single irrelevant demonstration is prepended...
https://arxiv.org/abs/2505.22630v1
candidates โ€˜Vietnameseโ€™ and โ€˜Thaiโ€™ for Example 1 in Table 17 have the correct class โ€˜languageโ€™, but โ€˜Southโ€™, โ€˜Ko- reaโ€™ in Example 5 in Table 17 do not have the correct class because the query is asking about con- tinent, not country. Step 4: Hypothesis Verification Finally, a context-based candidate is considered to sa...
https://arxiv.org/abs/2505.22630v1
center around 0 โ€“ suggesting that the computation of abstract class representations exists for zero-shot case, and is not influenced by the added irrelevant context. Logit attribution results for Mistral 7B are pre- sented in Figure 7, and for Pythia 6.9B in Figure 8. We remark that these plots follow a similar pattern...
https://arxiv.org/abs/2505.22630v1
the last token position, and we only allow models to attend to the query part. Similarly, in the query-dominant case, we set the attention scores corresponding to all tokens in the query to be โˆ’โˆž, allowing the models to only retrieve information from the context. To compare the knockout effect of the two criti- cal lay...
https://arxiv.org/abs/2505.22630v1
Lowโ€ = Two lower layers (<17), โ€œ2 Highโ€ = Two higher layers (>24). โ€œDiff.โ€ represents the probability difference, and โˆ†de- notes the change from the original setting. (Llama-3) I Ablation Studies I.1 Experimental Details Due to computational constraints, we cannot in- ference on the full C+Qdataset with 102M for larger...
https://arxiv.org/abs/2505.22630v1
P136 [X] plays [Y]. Others 859 P1376 [X], the capital city of [Y]. Place 179 P138 [X], which is named after [Y]. Others 461 P140 [X] is follower of [Y]. Others 432 P1412 [X] communicated in [Y]. Language 924 P159 [X] is headquartered in [Y]. Place 801 P17 [X], located in [Y]. Place 912 P176 [X], produced by [Y]. Compan...
https://arxiv.org/abs/2505.22630v1
of [Y]. Others 746 P364 The original language of [X] was [Y]. Language 756 P37 The official language of [X] is [Y]. Language 900 P39 [X], whose position is that of [Y]. Job 485 P407 The language of [X] is [Y]. Language 857 P413 [X] plays in the position of [Y]. Job 952 P449 [X] debuted on [Y]. Company 801 P463 [X] is a...
https://arxiv.org/abs/2505.22630v1
arXiv:2505.22645v1 [cs.CL] 28 May 2025Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinese Hanjia Lyu hlyu5@ur.rochester.edu University of Rochester Rochester, New York, USAJiebo Luo jluo@cs.rochester.edu University of Rochester Rochester, New York, USA Jian Kang jian.kang@ro...
https://arxiv.org/abs/2505.22645v1
Hong Kong, and Macau [ 31,57]. As Large Language Models (LLMs) have become integral to var- ious applications in daily life [ 40,53,64], it is increasingly im- perative to study their variance in behavior across languages and cultures [ 2,5,58,61], especially as these models become more multilingualโ€”although many remai...
https://arxiv.org/abs/2505.22645v1
using examples of regional terms from Mainland China and Taiwan: โ€ขSame term, same word: Terms may share the same word in both Mainland China and Taiwan, although some are written identically while others appear in different scripts: โ€“Same script: For instance, โ€œmilk teaโ€ is often written identi- cally in both regions. ...
https://arxiv.org/abs/2505.22645v1
occur, even between prima facie โ€œsimilarโ€ variants of a language. While the LLMs studied may be technically competent, they may not always be neutral or fair in their application of that competenceโ€”potentially perpetuating so- cietal inequalities or biases [ 7]. Our analyses are presented as a reproducible framework th...
https://arxiv.org/abs/2505.22645v1
leave the examination of the terms used in other regions such as Hong Kong and Macau to future work.integrated into various decision-making processes across hiring, such as resume screening and interviewee selection [ 15,45,51]. These concerns have prompted legal interventions, such as New York Cityโ€™s Local Law 144, wh...
https://arxiv.org/abs/2505.22645v1
the popularity of these names 2This scope does not account for other regional uses of Chinese scripts or linguistic variations in countries such as Malaysia, or Singapore. Consequently, we refrain from extrapolating our findings to these countries, as the results may not accurately reflect the complexities of Chinese l...
https://arxiv.org/abs/2505.22645v1
and apply the Benjamini-Hochberg correction [ 50]. 2.4 Language Models We benchmark 11 LLMs, which we categorize based on the primary language of the training corpora. Following Zhang and Li [62], we refer to the three LLM categories as English, Simplified and Tradi- tional Chinese-oriented LLMs. For the exact model va...
https://arxiv.org/abs/2505.22645v1
(comparing the โ€œSโ€ and โ€œTโ€ yellow plain bars for each LLM): All LLMs are significantly more likely to gen- erate misaligned responses when prompted in Traditional Chinese compared to Simplified Chinese (๐‘<.05). This suggests that, Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional...
https://arxiv.org/abs/2505.22645v1
in the response is accurate, but the expression used is uncommon. See Appendix B.4 for more details.the equivalent Simplified Chinese terms instead) has to do with a language imbalance in LLM training data [ 5,47], grounded in the fact that Simplified Chinese is more prevalent in online and global datasets [ 35]. Howev...
https://arxiv.org/abs/2505.22645v1
to Traditional Chinese appearances is extremely low among non-misaligned terms ( i.e., as expected, LLMs perform well at recovering Traditional Chinese terms that are well-represented in corpora), but this ratio is much higher among misaligned terms ( i.e., Traditional Chinese appearances of misaligned terms are underr...
https://arxiv.org/abs/2505.22645v1
tends to yield the highest rate of valid responses in this task across LLMs (even those that are not English-oriented); it remains the case that (even among only LLMs that have high response rates) the majority of LLMs select Taiwanese names. It is noteworthy that simply changing prompting language (while holding the c...
https://arxiv.org/abs/2505.22645v1
Chinese, or Traditional Chinese-orientedโ€”tend to select a valid Taiwanese name more often than a valid Mainland Chinese name for the regional name choice task (as indicated by the majority of points falling below the 50% dotted horizontal line for Mainland Chinese Name Rate). Furthermore, no LLMs display consistently l...
https://arxiv.org/abs/2505.22645v1
and Taiwanese name lists from this exclusion. FAccT โ€™25, June 23โ€“26, 2025, Athens, Greece Hanjia Lyu, Jiebo Luo, Jian Kang, and Allison Koenecke popularity in their respective regions. We define popularity based on the percentage of people in each region who bear that name, and bin names into ten distinct deciles based...
https://arxiv.org/abs/2505.22645v1
to candidate lists where both sets of 10 names have the same gender ratio, e.g., 7 male and 3 female names.Mainland Chinese name selection rate falls below 50%) are pre- sented in Tables 24, 25, and 26. To supplement these observational results, we now repeat the candidate name list experiment from Section 4.1, but thi...
https://arxiv.org/abs/2505.22645v1
Since the corresponding report for Mainland Chinese names [ 39] did not include gender information, we use GPT-4o-mini to infer the gender of each name and manually verify the labels. Additional details are provided in Appendix C.6. Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditiona...
https://arxiv.org/abs/2505.22645v1
restrict our candidate name list to only include these six names, and otherwise prompt in the same ways (requesting for one name to be chosen), running 8,000 trials of this experiment. Figure 4 illustrates that the selection bias fa- voring Taiwanese names is ameliorated when the names (but not scripts) are kept consta...
https://arxiv.org/abs/2505.22645v1
the training data can lead to over-fragmentation during tokenization. Moreover, Ahia et al . [3]note that script-specific linguistic features can further exac- erbate fragmentation. These factors indicate that tokenization is not merely a technical preprocessing step but a potential source of 9Exceptions, where Simplif...
https://arxiv.org/abs/2505.22645v1
to LLM-based disparities between writers of Simplified and Traditional Chinese. We first uncovered that underlying training data may be a driver of biases favoring Sim- plified Chinese in the regional term task; this points to a need for diversifying underlying training data and collecting niche data on regional terms....
https://arxiv.org/abs/2505.22645v1
(1992). [13] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. GLM: General Language Model Pretraining with Autoregressive Blank Infilling. (2022), 320โ€“335. [14] Philipp Ennen, Po-Chun Hsu, Chan-Jan Hsu, Chang-Le Liu, Yen-Chen Wu, Yin- Hsiang Liao, Chin-Tung Lin, Da-Shan Shiu...
https://arxiv.org/abs/2505.22645v1
Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, Yifan Xu, Weng Lam Tam, et al .2023. Alignbench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743 (2023). [33] Mapull. 2022. Chinese Pinyin Dictionary. https://github.com/mapull/chinese- ...
https://arxiv.org/abs/2505.22645v1
arXiv preprint arXiv:2403.01858 (2024). [50] David Thissen, Lynne Steinberg, and Daniel Kuang. 2002. Quick and easy imple- mentation of the Benjamini-Hochberg procedure for controlling the false positive rate in multiple comparisons. Journal of educational and behavioral statistics 27, FAccT โ€™25, June 23โ€“26, 2025, Athe...
https://arxiv.org/abs/2505.22645v1
of the North American Chapter of the Associa- tion for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers) , Marilyn Walker, Heng Ji, and Amanda Stent (Eds.). Association for Com- putational Linguistics, New Orleans, Louisiana, 15โ€“20. doi:10.18653/v1/N18-2003 [64] Wayne Xin Zhao, Kun Zhou, J...
https://arxiv.org/abs/2505.22645v1
any content or sentence structures that are not commonly used in Taiwan. Each reviewer read the translations and made their decisions independently. After review, all translations are confirmed to be frequently used in Taiwan. A.5 Power Analysis We examine whether a minimum difference of 5% exists between two proportio...
https://arxiv.org/abs/2505.22645v1
we decided to set the number of repeated trials at 15 for the regional term choice task. FAccT โ€™25, June 23โ€“26, 2025, Athens, Greece Hanjia Lyu, Jiebo Luo, Jian Kang, and Allison Koenecke LanguagePromptSimplified Chinese่ฏท้—ฎ'{definition}'ๆ˜ฏๆŒ‡ไป€ไนˆ๏ผŸ่ฏทโฝคโผ€ไธช่ฏๅ›ž็ญ”Traditional Chinese่ซ‹ๅ•'{definition}'ๆ˜ฏๆŒ‡ไป€้บผ๏ผŸ่ซ‹โฝคโผ€ๅ€‹่ฉžๅ›ž็ญ” Table 7: The first rephr...
https://arxiv.org/abs/2505.22645v1
deviation . considered not frequently used by all three reviewers are excluded, resulting in the removal of four terms out of 114. However, language evolves over time, and since the Cross-Strait vocabularies [ 26] were published in 2014, some terms may have become widely used in both regions. To explore this possibilit...
https://arxiv.org/abs/2505.22645v1
Figure 2, indicating that the use of GPT-4o to generate item definitions does not impact the findings. The Pearson correlation coefficients between the percent- age of correct, misaligned, and incorrect responses in Figure 2 and Figure 5 are 0.999 ( ๐‘<.001), 0.998 (๐‘<.001), and 0.997 ( ๐‘<.001), respectively. A.9 Add...
https://arxiv.org/abs/2505.22645v1
in technical reportsโ€”make it challenging to directly assess the impact FAccT โ€™25, June 23โ€“26, 2025, Athens, Greece Hanjia Lyu, Jiebo Luo, Jian Kang, and Allison Koenecke ST ST ST ST ST ST ST ST ST ST ST020406080100% Responses by CorrectnessCorrect response Misaligned response Incorrect response Qwen-1.5 Baichuan-2 Chat...
https://arxiv.org/abs/2505.22645v1
in Table 21. C.3 Statistics of Collected Names Mainland Chinese names are sourced from the name report pub- lished by the Ministry of Public Security of the Peopleโ€™s Republic of China in 2013 [ 39], while Taiwanese names are obtained from the name report published in Taiwan in 2018 [ 1]. It is important to note that ne...
https://arxiv.org/abs/2505.22645v1
be potential confounders, such as names of prominent business figures, politicians, or celebrities, that could skew the re- sults of the regional name choice task. To examine this possibility, we conducted the online-based name popularity experiment de- scribed in Section 4.3.2. As an illustrative example, consider two...
https://arxiv.org/abs/2505.22645v1
respect to scripts. C.7 Descriptive Word Extraction We prompt GPT-4o-mini to extract the descriptive words from LLM responses with the prompt: โ€œPlease determine if there are any adjectives describing the name {name} in the provided text: {text}. Do not include the adjectives in the name itself. If adjectives are Charac...
https://arxiv.org/abs/2505.22645v1
isolating variation in the first name. We repeat this experiment across three prompting languages: Simplified Chinese, Traditional Chinese, and English. For each name pair, we compare the token generation probabili- ties (raw and conditioned) of the shared last name. Table 35 shows the agreement rate between raw and co...
https://arxiv.org/abs/2505.22645v1
fast train ๆ™ฎๅฟซ/ๆ™ฎ้€šๅฟซ้€Ÿๅˆ—่ฝฆ 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 13,0,2 0,0,15 0,0,15 0,0,15 3,0,12 0,0,15 Rail police ไน˜่ญฆ 15,0,0 0,0,15 0,0,15 0,15,0 0,0,15 14,1,0 15,0,0 0,15,0 0,0,15 2,3,10 2,5,8 Railway police ้“่ญฆ 0,0,15 0,15,0 0,0,15 0,0,15 0,0,15 13,2,0 15,0,0 15,0,0 15,0,0 8,5,2 0,14,1 Maglev train ็ฃๆ‚ฌๆตฎๅˆ—่ฝฆ 15,0,0 15,0,0 0,0,...
https://arxiv.org/abs/2505.22645v1
congestion ไบค้€šๆ‹ฅๅ ต 9,0,6 0,0,15 15,0,0 0,0,15 15,0,0 12,0,3 0,0,15 1,14,0 0,0,15 1,0,14 5,3,7 Rear-end collision ่ฟฝๅฐพ 15,0,0 0,0,15 0,0,15 0,0,15 8,0,7 15,0,0 15,0,0 15,0,0 0,0,15 5,1,9 2,0,13 Alcohol tester ้…’็ฒพๆต‹่ฏ•ไปช 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 12,0,3 15,0,0 15,0,0 0,0,15 2,0,13 6,0,9 Driving school ้ฉพๆ ก 15,0,0 15,0,0 3,0...
https://arxiv.org/abs/2505.22645v1
15,0,0 0,0,15 13,0,2 15,0,0 13,0,2 14,1,0 15,0,0 15,0,0 0,0,15 7,0,8 2,0,13 Exhaust fan ๆŽ’้ฃŽๆ‰‡ 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 4,0,11 12,0,3 15,0,0 0,0,15 3,0,12 2,0,13 Defective product ๆฎ‹ๅ“ 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 2,0,13 0,0,15 0,0,15 0,0,15 2,0,13 2,0,13 False advertisement ่™šๅ‡ๅนฟๅ‘Š 15,0,0 0,0,15 0,0,15 0,0,15 ...
https://arxiv.org/abs/2505.22645v1
in yellow. In addition, the terms for which more than half of selected LLMs tend to misalign are also highlighted in yellow. Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinese FAccT โ€™25, June 23โ€“26, 2025, Athens, Greece Translation Regional Term Qwen-1.5 Baichuan-2 ChatGLM-...
https://arxiv.org/abs/2505.22645v1
police ็พฉ่ญฆ 0,0,15 0,0,15 0,0,15 15,0,0 0,0,15 3,0,12 0,6,9 15,0,0 0,0,15 4,0,11 0,0,15 Long-distance bus ๅฎข้‹ๆฑฝโพž 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 1,0,14 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 Private car ๅฎถๅบญโพž 0,15,0 0,15,0 0,7,8 0,15,0 0,0,15 0,15,0 0,15,0 0,15,0 0,15,0 1,11,3 0,14,1 Illegal taxi โฝฉ็‰Œโพž 0,0,15 0,15,0 0,0,15 0,0,...
https://arxiv.org/abs/2505.22645v1
0,15,0 0,15,0 5,8,2 4,2,9 Lobby manager โ‰คโ‰ฅ็ตฐ็† 15,0,0 0,0,15 0,0,15 0,0,15 0,0,15 4,0,11 15,0,0 0,0,15 0,0,15 2,0,13 0,0,15 Commercial-residential building ไฝๅ•†โผคๆจ“ 0,15,0 0,0,15 0,15,0 0,0,15 0,0,15 0,13,2 0,0,15 0,0,15 0,0,15 0,0,15 0,1,14 Duplex apartment ๆจ“ไธญๆจ“ 0,0,15 0,0,15 0,0,15 0,0,15 0,0,15 2,0,13 0,0,15 0,15,0 0,0,15 ...
https://arxiv.org/abs/2505.22645v1
3,12,0 0,0,15 0,3,12 0,0,15 Digital TV ๆ•ธไฝ้›ป่ฆ– 0,0,15 0,15,0 0,15,0 0,0,15 0,15,0 2,13,0 3,12,0 15,0,0 15,0,0 9,0,6 2,0,13 Video recorder ้Œ„ๅฝฑๆฉŸ 15,0,0 0,15,0 13,2,0 15,0,0 15,0,0 6,9,0 15,0,0 15,0,0 15,0,0 7,0,8 4,2,9 Digital camera ๆ•ธไฝ็›ธๆฉŸ 6,0,9 0,15,0 8,0,7 0,0,15 2,0,13 5,10,0 0,15,0 15,0,0 15,0,0 3,2,10 2,2,11 Sanitary pad...
https://arxiv.org/abs/2505.22645v1
Multiple names 19.0% 0.0% 0.0% Out-of-list name 11.0% 0.0% 100.0% Table 20: Breakdown of invalid response types across 100 sampled outputs for ChatGLM-2 ,Breeze , and GPT-4o when prompted in Simplified Chinese. FAccT โ€™25, June 23โ€“26, 2025, Athens, Greece Hanjia Lyu, Jiebo Luo, Jian Kang, and Allison Koenecke EnglishSim...
https://arxiv.org/abs/2505.22645v1
number of times the gender distribution appears in the experiment is 900, 1,260, and 180, respectively. We conduct a one-sided z-proportion test to examine whether the Mainland Chinese name selection rate is significantly below 50%. NS: Not significant, *: ๐‘<.05, **:๐‘<.01, ***:๐‘<.001. โ€œ-โ€ means there is no valid res...
https://arxiv.org/abs/2505.22645v1
9 female, 2 male / 8 female, or 3 male/ 7 female for each set of Taiwanese and Mainland Chinese names comprising the 20 candidate name list) used in the experiments of Section 4.3, when the models are prompted in English . The majority of LLMs tend to select Taiwanese names even when gender distributions are held const...
https://arxiv.org/abs/2505.22645v1
when the models are prompted in English. The number of times the gender distribution appears in the experiment is 900, 1,260, and 180, respectively. โ€œ-โ€ means there is no valid response. Model Simplified Chinese Traditional Chinese English Male % Significance # Valid Responses Male % Significance # Valid Responses Male...
https://arxiv.org/abs/2505.22645v1
็Ž‹ๆก‚่‹ฑ, ๅผ ๆก‚่‹ฑ ็Ž‹โฝŸๅ…ฐ, ๅผ โฝŸๅ…ฐ ๆŽโฝŸๅ…ฐ, ๅˆ˜โฝŸๅ…ฐ ็Ž‹ๆก‚ๅ…ฐ, ๅผ ๆก‚ๅ…ฐ, ๆŽๆก‚ๅ…ฐ ็Ž‹โฝŸๆข…, ๅผ โฝŸๆข… ๅผ ๅ‡ค่‹ฑ, ๆŽๅ‡ค่‹ฑ, ็Ž‹ๅ‡ค่‹ฑ ็Ž‹โฝŸ่‹ฑ, ๅผ โฝŸ่‹ฑ ็Ž‹ๅปบๅŽ, ๆŽๅปบๅŽ, ๅผ ๅปบๅŽ ๅผ โฝŸ็, ๆŽโฝŸ็ ๅผ ๅ…ฐ่‹ฑ, ็Ž‹ๅ…ฐ่‹ฑ ๅผ ๅฉทๅฉท, ็Ž‹ๅฉทๅฉท ๅผ ๅปบๅ›ฝ, ๆŽๅปบๅ›ฝ ๆŽ็ง€ๆข…, ๅผ ็ง€ๆข… ๆŽๅปบๅ†›, ๅผ ๅปบๅ†› ๅผ ๆตท็‡•, ๆŽๆตท็‡• ๆŽๅฟ—ๅผบ, ๅผ ๅฟ—ๅผบ ็Ž‹ไธฝไธฝ, ๅผ ไธฝไธฝ ๅผ ๆก‚่Šณ, ็Ž‹ๆก‚่Šณ ็Ž‹็ง€ๅŽ, ๆŽ็ง€ๅŽ ๆŽๆท‘่‹ฑ, ๅผ ๆท‘่‹ฑ ็Ž‹ๆก‚่Šฑ, ๆŽๆก‚่Šฑ ๅผ ๆท‘ๅ…ฐ, ๆŽๆท‘ๅ…ฐ ้™ณๆ€กๅ›, ๆž—ๆ€กๅ› ้™ณๅปบๅฎ, ๆž—ๅปบๅฎ ้™ณๆท‘ๆƒ , ๆž—ๆท‘ๆƒ  ้™ณ้บ—่ฏ, ๆž—้บ—่ฏ ็Ž‹ไฟŠๅ‚‘, โฟˆไฟŠๅ‚‘ ๅณๆ‰ฟ็ฟฐ, ๆŽๆ‰ฟ็ฟฐ ๆŽๆ‰ฟๆฉ, ๅณๆ‰ฟๆฉ Table 34: Combinations of last names with mu...
https://arxiv.org/abs/2505.22645v1
WebDancer: Towards Autonomous Information Seeking Agency Jialong Wuโˆ—, Baixuan Liโˆ—, Runnan Fangโˆ—, Wenbiao Yinโˆ—โ€ , Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Yong Jiangโ€ , Pengjun Xie, Fei Huang, Jingren Zhou Tongyi Lab , Alibaba Group Correspondence to: wujialongml@gmail.com {yinwenbiao.ywb,yongjiang.jy}@alibaba-...
https://arxiv.org/abs/2505.22648v1
autonomously . Observations from these interactions guide subsequent reasoning and actions until the task is completed. This process is optimized through end-to-end tool-augmented training. The ReAct framework [ 11] is the most suitable paradigm, as it tightly couples reasoning with action to facilitate effective learn...
https://arxiv.org/abs/2505.22648v1
systematic, end-to-end pipeline for building long-term information-seeking web agents. Extensive experiments on two web information seeking benchmarks, GAIA and WebWalkerQA, show the effectiveness of our pipeline and WebDancer (ยง4). We further present a comprehensive analysis covering data efficiency, agentic system ev...
https://arxiv.org/abs/2505.22648v1
order to search via search engine Sfor information Cnrelated to En. After that, we use LLMs ฯ€to restructure the obtained content into a new query Rn to replace the original entity in the question. The process can be signaled as: Rn=ฯ€(S(Cn)). This way, the new question Qn+1requires solving the sub-problem we have constr...
https://arxiv.org/abs/2505.22648v1
the generated trajectory, as they serve as valuable supervision signals. The LRMโ€™s intermediate reasoning process, denoted as, denoted as โ€œ<reasoning_content> โ€, is recorded as the current thought of the current step. Each constructed QA instance undergoes rejection sampling up to Ntimes to ensure quality and coherence...
https://arxiv.org/abs/2505.22648v1