text string | source string |
|---|---|
these intructions, we run the motion knowledge ground- ing pipeline on both commercial and open-source LLMs including Claude 3.5 Sonnet, GPT-4o, GPT- 4o-mini, GPT-3.5-turbo and Llama-3.1-70B.3 High-level Planning. To evaluate LLMs’ abil- ities to conceptually decompose motion instruc- tions, we introduce High-level Pla... | https://arxiv.org/abs/2505.21531v1 |
knowledge is more effectively accessed through incremental guidance rather than one-round generation. LLMHPS (piece_by_piece )HPS (in_one_go ) Claude 3.5 Sonnet 4.57 4.42 GPT-4o 4.68 4.55 GPT-4o-mini 4.67 3.93 GPT-3.5-turbo 3.50 3.33 Llama-3.1-70B 4.07 - Table 1: HPS for each tested LLM across two high-level planning s... | https://arxiv.org/abs/2505.21531v1 |
scores when animation quality falls below a cer- tain threshold. The best performing LLM, Claude 3.5 Sonnet, scores 1.28 points below the oracle animations, indicating considerable room for im- provement in motion understanding. The averaged kappa of 0.531 corresponds to moderate agreement according to Landis and Koch ... | https://arxiv.org/abs/2505.21531v1 |
through quanti- tative error analysis and representative case studies. 6.1 High-level Planning To better understand how LLMs perform differ- ently in HPS (Table 1), we count the numbers ofhigh-level plans with wrong or incomplete action descriptions generated using the piece_by_piece strategy (Table 5). The ranking of ... | https://arxiv.org/abs/2505.21531v1 |
Complexity. LLM performance in posi- tion prediction declines as motions become more complex with increased steps and body parts (Fig- ure A12). This degradation likely stems from two factors: LLMs’ difficulty in maintaining spatial re- lationships across extended movement sequences, and training data bias where humans... | https://arxiv.org/abs/2505.21531v1 |
al., 2024). bow (Figure 6f). In contrast, MoMask produces similar general bowing motions for both instruc- tions, suggesting its limited cultural understanding. Generating Raw Avatar Control Parameters. We further examine LLMs’ capabilities to directly generate SMPL control parameters, with prompt- ing strategies illus... | https://arxiv.org/abs/2505.21531v1 |
ethical consider- ations around consent and representation, as such systems could potentially reproduce and amplify biases in human movement patterns. Addition- ally, as these technologies become more sophisti- cated, there may be privacy concerns regarding the capture and reproduction of distinctive individual movemen... | https://arxiv.org/abs/2505.21531v1 |
Trans. Graphics (Proc. SIGGRAPH Asia) , 34(6):248:1–248:16. 9 Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. 2023. Human motion diffusion model. In The Eleventh In- ternational Conference on Learning Representations . Zhikai Zhang, Yitang Li, Haofeng Huang, Mingxian Lin, and L... | https://arxiv.org/abs/2505.21531v1 |
forward_to_upward_midline {m_avg_L_Collar: [0.0, 15.0, -10.0], m_avg_L_Shoulder: [0.0, 110.0, -30.0]} side_to_upward_back {m_avg_L_Collar: [0.0, -15.0, -10.0], m_avg_L_Shoulder: [0.0, -30.0, -30.0]} neutral_to_forward_midline {m_avg_L_Collar: [0.0, 0.0, 10.0], m_avg_L_Shoulder: [-45.0, 0.0, 100.0]} neutral_to_forward_s... | https://arxiv.org/abs/2505.21531v1 |
{m_avg_R_Knee: [45.0, 0.0, 0.0]} bent_at_90_degrees {m_avg_R_Knee: [90.0, 0.0, 0.0]} fully_bent {m_avg_R_Knee: [135.0, 0.0, 0.0]} LeftAnkleneutral {m_avg_L_Ankle: [0.0, 0.0, 0.0]} bent_upward {m_avg_L_Ankle: [-20.0, 0.0, 0.0]} bent_downward {m_avg_L_Ankle: [45.0, 0.0, 0.0]} tilted_inward {m_avg_L_Ankle: [0.0, 0.0, 30.0... | https://arxiv.org/abs/2505.21531v1 |
plan? <reflection_judgement> Do you think there 's need to replan this body part in order to achieve the goal final state in the high-level plan? Give your judgement. <correction> You think that: {reflection}. So the next position of {body part} should not be **{position}**.\newline Based on the thought, replan this bo... | https://arxiv.org/abs/2505.21531v1 |
to ask the nested question. "Head": { "question": "At the end of this step, is the head upright in the neutral position, tilted left, tilted right, tilted down, tilted up, turned left or turned right? Choose one from %s", "options": { "neutral": "neutral", "tilted_left": { "question": "Is the head tilted left slightly ... | https://arxiv.org/abs/2505.21531v1 |
Choose one from %s", "options": { "front": { "question": "Relative to the torso, is the left upper arm above the left shoulder, below the left shoulder, or neither? Choose one from %s", "options": { "above": "forward_to_upward_side", "below": "neutral_to_forward_side", "neither": "forward_to_side" } }, "behind": { "que... | https://arxiv.org/abs/2505.21531v1 |
right upper arm above the right shoulder, below the right shoulder, or neither? Choose one from %s", "options": { "above": "forward_to_upward_midline", "below": "neutral_to_forward_midline", "neither": "forward_to_midline" } }, "neither": { "question": "Relative to the torso, is the right upper arm in front of the body... | https://arxiv.org/abs/2505.21531v1 |
"LeftUpperLeg": { "question": "At the end of this step, is the left upper leg neutrally aligned with the body midline, straight forward, straight out to the side forming a right angle with the torso, or in other in- between positions? Choose one from %s", "options": { "neutral": "neutral", "forward": "forward", "side":... | https://arxiv.org/abs/2505.21531v1 |
"Relative to the torso, is the right upper leg in front of the body or behind the body? Choose one from %s", "options": { "front": { "question": "Relative to the torso, is the right upper leg above the right pelvis or below the right pelvis? Choose one from %s", "options": { "above": "forward_to_upward", "below": "neut... | https://arxiv.org/abs/2505.21531v1 |
"question": "At the end of this step, are the left toes in the neutral position or curled? Choose one from %s", "options": { "neutral": "neutral", "curled": { "question": "Are the left toes curled up or down? Choose one from %s", "options": { "curled_up": "curled_up", "curled_down": "curled_down" } } } }, "RightToes": ... | https://arxiv.org/abs/2505.21531v1 |
torso maintain a straight posture. twisted_left_slightly The left side of the waist moves slightly backward, and the right side moves slightly forward, forming a small angle to the left relative to the feet. The muscles on the right side extend mildly, while those on the left side contract slightly. twisted_left_fully ... | https://arxiv.org/abs/2505.21531v1 |
the torso. upward The left upper arm is lifted straight upwards, close to the ear, reaching towards the sky. side_elbowpit_forward The left upper arm is extended straight out to the side, forming a right angle with the torso (horizontally aligned with the shoulders). The elbow pit faces forward. side_elbowpit_upward Th... | https://arxiv.org/abs/2505.21531v1 |
ear, reaching towards the sky. side_elbowpit_forward The right upper arm is extended straight out to the side, forming a right angle with the torso (horizontally aligned with the shoulders). The elbow pit faces forward. side_elbowpit_upward The right upper arm is extended straight out to the side, forming a right angle... | https://arxiv.org/abs/2505.21531v1 |
very close to or touching the left shoulder; the muscles in the left upper arm are fully contracted. RightElbow neutral The right elbow is extended naturally, forming a straight line from the right shoulder to the right wrist. The right upper arm and right forearm create a nearly straight alignment of about 180 degrees... | https://arxiv.org/abs/2505.21531v1 |
forearm. The muscles on the thumb side of the right forearm contract slightly. tilted_towards_pinky_side The right wrist tilts laterally to form a small angle, less than 20 degrees, moving the little finger side of the right hand closer to the right forearm. The muscles on the pinky side of the right forearm contract s... | https://arxiv.org/abs/2505.21531v1 |
from the forward position. neutral_to_side The right upperleg is lifted from the neutral position and extended to the side, in the middle of neutral and side positions. neutral_to_forward_side The right upperleg is lifted from the neutral position and extended partly forward and slightly to the side, forming a diagonal... | https://arxiv.org/abs/2505.21531v1 |
The top of the left foot forms an acute angle, less than 90 degrees, with the left shin. The left toes point closer to the left shin; the muscles on the front of the left lower leg contract to lift the left toes. bent_downward The top of the left foot forms a steep obtuse angle, around 180 degrees, with the left shin. ... | https://arxiv.org/abs/2505.21531v1 |
contract to lift the right toes. curled_down The right toes curl downward, forming small downward angles with the sole of the right foot. The muscles on the bottom of the right foot and right toes contract to curl the right toes. 28 A.3 Motion Instructions We create twenty motion instructions for the main experiments. ... | https://arxiv.org/abs/2505.21531v1 |
graduate students, technical staff or researchers working on artificial intelligence at the same univer- sity, and participate voluntarily with above-average- wage compensation. Tasks are designed to be safe and unbiased, with clear instructions and reason- able time commitments. Participants are informed of the study’... | https://arxiv.org/abs/2505.21531v1 |
human evaluator to read. This page of the document is followed by evaluation rubrics and 6 examples covering different WBS and BPQ. Figure A6: Sample form in human evaluation of the animations from the complete generation. At the header of each form, we provide a link to the illustrative docu- ment shown in Figure A5. ... | https://arxiv.org/abs/2505.21531v1 |
3.00 (0.00, 0.00) 4.67 (0.47, 0.22) 4.67 (0.47, 0.22) 1.67 (0.47, 0.22) 3.33 (0.47, 0.22) 14 5.00 (0.00, 0.00) 4.67 (0.47, 0.22) 5.00 (0.00, 0.00) 3.00 (0.00, 0.00) 4.67 (0.47, 0.22) 15 5.00 (0.00, 0.00) 4.33 (0.47, 0.22) 5.00 (0.00, 0.00) 4.33 (0.47, 0.22) 2.67 (0.47, 0.22) 16 4.33 (0.47, 0.22) 5.00 (0.00, 0.00) 3.67 ... | https://arxiv.org/abs/2505.21531v1 |
0.00) 0.4896 (0.01, 0.00) 5 0.7812 (0.00, 0.00) 0.7812 (0.00, 0.00) 0.6328 (0.01, 0.00) 0.6797 (0.05, 0.00) 0.5079 (0.09, 0.01) 6 0.8938 (0.01, 0.00) 0.8625 (0.01, 0.00) 0.7875 (0.01, 0.00) 0.7875 (0.00, 0.00) 0.7375 (0.01, 0.00) 7 0.7916 (0.02, 0.00) 0.8021 (0.01, 0.00) 0.7396 (0.01, 0.00) 0.8646 (0.01, 0.00) 0.7812 (... | https://arxiv.org/abs/2505.21531v1 |
0.00) 0.5209 (0.04, 0.00) 15 0.7000 (0.06, 0.00) 0.6562 (0.04, 0.00) 0.4937 (0.02, 0.00) 0.3562 (0.01, 0.00) 0.3938 (0.06, 0.00) 16 0.4609 (0.04, 0.00) 0.5468 (0.03, 0.00) 0.5312 (0.03, 0.00) 0.4063 (0.06, 0.00) 0.5469 (0.02, 0.00) 17 0.6250 (0.08, 0.01) 0.5625 (0.00, 0.00) 0.5625 (0.02, 0.00) 0.5625 (0.02, 0.00) 0.343... | https://arxiv.org/abs/2505.21531v1 |
(0.40, 0.16) 2.80 (0.75, 0.56) 1.80 (0.40, 0.16) 1.60 (0.80, 0.64) 5.00 (0.00, 0.00) 3 4.00 (0.00, 0.00) 1.00 (0.00, 0.00) 2.60 (0.49, 0.24) 1.00 (0.00, 0.00) 2.40 (1.02, 1.04) 5.00 (0.00, 0.00) 4 2.60 (1.50, 2.24) 3.20 (0.40, 0.16) 3.60 (1.02, 1.04) 1.00 (0.00, 0.00) 1.40 (0.49, 0.24) 4.60 (0.49, 0.24) 5 3.40 (1.20, 1... | https://arxiv.org/abs/2505.21531v1 |
Lifted in the air\n2. Hands: Holding the right shoe" } ] Figure A10: Example of incomplete action descriptions, generated by GPT-4o. In this example, the plan does not specify the action of the left arm in order for the eyes to see the watch. [motion instruction] look down to check the time of the watch on the left wri... | https://arxiv.org/abs/2505.21531v1 |
and begins moving forward." }, { "step_number": 2, "time_range": [1, 2.5], "movement": "Swing the golf club downwards and across the front body, moving towards the left side.", "initial_state": "The golf club is raised upwards and twisting forward.", "final_state": "The golf club is positioned over the left shoulder, w... | https://arxiv.org/abs/2505.21531v1 |
arXiv:2505.21544v1 [cs.CV] 24 May 2025Vision Meets Language: A RAG-Augmented YOLOv8 Framework for Coffee Disease Diagnosis and Farmer Assistance Semanto Mondal1* 1*University of Naples Federico II, Via Cinthia 21, Italy, 80126, Naples, Italy. Corresponding author(s). E-mail(s): semanto.mondal@unina.it; Abstract As a so... | https://arxiv.org/abs/2505.21544v1 |
Farmers are utilising different spraying techniques to utilise pesticides which can lead to spray drift and losses of pesticides into surface waters and the environment. On the other hand, pesticide application also results in residues, which are an acute food-related concern. More than 45% of food products are found t... | https://arxiv.org/abs/2505.21544v1 |
neurosymbolic approach where yolo and LLM are the neural components, and the external knowledge-based is the symbolic component, which provides context to the LLM to generate guided feedback. 2 Related Work The integration of deep learning techniques into agriculture has significantly advanced plant disease detection, ... | https://arxiv.org/abs/2505.21544v1 |
RAG, and soft-prompting. Their study revealed that RAG approaches outperformed both unmodified and fine-tuned GPT-3.5 models, especially when combined with basic soft prompts. This underscores the effectiveness of RAG in improving LLM outputs, even with limited datasets and technical expertise. The use of large languag... | https://arxiv.org/abs/2505.21544v1 |
named retrieval. These text chunks are used as prompts to the LLM model. Based on the context and user query, the LLM model can create a more precise and creative answer, which can be referred to as generation. This is not possible with other types of chatbots. The working of the RAG-based chatbot can be divided into t... | https://arxiv.org/abs/2505.21544v1 |
pipeline. When a user uploads an image through the Streamlit interface, it is passed directly to the trained YOLOv8 model for inference. The model outputs the predicted disease label(s) along with the corresponding bounding boxes, which are then visualised on the frontend. These predicted disease classes serve as input... | https://arxiv.org/abs/2505.21544v1 |
For this work, we have trained the YOLOV8 model using both versions of the dataset and observed that the best performance was obtained when the model was trained using the original version of the dataset. 5.2 Dataset Overview The BRACOL dataset comprises 1,747 images of Arabica coffee leaves, capturing both healthy spe... | https://arxiv.org/abs/2505.21544v1 |
Remedy Generation and Follow-up Questions After a leaf image is uploaded through the Streamlit-based user interface (Figure 4a), the system first performs disease detection using the YOLOv8 model (Figure 4b). If a disease is detected, the predicted class label is used to form a query that is passed to the RAG (Retrieva... | https://arxiv.org/abs/2505.21544v1 |
L.C.C.: The 2021 european union report on pesticide residues in food. efsa (2023) [3] Berger, L.T., Doll, D., Schwitzky, E., Lavelle, K., Skalsky, M., Spinelli, F.: Sus- tainable ways to reduce pesticides in pome and stone fruit production. EIP (2022) [4] Ultralytics: Models supported by ultralytics. Ultralytics (2024)... | https://arxiv.org/abs/2505.21544v1 |
arXiv:2505.21548v1 [physics.soc-ph] 25 May 2025Fluent but Culturally Distant: Can Regional Training Teach Cultural Understanding? Dhruv Agarwal Cornell University da399@cornell.eduAnya Shukla Cornell University as2933@cornell.edu Sunayana Sitaram Microsoft Research Sunayana.Sitaram@microsoft.comAditya Vashistha Cornell... | https://arxiv.org/abs/2505.21548v1 |
with the cultural com- petence required to serve their intended users? In- sights from this question can inform the design of future multilingual and multicultural training regimes. We address this question through a detailed case study of India as a non-Western culture. We com- pare several Indic LLMs—models explicitl... | https://arxiv.org/abs/2505.21548v1 |
systems (Arora et al., 2023; Masoud et al., 2025). For example, Tao et al. (2024) find a strong Anglo- Protestant drift in GPT-family models. Recent mul- tilingual vision–language benchmarks such as M5 (Schneider and Sitaram, 2024) extend these assess- ments to multimodal settings. Efforts towards cultural alignment. I... | https://arxiv.org/abs/2505.21548v1 |
to show a form of exasperation? A. yah B. yo C. eh D. leh. We filter this dataset to include 46 India-specific and 20 US-specific questions. Each model’s accuracy is measured on both sets. Cultural Adaptation. While CulturalBench di- rectly probes raw cultural knowledge in a model, NormAd (Rao et al., 2024) tests a mod... | https://arxiv.org/abs/2505.21548v1 |
models with poor instruction- following, added explicit formatting instructions.2 GlobalOpinionQA. We follow the methodology of Durmus et al. (2023) to measure how closely a model’s opinion distribution aligns with that of Indian and American respondents. LetQ=q1, . . . , q nbe the set of questions and Oq=o1, . . . , o... | https://arxiv.org/abs/2505.21548v1 |
human baseline. Impact of prompting strategies. Figure 6 in Ap- pendix A.2 compares each model’s Euclidean dis- tance to India under the three prompting strategies. Demographic prompting pulls global models 15% closer to India on average, yet Indic models are relatively unaffected (only 3% closer). This sug- gests that... | https://arxiv.org/abs/2505.21548v1 |
experiment requires accessing raw logits, we could not run it for GPT-4O.So nCAD tells us “ what fraction of the possible tilt did the model use? ” Values fall neatly in [−1,1] where +1 means perfect Indian alignment, –1 per- fect US alignment, 0 neutrality. Figure 2(b) plots mean nCAD. Allmodels now show a statistical... | https://arxiv.org/abs/2505.21548v1 |
reduces to 15.8% (14% for Indic models). While Cross-lingual prompting further reduces it to 13.6% (11%), this improvement is offset by lower accu- racy on both cultures, indicating no real gain in Indian knowledge. Thus, our prompting strategies— designed to give models a fair chance at eliciting cultural knowledge–ar... | https://arxiv.org/abs/2505.21548v1 |
the performance disparity being statistically significant for most models (two proportions z-test with Bonferroni correction). Indic models struggle with cultural reasoning. In the two reasoning-heavy settings ( Country and Value+Country ), most Indic models perform only slightly better than random chance on Indian que... | https://arxiv.org/abs/2505.21548v1 |
supervised fine-tuning (SFT) on local-language data leaves cultural knowledge un- changed or degraded (Table 1). This suggests that fine-tuning may obscure subtle cultural priors that the base models might have acquired. Technical reports reveal that almost every Indic model relies on instruction sets that are either t... | https://arxiv.org/abs/2505.21548v1 |
training objectives that prioritize cultural grounding alongside linguistic ability. Without such efforts, regional models risk replicating the very biases they aim to avoid. 6https://www.karya.in Limitations Our analysis is constrained by the availability of regional models. DeepSeek is the only frontier- scale sovere... | https://arxiv.org/abs/2505.21548v1 |
Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, and Bi Puranen. 2022. World values survey (1981-2022). trend file. Shreya Havaldar, Sunny Rai, Bhumika Singhal, Langchen Liu, Sharath Chandra Guntuku, and Lyle Ungar. 2023. Multilingual language models are not multicultural: A case study in emotion. Preprint , ar... | https://arxiv.org/abs/2505.21548v1 |
Qadri, Renee Shelby, Cynthia L. Bennett, and Emily Denton. 2023. Ai’s regimes of representation: A community-centered study of text-to-image models in south asia. In 2023 ACM Conference on Fairness, Accountability, and Transparency , FAccT ’23, page 506–517. ACM.Abhinav Rao, Akhila Yerukola, Vishwa Shah, Katha- rina Re... | https://arxiv.org/abs/2505.21548v1 |
thing. In fact, it might be desirable for a global model not to be aligned to any culture than to be strongly aligned to one specific culture.Is drinking alcohol morally acceptable, morally unac- ceptable, or not a moral issue? Answer Option India USA (A) Morally acceptable 10% 32% (B) Morally unacceptable 70% 16% (C) ... | https://arxiv.org/abs/2505.21548v1 |
arXiv:2505.21549v1 [cs.CV] 25 May 2025Distill CLIP (DCLIP): Enhancing Image-Text Retrieval via Cross-Modal Transformer Distillation Daniel Csizmadia1∗Andrei Codreanu1Victor Sim1Vighnesh Prabeau1 Michael Lu1†Kevin Zhu1†Sean O’Brien1†Vasu Sharma1,2† 1Algoverse AI Research2Carnegie Mellon University csizmadiadaniel@vt.edu... | https://arxiv.org/abs/2505.21549v1 |
models. Our contributions are: 1) a lightweight distillation framework using YOLO-based region features to improve visual grounding without full retraining, 2) an asymmetric student-teacher architecture that preserves CLIP’s zero-shot space while improving retrieval performance, and 3) empirical results showing strong ... | https://arxiv.org/abs/2505.21549v1 |
of the teacher’s characteristics and training process is also highlighted by research em- phasizing teacher consistency [ 36] and perspectives viewing KD as a teacher-student co-optimization problem [ 69]. DCLIP’s approach aligns with this philosophy: its "meta-teacher" is not an off-the-shelf model but is specifically... | https://arxiv.org/abs/2505.21549v1 |
semantically enriched embeddings with bidirectional attention and region-level supervision. With the finetuned cross modal attention in the meta teacher, we create a semantically aligned distillation target for the student. However, using this teacher directly would require paired image-text inputs at inference time, w... | https://arxiv.org/abs/2505.21549v1 |
loss plays a pivotal role in preserving the original CLIP structure by ensuring the student does not overfit to the teacher’s embeddings. Without this regularization, the student would risk collapsing to a modality-dependent encoder that sacrifices CLIP’s zero-shot flexibility. The cosine losses guide the student to ad... | https://arxiv.org/abs/2505.21549v1 |
quality by computing the average precision for each query and averaging over all queries. It rewards models that rank relevant items higher. Zero-Shot Classification. Zero-shot classification assesses the model’s ability to generalize to unseen classes without additional training. We report Top-1 and Top-5 accuracy on ... | https://arxiv.org/abs/2505.21549v1 |
and a straightforward aggregation strategy, it consistently adapts large vision models to specialized retrieval tasks while maintaining strong out-of-domain generalization. ModelImageNet-1K CIFAR-10 CIFAR-100 Top-1 Acc (%) Top-5 Acc (%) Top-1 Acc (%) Top-5 Acc (%) Top-1 Acc (%) Top-5 Acc (%) ViT-B/32 (Base CLIP) 60.08 ... | https://arxiv.org/abs/2505.21549v1 |
trade-off in Figure 4. 0.1 0.2 0.3 0.4 MSCOCO T I Recall@1 0.350.400.450.500.550.600.65ImageNet Zero-Shot Top-1 Accuracy Zero-Shot Classification vs. Text-Image Retrieval Pareto Curve 1 2 3 4 5 67 8 DCLIP (training progression) CLIP ViT-B/16 CLIPSelf ViT-B/16 RegionCLIP FineCLIP 2.5M TinyCLIP 8M Approximate Pareto Fron... | https://arxiv.org/abs/2505.21549v1 |
and cross-attention are retained. This indicates that smaller models benefit from strong external guidance to shape their embedding space effectively. Conversely, our findings with ViT-L/14 reveal that such guidance can become a liability: large models tend to overfit quickly to their teacher, limiting generalization. ... | https://arxiv.org/abs/2505.21549v1 |
learning. Overall, DCLIP demonstrates that careful distillation of spatial and semantic knowledge from CLIP can yield lightweight yet competitive alternatives for real-world deployment. 9 Acknowledgments and Disclosure of Funding Use unnumbered first level headings for the acknowledgments. All acknowledgments go at the... | https://arxiv.org/abs/2505.21549v1 |
Makhzani, A., Kramar, M., V orontsov, E., and Miotto, R. (2023). Adapting Large Vision-Language Models to Medical Image Understanding. arXiv preprint arXiv:2304.07193 . [14] Federico, F., Deza, A., Kreiman, G., and Vaziri, A. (2024). Optimal visual representations for generative text-to-image models: Lessons from a com... | https://arxiv.org/abs/2505.21549v1 |
European conference on computer vision , pages 740–755. Springer, Cham. [28] Young, P., Lai, A., Hodosh, M., and Hockenmaier, J. (2014). From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics , 2... | https://arxiv.org/abs/2505.21549v1 |
H., Miller, J., Hajishirzi, H., Farhadi, A., and Schmidt, L. (2021). OpenCLIP. Zenodo . https://doi.org/10.5281/zenodo.5143773 . [45] Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y ., and Gao, J. (2019). ERNIE-ViL: Knowledge Enhanced Vision-Language Pre-training for Visual Question Answering and Imag... | https://arxiv.org/abs/2505.21549v1 |
., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., and Gao, J. (2022). Grounded Language-Image Pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 10965–10975. [59] Zhong, Y ., Li, L.H., Zhang, P., Yang, J., Li, C., Yuan, L., Zhang, L., and Gao, J. ... | https://arxiv.org/abs/2505.21549v1 |
Contrastive Contrastive Student Loss Contrastive + Cosine Contrastive + Cosine + ZS Anchor Loss Batch Size 32 32 Hidden Dimension 512 768 Table 6: Comparison of training hyperparameters and architectural choices between ViT-B and ViT-L during DCLIP distillation. Aggregation method refers to how region embeddings are av... | https://arxiv.org/abs/2505.21549v1 |
end-to-end—including fine-tuning of the student and teacher—on a single NVIDIA RTX 2070 Super (8 GB) in roughly 1.5 hours per epoch. Even the larger ViT-L/14 models train in about 2 hours per epoch on an NVIDIA RTX 5090 (32 GB). These results show that DCLIP’s meta-teacher distillation can be conducted within reasonabl... | https://arxiv.org/abs/2505.21549v1 |
used. D Evaluation Metric Definitions (From Section 5.1 of provided document) Recall@K Recall@K measures the proportion of queries for which the correct match appears in the top-K retrieved results. Recall@ K=1 NNX i=1Success@K (i) where Nis the number of queries, and Success@K (i) = 1 if the correct item for query iis... | https://arxiv.org/abs/2505.21549v1 |
In previous experiments we utilized a image text matching(ITM) loss in both the student and the teacher. The ITM loss, utilized by BLIP, creates a binary classification head that is trained based on hard negatives that allows the classifier to make decisions on if it thinks that caption belongs with that specific image... | https://arxiv.org/abs/2505.21549v1 |
arXiv:2505.21578v1 [cs.CL] 27 May 2025Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use Titouan Parcollet, Yuan Tseng, Shucong Zhang, Rogier van Dalen AI Center Cambridge, Samsung, United Kingdom {t.parcollet,s1.zhang,r.vandalen }@samsung.com Abstrac... | https://arxiv.org/abs/2505.21578v1 |
of the data unreli- able for training and evaluation. SpeechStew [10], OWSM [11], and Gigaspeech [12] are covered by non-commercial licenses or copyrights forbidding industrial research. Common V oice [13] is too small on its own. Finally, most of these datasets also do not offer any validation or test sets for proper ... | https://arxiv.org/abs/2505.21578v1 |
3.0 50 500 6,100 0 1.5 1.5 Libriheavy Read, clean, native 50,000 CC BY 4.0 50 500 11,000 11,000 0 0 LibriSpeech Read, clean, native 960 CC BY 4.0 0 0 0 0 5 5 Total hours 128,110 250 2,500 25,150 13,150 16.5 16.5 acoustic conditions with standardised text normalisation and carefully crafted validation and testing sets. ... | https://arxiv.org/abs/2505.21578v1 |
of which permit commercial use (see Ta- ble 1). The code to reproduce the dataset is distributed under the Apache 2.0 license. 3.1. Audio data selection and corpus structure The Loquacious set contains four training splits and two evaluation sets: small (250 hours), medium (2,500 hours), large (25,000 hours), clean (13... | https://arxiv.org/abs/2505.21578v1 |
babble noise and reverberation. Half of the dataset can be con- sidered as clean speech as V oxPopuli and Libriheavy have clean recording conditions. V oxPopuli adds heavily accented speech to it (EU parliament with native and non-native speakers). The other half of the dataset can be considered as potentially noisy. I... | https://arxiv.org/abs/2505.21578v1 |
already almost contains one hundred thou- sand speakers, it is fair to assume that the real number is well above this value. 4. Speech Recognition Experiments This section details the experimental protocol (section 4.1) as well as the speech recognition results (section 4.2) ob- tained with the Loquacious Set on variou... | https://arxiv.org/abs/2505.21578v1 |
Set, the WER drops from 24.3% to 7.5% when switching from the small data split with 25M parameters conformer to the full dataset and 480M parameters. This trend is also observed in all available test sets. Secondly, it is interesting to note that not all test sets benefit equally from adding more data. In the case of V... | https://arxiv.org/abs/2505.21578v1 |
B. Paul and J. Baker, “The design for the wall street journal- based csr corpus,” in Speech and Natural Language: Proceed- ings of a Workshop Held at Harriman, New York, February 23-26, 1992 , 1992. [2] J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett, “Darpa timit acoustic-phonetic continous ... | https://arxiv.org/abs/2505.21578v1 |
Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” arXiv preprint arXiv:1912.06670 , 2019. [14] M. Ravanelli, T. Parcollet, A. Moumen, S. de Langen, C. Subakan, P. Plantinga, Y . Wang, P. Mousavi, L. D. Libera, A. Ploujnikov, F. Paissan, D. Borra, S. Zaiem, Z. Zhao, ... | https://arxiv.org/abs/2505.21578v1 |
arXiv:2505.21598v1 [cs.CL] 27 May 2025Rethinking Data Mixture for Large Language Models: A Comprehensive Survey and New Perspectives Yajiao Liu1, Congliang Chen1, Junchi Yang1,2, Ruoyu Sun*1,2 1The Chinese University of Hong Kong, Shenzhen 2Shenzhen Research Institute of Big Data {yajiaoliu, congliangchen}@link.cuhk.ed... | https://arxiv.org/abs/2505.21598v1 |
during the target model’s training process. As the cost of training LLMs continues to rise with the increasing scale and diverse sources of data, data mixture has be- come increasingly important. Meanwhile, many new studies—beyond the scope of Albalak et al. (2024)—have gradually emerged. Accordingly, we believe it is ... | https://arxiv.org/abs/2505.21598v1 |
trated in Algorithm 2. During the target model training, the domain weights are dynamically opti- mized to effectively and efficiently steer the target model toward optimal performance. Most online methods can be regarded as online variants of exist- ing offline approaches. For example, certain offline methods can be t... | https://arxiv.org/abs/2505.21598v1 |
ˆℓval,j(θ) =ℓ(ˆDval,j;θ). We provide a detailed explanation of static weights anddynamic weights in Appendix A to clarify how different methods leverage them to ap- proximate the best-performing model θ⋆. 3 Offline Methods Offline methods assume the existence of optimal static domain weights w⋆for a given corpus com- 2... | https://arxiv.org/abs/2505.21598v1 |
(Chen et al., 2024) PT OthersODM (Albalak et al., 2023) PT ADO (Jiang et al., 2024) PT Table 1: Overview of existing data mixture methods discussed in this paper. In the last column, we report the specific training setting for each method, as described in the corresponding original paper. SFT denotes Supervised Fine-Tu... | https://arxiv.org/abs/2505.21598v1 |
=cj+bjexpPk i=1Aijwi c,b∈Rk,A∈Rk×kw⋆= min wPk j=1hjℓval,j(w) BiMix ℓval,j(wj, s) =Aj wαj j Bj sβj+Cj A, B, C, α, β ∈Rkw⋆= min w1 kPk j=1ℓval,j(wj, s)s.t.Pwj= 1 AutoScale ℓval,j(wj, D) = Dj 0+wjD−γj+cj D0, γ, c∈Rkw⋆= min w1 kPk j=1ℓval,j(wj, D)s.t.Pwj= 1 RegMix ℓval,j(w) =Mj(w) Regression model {Mj}k j=1w⋆= min wℓ... | https://arxiv.org/abs/2505.21598v1 |
performance for any given w; (3)optimizing the objective using appropriate al- gorithms to identify the optimal domain weights w⋆= minw∈∆kg(f1(w), ..., f k(w)). Function fitting-based methods reduce the bi- level optimization problem to a single-level one by fitting mixing laws, which can be solved us- ing gradient-bas... | https://arxiv.org/abs/2505.21598v1 |
face (1) model transfer issue : Albalak et al. (2023) demonstrates that the optimal domain weights w⋆may not trans- fer well across models when adopting new model architectures or different tokenizers. Algorithm- based methods also face (2) model scale issue :the optimal weights derived from a small proxy model fail to... | https://arxiv.org/abs/2505.21598v1 |
where ηis the learning rate. The dynamic domain weights wtare updated every tupdate steps based on Vt i, with ζ denoting the learning rate for updating wt. Velocitune (Luo et al., 2024) refers to Vt ias the learning velocity of the target model on domain i, which we interpret as a measure of learning progress. Method T... | https://arxiv.org/abs/2505.21598v1 |
uses the fitted laws to dynamically adjust do- main weights wt. We provide a comparison be- tween Skill-It and Aioli in Table 4. 4.3 Others ODM (Albalak et al., 2023) and ADO (Jiang et al., 2024) adopt alternative strategies for dynamically adjusting domain weights during the training of the target model, differing fro... | https://arxiv.org/abs/2505.21598v1 |
the last token of the training sample, and the loss trajectory of the training sample over multiple runs. They find that using the loss trajec- tory as the sample feature significantly outperforms the other metrics. 5.3 Loss metrics vs. Task performance To determine the optimal domain weights, loss met- rics, such as t... | https://arxiv.org/abs/2505.21598v1 |
data expert models. Preprint , arXiv:2502.15950. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international confer- ence on machine learning , pages 41–48. Mayee F. Chen, Michael Y . Hu, Nicholas Lourie, Kyunghyun Cho, and Christopher R... | https://arxiv.org/abs/2505.21598v1 |
Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Si- monyan, Erich Elsen, Jack W. Rae, Oriol Viny... | https://arxiv.org/abs/2505.21598v1 |
Shizhe Diao, Renjie Pi, Jipeng Zhang, Chi Han, and Tong Zhang. 2024a. Lisa: Lay- erwise importance sampling for memory-efficient large language model fine-tuning. Preprint , arXiv:2403.17919. Rui Pan, Jipeng Zhang, Xingyuan Pan, Renjie Pi, Xi- aoyu Wang, and Tong Zhang. 2024b. Scalebio: Scal- able bilevel optimization ... | https://arxiv.org/abs/2505.21598v1 |
Tim Cera, Tim Leslie, Tiziano Zito, Tom Krauss, Utkarsh Upad- hyay, Yaroslav O. Halchenko, and Yoshiki Vázquez- Baeza. 2020. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature Methods , 17(3):261–272. Jason Wei, Maarten Bosma, Vincent Y . Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, A... | https://arxiv.org/abs/2505.21598v1 |
i∈[k], where |Dtrain,i|denotes the number of to- kens in each training domain data Dtrain,i. The domain weights in these methods remain fixed rather than being optimized and thus they do not rely on proxy models. Considering that they can be implemented by omitting line 3 from Framework 1, and for the sake of completen... | https://arxiv.org/abs/2505.21598v1 |
style iterative update: θt+1=θt−ηdt, ηis the learning rate, dt=P iwt i∇ℓt train,i(θ). There are several approaches (Désidéri, 2012; Liu et al., 2021a,b) targeting to improve the "worst-case improvement" among all tasks, where ft i(θ) = ℓt train,i(θ)−ℓt train,i+1(θ). Yu et al. (2020) propose a form of gradient surgery t... | https://arxiv.org/abs/2505.21598v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.