text string | source string |
|---|---|
Tech- nologies, Volume 1 (Long and Short Papers) , pages 4149–4158. Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open founda- tion and fine-tuned chat models. arXiv preprint arXiv:2307.... | https://arxiv.org/abs/2505.20971v1 |
Badr AlKhamissy, Gargi Ghosh, Mona T. Diab, and Asli Celikyilmaz. 2022. ALERT: adapting language mod- els to reasoning tasks. CoRR , abs/2212.08286. Jing Zhang, Xiaokang Zhang, Jifan Yu, Jian Tang, Jie Tang, Cuiping Li, and Hong Chen. 2022. Subgraph retrieval enhanced model for multi-hop knowledge base question answeri... | https://arxiv.org/abs/2505.20971v1 |
straints are involved. A.2.3 EM as the Principled Approach for Latent Variables •EM is the standard, principled approach for parameter estimation in models with latent variables. •It provides a formal framework to optimize the Reasoner (generating the latent NL chain) and the Aligner (grounding the chain to the KG) wit... | https://arxiv.org/abs/2505.20971v1 |
logEz∼pψ(z|G,q) pw(a|q, z) 13 is approximated by log 1 KKX k=1pw(a|q, zk) . We then take gradients (w.r.t. w) and update w so that pw(a|q, z)is more likely to produce the correct afor the sampled Graph-aware Reasoning Chains. 3.Result. After updating w, Responser pw(a| q, z)is better aligned with whatever Graph- aw... | https://arxiv.org/abs/2505.20971v1 |
Li et al., 2023; Nie et al., 2024); they seldom perform task-specific optimization that directly targets the core reasoning mechanism. Accuracy, reliability, and complexity han- dling. The synergy between NL reasoning, KG-constrained alignment, and EM -guided super- vised fine -tuning translates into markedly higher ac... | https://arxiv.org/abs/2505.20971v1 |
both alignable and answer -bearing, RAR avoids this brittleness and achieves the quality gains visualized in Fig. 4 of the submission. In summary, the collaboration of an explicit NL Reasoner, a KG -constrained Aligner, and EM-based optimization endows RARwith a distinc- tive combination of interpretability, flexibilit... | https://arxiv.org/abs/2505.20971v1 |
decoding process. This high-performance infrastructure allows us to efficiently retrieve next- hop candidates for token constraints directly from the knowledge graph. E Case Study of Different Iteration Steps In Fig. 7, we provide two examples to investigate the effect of iteration steps of the EM algorithm. F Template... | https://arxiv.org/abs/2505.20971v1 |
for the given question. Ensure the thinking process is aligned with triples in the knowledge base. Question: <Question> Query entities: <Question Entities> ============================= LLM Output ================================ Thinking process: <Reasoning Chain> Align Process: <Knowledge Path> Figure 10: The prompt ... | https://arxiv.org/abs/2505.20971v1 |
arXiv:2505.20976v1 [cs.CL] 27 May 2025Contrastive Learning on LLM Back Generation Treebank for Cross-domain Constituency Parsing Peiming Guo‡*, Meishan Zhang‡, Jianling Li§, Min Zhang‡, Yue Zhang¶† ‡Institute of Computing and Intelligence, Harbin Institute of Technology (Shenzhen), China §School of New Media and Commun... | https://arxiv.org/abs/2505.20976v1 |
and then the chart- based parser annotates a pseudo treebank on them. Integrating with self-training, this two-stage ap- proach gains state-of-the-art cross-domain con- stituency parsing performance. However, this approach indirectly uses LLM for treebank annotation, not taking full advantage of LLM abilities in domain... | https://arxiv.org/abs/2505.20976v1 |
bank, conventional parsers, masked language mod- eling pre-training, previous cross-domain methods and large language models1. 2 Related Work Cross-domain Constituency Parsing. Con- stituency parsing is an important and fundamental task in computational linguistics, which has not been completely solved. The main challe... | https://arxiv.org/abs/2505.20976v1 |
and negative instances by other examples in the same batch. However, our proposed strategy takes the left child, right child, parent and brother nodes as positive instances, and the fifteen corresponding invalid spans as negative instances. 3 Method In this section, the process of LLM back genera- tion (§ 3.1) is first... | https://arxiv.org/abs/2505.20976v1 |
Iam proundDemonstration LLM Back Generation TreeFigure 2: Overview of LLM Back Generation. We first extract the target domain constituency tree and domain keywords. Then we mask all sentential words except domain keywords from the constituency tree. Finally, the LLM back generates the whole syntax tree based on the mas... | https://arxiv.org/abs/2505.20976v1 |
learning strategy is based on span-level not example-level, which can significantly expand pre-training data. Specifi- cally, we start with the introduction of the positive and negative instances for each constituent span (i, j)in the LLM back generation tree. The goal of the chart-based parser is to distinguish all va... | https://arxiv.org/abs/2505.20976v1 |
span with the start index iand end index j, our contrastive objective Lis based on its span representation r: L=−X m∈(i,j)+logef(r,r+ m) P n∈(i,j)−ef(r,r+ m)+ef(r,r− n), where findicates the cosine similarity function divided by temperature factor τ. We pre-train BERT (Devlin et al., 2019), parti- tioned transformer an... | https://arxiv.org/abs/2505.20976v1 |
We use gpt-3.5-turbo andgpt-4 to generate bracketed parse trees with in-context-learning (Brown et al., 2020), where 10 constituency tree examples from the source treebank PTB are prepended before the testing instance as demonstrations. Notably, the outputs can contain numerous errors, includ- ing unmatched brackets, o... | https://arxiv.org/abs/2505.20976v1 |
transfer to constituent recognition. CTPT significantly im- proves the parsing performance across five domains compared with DAPT and NOPT. The observation suggests that our span-level contrastive learning pre-training effectively acquires constituent knowl- edge from the LLM back generation treebank. Finally, we make ... | https://arxiv.org/abs/2505.20976v1 |
decreases gradually. A reasonable explana- tion might be that fewer retained domain keywords cause LLM generation to be freer. Therefore, the final LLM back generation treebank will shift from the target domain. Third, our contrastive learn- ing pre-training strategy significantly improves the results compared with DAP... | https://arxiv.org/abs/2505.20976v1 |
tree as the input instruction and fills in missing words to generate the LLM back generation tree- bank. Besides, we presented a span-level con- trastive learning pre-training strategy for the LLM back generation treebank. To the best of our knowl- edge, this is the first work to introduce contrastive learning into con... | https://arxiv.org/abs/2505.20976v1 |
Computational Linguistics . Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tec... | https://arxiv.org/abs/2505.20976v1 |
McClosky, Eugene Charniak, and Mark Johnson. 2008. When is self-training effective for parsing? In Proceedings of the 22nd International Conference on Computational Linguistics . David McClosky, Eugene Charniak, and Mark Johnson. 2010. Automatic domain adaptation for parsing. In Human Language Technologies: The 2010 An... | https://arxiv.org/abs/2505.20976v1 |
for dense document retrieval. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . Sen Yang, Leyang Cui, Ruoxi Ning, Di Wu, and Yue Zhang. 2022. Challenges to open-domain con- stituency parsing. In Findings of the Association for Computational Linguistics: ACL 2022 . Ruoyu Zhang, ... | https://arxiv.org/abs/2505.20976v1 |
each constituency tree in a batch, only 20% con- stituents are sampled as examples to compute the contrastive learning loss. Temperature factor τ is 0.05. For constituency parsing fine-tuning, we use the AdamW algorithm with learning rate 1e-5, batch size 64, and linear learning rate warmup over the first 400 steps to ... | https://arxiv.org/abs/2505.20976v1 |
generated corpus is 20 times more than ours, which leads to an unaffordable and un- avoidable high cost. We guess it is an important reason why they do not apply more powerful GPT- 4 to generate sentences. For another thing, our parser can handle sentences from all five domains simultaneously. But Li et al. (2023) trai... | https://arxiv.org/abs/2505.20976v1 |
arXiv:2505.20977v1 [cs.CL] 27 May 2025Evaluating and Steering Modality Preferences in Multimodal Large Language Model Yu Zhang1,2Jinlong Ma1Yongshuai Hou2Xuefeng Bai1Kehai Chen1,2 Yang Xiang2Jun Yu1Min Zhang1,2 1Harbin Institute of Technology, Shenzhen, China 2Peng Cheng Laboratory, Shenzhen, China yuzhang2717@gmail.co... | https://arxiv.org/abs/2505.20977v1 |
In this paper, we provide a controlled setup to study the modality preference in MLLMs. As shown in the left panel of Figure 1, we introduce a modality context conflict evaluation setting, where an MLLM is asked to answer a question based on a pair of contrasting evidence. In this way, we can determine the modality pre... | https://arxiv.org/abs/2505.20977v1 |
one common approach involves evaluating model performance on unimodal inputs—comparing results when only text or only image inputs are provided for the same task [ 12]. However, this method inherently overlooks how models handle genuinely multimodal scenarios where both modalities are presented simultaneously, thus lim... | https://arxiv.org/abs/2505.20977v1 |
LLMs to generate textual contexts ctthat conflict with the vision context cvin relation to the question q.AvandAtare answers based on the vision context or textual context. 3.2 Semi-automated Data Construction Pipeline In this section, we introduce our semi-automated data construction pipeline, which follows a meticu- ... | https://arxiv.org/abs/2505.20977v1 |
dataset is organized into eight categories: positional reasoning, counting and color, attribute recognition, sentiment analysis, activity recognition, sports recognition, and object recognition, with 250 samples in each category. Details of the data annotation format along with sample cases and dataset statistics is pr... | https://arxiv.org/abs/2505.20977v1 |
can be controlled. This investigation helps uncover the underlying mechanisms of modality preference and enables us to apply these insights to downstream tasks. 4.1 Which Modality Do MLLMs Prioritize? Before evaluating modality preference, we first assess the ability of MLLMs to answer questions accurately given a sing... | https://arxiv.org/abs/2505.20977v1 |
and LLaV A-OneVision families to investigate the relationship between model size and modality preference. As shown in the left panel of Figure 3, we observe that for all model families, the preference for the vision modality increases with the model size. And the Qwen2.5VL and InternVL3 models exhibit a significant pre... | https://arxiv.org/abs/2505.20977v1 |
that MLLMs may implicitly associate modality reliability with surface-level quality signals (e.g., noise or errors), which then influences their modality weighting during reasoning. Details of the data generation, experimental design, and full results are provided in Appendix C.2. 30 20 10 0 10 20 30 PC1020406080100120... | https://arxiv.org/abs/2505.20977v1 |
representations in response to each modality preference prompt. It then computes and scales the direction of modality preference that expresses a preference for a specific modality. Modality Preference Steering selects the target layer during the second inference and adds the scaled modality preference direction to the... | https://arxiv.org/abs/2505.20977v1 |
expressing the text modality at layer lfor a given query as: ut ℓ=1 NNX i xt i,ℓ−xv i,ℓ . (1) Averaging over different queries allows us to capture the activation values most closely associated with modality preference requests, independent of questions. As shown in the right panel of Figure 5, we compute the absolut... | https://arxiv.org/abs/2505.20977v1 |
complex reasoning through intermediate steps; and Few-shot uses a few examples to guide the model in answering. For detailed method implementation and the results of LLaV A-OneVision-7B and InternVL3-14B, refer to Appendix D.2. As shown in Table 2, the proposed method significantly outperforms the baseline approaches, ... | https://arxiv.org/abs/2505.20977v1 |
to downstream tasks, such as hallucination mitigation and multimodal machine translation. 11 References [1]Diff-erank: A novel rank-based metric for evaluating large language models. arXiv preprint arXiv:2401.17139 , 2024. 3 [2]Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tas... | https://arxiv.org/abs/2505.20977v1 |
Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. arXiv preprint arXiv:2401.13919 , 2024. 1 [17] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language ... | https://arxiv.org/abs/2505.20977v1 |
the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 26296–26306, June 2024. 4, 18, 21 [30] Jiazhen Liu, Yuhan Fu, Ruobing Xie, Runquan Xie, Xingwu Sun, Fengzong Lian, Zhanhui Kang, and Xirong Li. Phd: A chatgpt-prompted visual hallucination evaluation dataset. arXiv preprint arXiv:2403.111... | https://arxiv.org/abs/2505.20977v1 |
, 2023. 3 [45] Chen Henry Wu, Neil Kale, and Aditi Raghunathan. Why foundation models struggle with cross-modal context. In ICLR 2025 Workshop on Foundation Models in the Wild . 3, 5, 7 [46] Muling Wu, Wenhao Liu, Xiaohua Wang, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaoqing Zheng, and Xuanji... | https://arxiv.org/abs/2505.20977v1 |
investigates the modality preference in multimodal large language models (MLLMs) using a controlled experiment setup with a modality conflict dataset. In constructing the dataset, we employs LLaV A1.5-7B and QwenVL-7B to filter samples and ensure that most models could answer questions correctly based on a single modal... | https://arxiv.org/abs/2505.20977v1 |
within the context. Only one alternative answer should be generated. Caption: { caption } Question: { question } Output the new answer enclosed in <answer> </answer> and the context enclosed in <context> </context> tags. Conflict Context Generation for count task using GPT-4o-mini Instruction: # Given the caption of an... | https://arxiv.org/abs/2505.20977v1 |
improve clarity, regenerate the problematic part of the sample (e.g., the question or context), or discard the sample entirely if it cannot be reasonably corrected. To ensure consistency and reduce subjectivity, each category (i.e., vision-aligned, text-aligned, and conflict) is independently verified by three trained ... | https://arxiv.org/abs/2505.20977v1 |
park filled with lush green trees . The contrast between the brown tower and the sur- rounding greenery creates a picturesque scene. Ques- tion: What color are the trees? Assistant: <output> vision-based Answer: white Text-based Answer: green<image> is a placeholder for below image User: <image> Conflict Text Context: ... | https://arxiv.org/abs/2505.20977v1 |
excluded this confounding factor from the analysis. Details of results for modality preference evaluation We provide the results of modality prefer- ence for several models in Table 1 in the main text. More detailed modality preference evaluation results are presented in Table 7. We evaluated 18 models, including 17 op... | https://arxiv.org/abs/2505.20977v1 |
modality preference with surface-level quality signals (e.g., noise or errors), which then influences their modality weighting during reasoning. We provide the results of modality preference using either the noisy image context or the text context with grammer error in Figure 4 in main text. We also provide more result... | https://arxiv.org/abs/2505.20977v1 |
LLaV A-OneVision-7B. As shown in Table 6, the proposed method significantly outperforms the baseline in both directions of modality preference. D.3 Downstream Task Applications By steering the model’s modality preference towards the vision modality, we reduce vision hallu- cinations in MLLMs. Conversely, by steering to... | https://arxiv.org/abs/2505.20977v1 |
60.0/29.6 28.0/63.5 QwenVL-7B 35.2/43.6 15.2/63.2 11.2/78.0 16.4/45.6 2.8/73.6 34.0/40.8 19.2/61.6 35.2/31.2 21.1/54.7 Qwen2VL-7B 26.4/58.0 12.4/82.8 0.8/95.6 13.2/80.4 4.0/93.6 16.0/78.8 11.6/83.6 38.0/54.0 15.3/78.3 Qwen2.5VL-7B 65.6/12.8 45.2/46.0 18.0/68.8 46.4/38.0 51.6/39.6 70.8/20.0 42.0/43.6 77.6/14.0 52.2/35.4... | https://arxiv.org/abs/2505.20977v1 |
visual contexts are provided and the instruction explicitly directs the model to answer based on the textual modality. Values are reported as vision-based accuracy/text-based accuracy for each model. Model Sport Attribute Sentiment Positional Counting Color Activity Object Avg LLaMA Vision 97.6 97.2 99.6 99.2 97.2 96.0... | https://arxiv.org/abs/2505.20977v1 |
98.8 98.4 99.6 100.0 98.6 SPHINX-V2-1K 98.4 97.6 99.2 98.8 97.6 98.4 98.4 99.6 98.5 LLaV A1.6-7B 98.0 90.8 98.8 96.8 96.8 97.2 96.0 96.4 96.4 Table 12: Accuracy of question answering in the MC2dataset when only unimodal noisy textual context is provided. Model Sport Attribute Sentiment Positional Counting Color Activit... | https://arxiv.org/abs/2505.20977v1 |
access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] See details in Suppl. 6.Experimental setting/details Question: Does the paper specify all ... | https://arxiv.org/abs/2505.20977v1 |
arXiv:2505.20981v1 [cs.CV] 27 May 2025RefA V: Towards Planning-Centric Scenario Mining Cainan Davidson1, Deva Ramanan1, Neehar Peri1 1Carnegie Mellon University Abstract Autonomous Vehicles (A Vs) collect and pseudo-label terabytes of multi-modal data localized to HD maps during normal fleet testing. However, identifyi... | https://arxiv.org/abs/2505.20981v1 |
propose RefA V , a large-scale Preprint. Figure 1: Scenario Mining Problem Setup. Given a natural language prompt such as vehicle making left turn through ego-vehicle’s path while raining , our problem setup re- quires models to determine whether the described scenario occurs within a 20-second driving log, and if so, ... | https://arxiv.org/abs/2505.20981v1 |
large-scale dataset designed to evaluate VLMs on 3D scene understanding and spatio-temporal localization. Our extensive experiments highlight the limitations of current methods, and demonstrate the effectiveness of our proposed program synthesis-based approach. Lastly, we host a challenge at CVPR 2025 in conjunction wi... | https://arxiv.org/abs/2505.20981v1 |
UP-VL [45] distills CLIP features into LiDAR data to generate amodal cuboids. Recent work uses 2D VLMs to generate 3D pseudo-labels for open-vocabulary perception, enabling zero-shot LiDAR panoptic segmentation [46, 60] and 3D object detection [22, 48, 38]. While traditional grounding methods [69] struggle with complex... | https://arxiv.org/abs/2505.20981v1 |
(e.g. car,pedestrian ,bicycle ) and do not have HD maps, making it difficult to evaluate diverse multi-agent interactions. Although A V2 annotates objects up to 150m away from the ego vehicle, we clip all object tracks at 50m. We find that current 3D perception models struggle with long-range detection and tracking [49... | https://arxiv.org/abs/2505.20981v1 |
more details about our annotation workflow. Dataset Statistics. As shown in Table 1, our dataset uniquely addresses the task of spatio-temporal scenario mining, whereas other datasets focus on object detection, referring multi-object tracking (RMOT), or visual question answering (VQA). Further, RefA V is uniquely built... | https://arxiv.org/abs/2505.20981v1 |
interactions. For a given natural language prompt, we use an LLM to parse the referred object class and only keep 3D tracks from this class. For example, for the natural language query find all cars turning left , we remove all predicted tracks except cars . Interestingly, RefBlind is a surprisingly strong baseline, ev... | https://arxiv.org/abs/2505.20981v1 |
it better aligned with human visual evaluations of tracking performance. HOTA-Temporal extends the standard HOTA metric by only considering the referred timestamps within a track as true positives. HOTA-Track is similar to HOTA-Temporal, but it does not penalize methods for incorrectly predicting the start and end of a... | https://arxiv.org/abs/2505.20981v1 |
names alone. However, RefClassify underperforms RefBlind when using predicted tracks (e.g., LE3DE2E scores 17.2 vs. 19.2 HOTA-Temporal). Finally, our proposed RefProg baseline significantly outperforms all other zero-shot methods. Notably, RefProg with LE3DE2E tracks achieves a 32.8% improvement in HOTA-Temporal over R... | https://arxiv.org/abs/2505.20981v1 |
interactions between vulnerable road users and vehicles at a crowded intersection. Scenario (b) presents an atypical instance of a common multi-agent interaction (e.g. pedestrian walking a dog ). In (c), we show a complex ego-vehicle trajectory that involves multiple moving vehicles. Scenario (d) illustrates an example... | https://arxiv.org/abs/2505.20981v1 |
IEEE/CVF conference on computer vision and pattern recognition . 2020, pp. 11621–11631. [5] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. “Segment anything in 3d with nerfs”. In: Advances in Neural Information Processing Systems 36 (2024). [6] Tzoulio Cha... | https://arxiv.org/abs/2505.20981v1 |
the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops . Jan. 2024, pp. 930–938. [20] Napat Karnchanachari, Dimitris Geromichalos, Kok Seang Tan, Nanxiang Li, Christopher Eriksen, Shakiba Yaghoubi, Noushin Mehdipour, Gianmarco Bernasconi, Whye Kit Fong, Yiluan Guo, et al. “Towards learning-b... | https://arxiv.org/abs/2505.20981v1 |
Vision . Springer. 2024, pp. 38–55. [35] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. “Segment any point cloud sequences by distilling vision foundation models”. In:Advances in Neural Information Processing Systems 36 (2023), pp. 37193–37229. [36] Zhijian Liu, Haot... | https://arxiv.org/abs/2505.20981v1 |
Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. “Learning Transferable Visual Models From Natural Language Supervision”. In: Proceedings of the 38th International Conference on Machine Learning . 12 [52] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindle... | https://arxiv.org/abs/2505.20981v1 |
base- line and new evaluation metrics”. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE. 2020, pp. 10359–10366. [67] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Ponte... | https://arxiv.org/abs/2505.20981v1 |
z> in meters) of <1.422, 0, 0.25>. This translation represents the offset from the centroid of the ego-vehicle to the ego-vehicle reference coordinate located near the center of its rear axle. All Tracks Oracle. Our naive baseline labels all ground truth tracks as referred objects . This means that all logs and timesta... | https://arxiv.org/abs/2505.20981v1 |
object segmentation. We perform all operations at the native image resolution. Next, we project all LIDAR points within each camera’s frustum onto the corresponding image plane and retain only the points that fall within the SAM2 segmentation mask. From the retained points, we compute the medoid. We then apply DBSCAN t... | https://arxiv.org/abs/2505.20981v1 |
we downsample all output tracks to 2 Hz for evaluation. B Scenario Mining Annotation Tool In order to manually annotate interesting scenarios, we build a custom web app with Claude 3.7 (cf. Fig 5). To construct a scenario, the user selects an object by clicking on a point within the image frame, writes a natural langua... | https://arxiv.org/abs/2505.20981v1 |
2 100.0 64.2 68.8 80.6 80.1 Le3DE2E Det. [8] + AB3DMOT [66] 10 78.0 48.3 49.2 73.1 73.6 Le3DE2E Det. [8] + AB3DMOT [66] 2 73.0 45.2 44.1 69.3 69.5 Valeo4Cast Det. [71] + AB3DMOT [66] 10 74.8 42.8 45.5 71.6 71.8 Valeo4Cast Det. [36] + AB3DMOT [66] 2 69.9 45.2 44.1 69.3 69.5 BEVFusion Det. [36] + AB3DMOT [66] 10 76.0 46.... | https://arxiv.org/abs/2505.20981v1 |
’, ’backward ’, ’left ’, ’right ’). 25 min_number : Minimum number of objects to identify in the 26 direction per timestamp . Defaults to 1. 27 max_number : Maximum number of objects to identify in the 28 direction per timestamp . Defaults to infinity . 29 within_distance : Maximum distance for considering an object in... | https://arxiv.org/abs/2505.20981v1 |
directory containing scenario logs and 103 data . 104 category : the category of objects to return 105 106 Returns : 107 dict : A scenario dict that where keys are the unique id ( uuid ) 108 of the object and values are the list of timestamps the object 109 is in view of the ego - vehicle . 110 111 Example : 112 trucks... | https://arxiv.org/abs/2505.20981v1 |
( vehicles , log_dir , 187 direction =’ left ’) 188 """ 189 21 190 191def has_lateral_acceleration ( 192 track_candidates :dict , 193 log_dir :Path , 194 min_accel =-np.inf , 195 max_accel =np. inf) -> dict : 196 """ 197 Objects with a lateral acceleartion between the minimum and 198 maximum thresholds . Most objects w... | https://arxiv.org/abs/2505.20981v1 |
The tracks that could be heading toward 275 another tracks 276 related_candidates : The objects to analyze to see if the 277 track_candidates are heading toward 278 log_dir : Path to the directory containing scenario logs 279 and data . 280 angle_threshold : The maximum angular difference between 281 the velocity vecto... | https://arxiv.org/abs/2505.20981v1 |
candidate must 368 be from the pedestrian crossing . A distance of zero means 369 that the object must be within the boundaries of the 370 pedestrian crossing . 371 372 Returns : 373 Filtered scenario dictionary where keys are track UUIDs 374 and values are lists of timestamps . 375 376 Example : 377 vehicles_at_ped_cr... | https://arxiv.org/abs/2505.20981v1 |
(e.g., pedestrians 466 or vehicles ) to check for crossings . 467 log_dir : Path to scenario logs . 468 direction : specifies the axis and direction the 469 half midplane extends from 470 in_direction : which direction the related candidate has 471 to cross the midplane for it to be considered a crossing 472 forward_th... | https://arxiv.org/abs/2505.20981v1 |
that the positive tracks are 553 traveling in relative to the related candidates 554 " opposite " indicates the track candidates are traveling in 555 a direction 556 135 -180 degrees from the direction the related candidates 557 are heading toward . 558 " same " indicates the track candidates that are traveling 559 in ... | https://arxiv.org/abs/2505.20981v1 |
is in 637 a drivable area . 638 639 Example : 640 buses_in_drivable_area = in_drivable_area (buses , log_dir ) 641 """ 642 28 643 644def on_road ( 645 track_candidates :dict , 646 log_dir : Path ) ->dict : 647 """ 648 Identifies objects that are on a road or bike lane . 649 This function should be used in place of in_d... | https://arxiv.org/abs/2505.20981v1 |
, 737 peds_not_on_pedestrian_crossing ]) 738 """ 739 740 741def scenario_or ( scenario_dicts : list [ dict ]): 742 """ 743 Returns a composed scenario where that tracks all objects and 744 relationships in all of the input scenario dicts . 745 746 Args : 747 scenario_dicts : the scenarios to combine 748 749 Returns : 7... | https://arxiv.org/abs/2505.20981v1 |
arXiv:2505.20995v1 [cs.CL] 27 May 2025Articulatory strategy in vowel production as a basis for speaker discrimination Justin J. H. Lo1, Patrycja Strycharczuk2, Sam Kirkham1 1Linguistics and English Language, Lancaster University, United Kingdom 2Linguistics and English Language, University of Manchester, United Kingdom... | https://arxiv.org/abs/2505.20995v1 |
Establishing the individuality of such dif- ferences could contribute to the explanations of the phonetic sources of individuality in human speech. To this end, we investigate whether tongue shape features in vowel production can be the basis of successful speaker dis- crimination. V owels have long been of considerabl... | https://arxiv.org/abs/2505.20995v1 |
pre-trained unidirec- tional MobileNet1.0 model, version 1.1.0, to automatically la- bel 11 points corresponding to key anatomical landmarks on the midsagittal tongue contour, which were exported as Cartesian coordinates (in mm) and rotated on the occlusal plane. For the purpose of the current paper, we focus on the da... | https://arxiv.org/abs/2505.20995v1 |
SDs higher (solid red) and lower (dashed blue). The effects of the first three PCs in the shape analysis are summarised in Figure 2. All three PCs can be characterised as representing variation in the degree of tongue curvature, centred at progressively more anterior parts of the tongue. PC1 targets the posterior part ... | https://arxiv.org/abs/2505.20995v1 |
as previously men- tioned, high PC3 scores represent a tongue shape in which the anterior part of the tongue is concave and the dorsum is raised, seen, for instance, in a velarised [ ë]. Example data in [24] sug- gest that, in liquid production, this feature can be somewhat idiosyncratic. Assuming that similar variatio... | https://arxiv.org/abs/2505.20995v1 |
EER and Cllr, while PC3 slightly outperformed PC2 with a lower Cllr. Combinations of two PCs resulted in im- proved system performance over their individual components, with the combination of PC1 and PC3 producing the lowest EER and Cllr, followed closely by the combination of PC2 and PC3. However, the system with all... | https://arxiv.org/abs/2505.20995v1 |
root articulation are both present in our data. Figure 4 provides a clear illustration with two speakers with the most extreme mean size-and-shape PC2 scores. It does appear to be the case, however, that there is greater scope for individuality in the shape of the anterior part of the tongue. We believe this can be att... | https://arxiv.org/abs/2505.20995v1 |
speaker discrimination. These features, in- dividually or in combination, perform at a level on par with LTFDs [31, 32], which also aim to capture speaker specificity in articulatory setting. While the controlled nature of our materi- als and differences in data modelling techniques mean that such a comparison needs to... | https://arxiv.org/abs/2505.20995v1 |
C. Grigoras, “A case for formant analysis in forensic speaker identification,” Int. J. Speech Lang. Law , vol. 12, no. 2, pp. 143–173, 2005. [16] K. N. Stevens and A. S. House, “Development of a quantitative description of vowel articulation,” J. Acoust. Soc. Am. , vol. 27, pp. 484–493, 1955. [17] G. Fant, Acoustic The... | https://arxiv.org/abs/2505.20995v1 |
arXiv:2505.21003v1 [cs.CL] 27 May 2025Uncertainty Unveiled: Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models? Yifei Wang1,2, Yu Sheng1,2, Linjing Li1,2*, Daniel Zeng1,2 1State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of... | https://arxiv.org/abs/2505.21003v1 |
examine the impact of increasing scales of in-context examples on LLMs’ confidence through faithful uncertainty quantification (UQ) approaches. By incorporat- ing model parameters, configurations, and various demonstration sets, we approximate the predictive distribution in the output space. Then we compute entropy to ... | https://arxiv.org/abs/2505.21003v1 |
and uncertainty in its prediction, called total uncertainty. Total uncertainty can be decomposed into two key components: epistemic (model) uncertainty and aleatoric (data) uncertainty (Hou et al., 2024; Valdenegro-Toro and Mori, 2022). The advent of large language models (LLMs) has introduced new challenges in quantif... | https://arxiv.org/abs/2505.21003v1 |
over the output space, we obtain L×|Y| probability matrix AL×|Y|, from which wecompute the entropy as follows: TU=−H σ "LX l=1P(y|x, zl 1:N)# y∈|Y| where σis a normalization function that ensures the sum of probabilities equals one, and H=P ip(x)log(p(x)). Some studies indicate that log- its may be uncalibrat... | https://arxiv.org/abs/2505.21003v1 |
5.98 𝐙𝐙𝟏𝟏:𝐍𝐍𝟑𝟑 0.00 0.00 0.00 0.00 6.42 𝐙𝐙𝟏𝟏:𝐍𝐍𝟒𝟒 0.00 0.00 0.00 0.00 6.19 𝐙𝐙𝟏𝟏:𝐍𝐍𝟓𝟓 0.00 0.00 0.00 0.00 7.32 𝐙𝐙𝟏𝟏:𝐍𝐍𝟔𝟔 0.00 0.00 0.00 0.98 5.09EU TUFigure 3: A workflow for uncertainty quantification and decomposition under many-shot ICL settings, involves the following components: a LL... | https://arxiv.org/abs/2505.21003v1 |
reduces LLMs’ uncertainty across models and datasets. As shown in Figs.4 and 5, the results 1Model weights are loaded at float16 precision. 2the Area Under the Receiver Operating Characteristic 2021222324252627 Llama-3.1-8B0.000.250.500.751.00TU 2021222324252627 Mistral-7B-v0.20.000.250.500.751.00 2021222324252627 Qwen... | https://arxiv.org/abs/2505.21003v1 |
under k-shot ICL. Despite 8.65% of cases experi- encing heightened uncertainty with longer inputs in 128-shot learning, this effect minimally impacts overall model performance, as reflected by the small absolute values of ∆Acc. Crucially, the tran- sition from few-shot (e.g., 4-shot) to many-shot ICL demonstrates a mar... | https://arxiv.org/abs/2505.21003v1 |
difficulty for current large models in effectively comprehending long texts with complex structures. Additional Information Reduces EU. To val- idate that additional examples enhance the infor- mational content and yield a clearer βfor models (as shown in Fig.7), we observe that only diverse examples effectively reduce... | https://arxiv.org/abs/2505.21003v1 |
simplicity, detailed computations within theMHSA sublayer, such as the projection matrices WQ,K,V,O , and the splitting-merging operations across attention heads, are omitted here. Each de- coder block, therefore, maintains two distinct resid- ual pathways: one emerging from the MHSA, h(l) i, and the other from MLP sub... | https://arxiv.org/abs/2505.21003v1 |
answer from about 22nd layer onward, maintaining this highest probability thereafter. Si- multaneously, it demonstrates robustness by main- taining near-zero probabilities for incorrect options, effectively eliminating the influence of distractors on the model’s final prediction. CMQA 4-shot 32-shot 64-shot 128-shot Ll... | https://arxiv.org/abs/2505.21003v1 |
scenarios. Nevertheless, applying ICL to rationale- intensive reasoning and generative contexts re- mains a promising direction. Future investigations should assess the reliability and trustworthiness of ICL in these domains, as advancements in this area could not only enhance task-solving performance but also broaden ... | https://arxiv.org/abs/2505.21003v1 |
generative LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 7752–7767, Bangkok, Thailand. Association for Computational Linguistics. Nora Belrose, Zach Furman, Logan Smith, Danny Ha- lawi, Igor Ostrovsky, Lev McKinney, Stella Bider- man, ... | https://arxiv.org/abs/2505.21003v1 |
Punta Cana, Dominican Republic. Association for Computational Linguistics. Albert Gu and Tri Dao. 2024. Mamba: Linear-time sequence modeling with selective state spaces. Tahmid Hasan, Abhik Bhattacharjee, Md. Saiful Is- lam, Kazi Mubasshir, Yuan-Fang Li, Yong-Bin Kang, M. Sohel Rahman, and Rifat Shahriyar. 2021. XL- su... | https://arxiv.org/abs/2505.21003v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.