text string | source string |
|---|---|
collaborative board game Forbidden Island MRDA Corpus (Shriberg et al., 2004)Dialog Act Recognition Audio, Text 75 meetings with an average of about six participants per meeting MIntRec2.0 (Zhang et al., 2024a)Dialog Act Recognition Audio, Video, Text1,245 conversations, 15,040 utterances from TV series: Su- perstore, ... | https://arxiv.org/abs/2505.18845v1 |
(Bhatnagar et al., 2022) Dialog Summary SAMSum (Gliwa et al., 2019) ROUGE-1 45 Bi-Level CL (Huang et al., 2022) Disentanglement Ubuntu IRC (Kummerfeld et al., 2019) F1 Score 70.4 CLUCDD (Gao et al., 2023) Disentanglement Ubuntu IRC (Kummerfeld et al., 2019) F1 Score 58.42 DiHRL (Li et al., 2024a) Disentanglement Ubuntu... | https://arxiv.org/abs/2505.18845v1 |
Journal of Data-centric Machine Learning Research (2025) Submitted 5/25; Revised 10/23; Published 11/23 Signal, Image, or Symbolic: Exploring the Best Input Representation for Electrocardiogram-Language Models Through a Unified Framework William Han1wjhan@andrew.cmu.edu Chaojing Duan2chaojing.duan@ahn.org Zhepeng Cen1z... | https://arxiv.org/abs/2505.18847v1 |
Electrocardiogram-Language Models (ELMs) (Zhao et al., 2024; Lan et al., 2025; Liu et al., 2024b; Han et al., 2024a). ELMs offer enhanced flexibility and interpretability, mirroring the process by which cardiac electrophysiologists investigate ECGs. Although ELMs are in their early stages and do not yet match the exper... | https://arxiv.org/abs/2505.18847v1 |
studies that highlight the performance of each input representation when varying the LLM, ECG length, and sequence length. 3.We open source a comprehensive, unified framework for benchmarking ELMs across 4 input representations, 6 datasets, 4 ECG-specific encoders, 3 general, pretrained encoders, and 3 LLMs. We hope th... | https://arxiv.org/abs/2505.18847v1 |
image representations of ECG signals (Liu et al., 2024b; Lan et al., 2025), denoted as Ximg. Although Ximgis the format used by cardiac electrophysiol- ogists to analyze ECG recordings, it poses several challenges for deep learning systems (Ni et al., 2025). While this approach has gained popularity in the general time... | https://arxiv.org/abs/2505.18847v1 |
Liu et al., 2024b). In the LLaVA-style approach (Liu et al., 2023, 2024b), fECGis frozen, and a learnable projection matrix W∈Rh×dis introduced, where his the hidden dimension of the LLM. The latent vector z=fECG(X) is projected to z′=Wz, concatenated with the embedded query Q, and fed into the LLM to generate the resp... | https://arxiv.org/abs/2505.18847v1 |
way for comparability, largely following (Han et al., 2024a). We first adjust all lead arrangements to be consistent with PTB-XL (i.e., [I, II, III, aVL, aVR, aVF, V1–V6]). Powerline noise is removed using bidirectional notch filters at 50 Hz and 60 Hz with a quality factor (Q) of 30. A fourth-order Butterworth bandpas... | https://arxiv.org/abs/2505.18847v1 |
al., 2024a), ST-MEM (Na et al., 2024), MLAE (Zhang et al., 2023), and MTAE (Zhang et al., 2023). For general, pretrained encoders, we utilize CLIP (Radford et al., 2021), ViT (Dosovitskiy et al., 2021), and SigLIP (Zhai et al., 2023). 3.5 Large Language Models In this study, we utilize the Llama 3.2 1B Instruct (Gratta... | https://arxiv.org/abs/2505.18847v1 |
n, sn. A representative example of the token template is given in Appendix B. To ensure that training focuses only on the desired targets—specifically, the assistant responses and their corresponding end-of-turn tokens—we define a target labeling function ℓ:{1,2, . . . , T } → V ∪ {− 100}, where Vis the extended vocabu... | https://arxiv.org/abs/2505.18847v1 |
We begin with the main results on four datasets, highlighting statistically significant differences, then report performance on the PULSE ECG-Bench benchmark. We also assess the robustness of each representation under signal perturbations. We then go into a series of ablation studies to highlight performance difference... | https://arxiv.org/abs/2505.18847v1 |
= Perturbed – Baseline) averaged over 5 random seeds on the ECG-Chat Instruct dataset. Representation Training Paradigm F(∗)∆ (Perturbed – Baseline) BLEU-4 ROUGE-L METEOR BERTScore F1 Accuracy Xsig 2-Stage ScratchMERL -0.07 -0.05 -0.08 -0.08 +0.05 ST-MEM +0.04 -0.03 -0.01 -0.01 +0.01 MLAE +0.01 -0.07 -0.05 -0.09 -0.09 ... | https://arxiv.org/abs/2505.18847v1 |
±0.06 Qwen 2.5 1.5B InstructXsig 2-Stage ScratchMERL 16.12 ±0.04 60.33 ±0.05 53.87 ±0.07 95.21 ±0.01 6.89 ±0.06 ST-MEM 15.42 ±0.02 58.63 ±0.06 50.74 ±0.05 95.74 ±0.02 6.52 ±0.03 MLAE 15.34 ±0.03 58.51 ±0.06 50.86 ±0.05 95.23 ±0.01 6.53 ±0.05 MTAE 15.49 ±0.08 59.48 ±0.03 51.36 ±0.04 95.28 ±0.01 6.72 ±0.07 X∗ sig End-to-... | https://arxiv.org/abs/2505.18847v1 |
Subsection 3.6). Although lengths of 1024–2048 tokens are modest by modern standards, practitioners with tight computational budgets may prefer the End-to-End LLaVA models—they avoid training an ECG encoder from scratch and remain robust at shorter sequence lengths. Table 5: Mean results over 5 random seeds with standa... | https://arxiv.org/abs/2505.18847v1 |
lengths, re-using the L= 1250 setting from Figure 2 and Table 8. As shown in Table 6, ECG-Byte consistently outperforms all baselines, with its advantage growing as Lincreases. This is intuitive since 2-Stage Scratch andEnd-to-End LLaVA methods compress the ECG into one token, while ECG-Byte compresses the signal while... | https://arxiv.org/abs/2505.18847v1 |
approach consistently achieves the highest mean scores and secures the greatest number of significant wins in two-sample z-tests (Tables 1). This performance gap widens with longer ECG segments ( L= 2500) and remains robust under realistic perturbations (Table 3), suggesting that the model effectively leverages compres... | https://arxiv.org/abs/2505.18847v1 |
series approximation, 2024b. URL https://arxiv.org/abs/2411.15209 . Seokmin Choi, Sajad Mousavi, Phillip Si, Haben G. Yhdego, Fatemeh Khadem, and Fate- meh Afghah. Ecgbert: Understanding hidden language of ecgs with self-supervised rep- resentation learning, 2023. Tri Dao. Flashattention-2: Faster attention with better... | https://arxiv.org/abs/2505.18847v1 |
Hosseini, Sahana Chennabasappa, Sanjay Singh, Sean Bell, Seohyun Sonia Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Raparthy, Sheng Shen, Shengye Wan, Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whitman, Sten Sootla, Stephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herm... | https://arxiv.org/abs/2505.18847v1 |
Mihir Patel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang, Miquel Jubert Hermoso, Mo Metanat, Mohammad Rastegari, Mun- ish Bansal, Nandhini Santhanam, Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, Nikhil Mehta, Nikolay Pavlovich Laptev, Ning Dong, Norman ... | https://arxiv.org/abs/2505.18847v1 |
record dataset. Scientific Data, 10, 01 2023. doi: 10.1038/s41597-022-01899-x. Mark Johnson. Counties most in need of cardiologists are the most likely to have none, 07 2024. URL https://www.washingtonpost.com/science/2024/07/29/ cardiologists-rural-counties-shortage/ . Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B... | https://arxiv.org/abs/2505.18847v1 |
cardiovascular disease through language models? In Andreas Vlachos and Isabelle Augenstein, editors, Findings of the Association for Com- putational Linguistics: EACL 2023 , pages 442–453, Dubrovnik, Croatia, May 2023a. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-eacl.33. URL https://aclan... | https://arxiv.org/abs/2505.18847v1 |
Cecilia Mascolo, and Aaqib Saeed. Electrocardiogram- language model for few-shot question answering with meta learning, 2024a. URL https: //arxiv.org/abs/2410.14464 . Jialu Tang, Tong Xia, Yuan Lu, Cecilia Mascolo, and Aaqib Saeed. Electrocardiogram re- port generation and question answering via retrieval-augmented sel... | https://arxiv.org/abs/2505.18847v1 |
Greenspan, Benjamin Glicksberg, Jagat Narula, and Girish Nadkarni. Heartbeit: Vision transformer for electrocardiogram data improves diagnostic performance at low sample sizes, 2022. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention i... | https://arxiv.org/abs/2505.18847v1 |
it as a general assistant. Appendix B. Input Templates We provide the input template for each LLM used in our study. We utilize the conversational templates provided by FastChat (Zheng et al., 2023). Llama 3.2 1B Instruct Conversation Template <|begin oftext|><|start header id|>system <|endheader id|> qsys<|eotid|><|st... | https://arxiv.org/abs/2505.18847v1 |
BertScore F1 Accuracy ECG-Chat PretrainXsig 2-Stage ScratchMERL 4.77 ±0.06 44.33±0.11 26.86 ±0.16 93.06 ±0.02 11.12 ±0.10 ST-MEM 6.93 ±0.04 32.83 ±0.05 23.83 ±0.08 90.82 ±0.01 0.58 ±0.09 MLAE 7.64 ±0.10 36.28 ±0.07 24.35 ±0.08 91.65 ±0.01 3.67 ±0.11 MTAE 5.71 ±0.08 31.36 ±0.14 20.09 ±0.18 90.90 ±0.02 1.61 ±0.12 X∗ sig ... | https://arxiv.org/abs/2505.18847v1 |
arXiv:2505.18853v1 [cs.CL] 24 May 2025Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation Alexander Shabalin HSE University Constructor University amshabalin@hse.ruViacheslav Meshchaninov HSE University Constructor University vmeshchaninov@hse.ruDmitry Vetrov Constructor University dvetrov@constructor... | https://arxiv.org/abs/2505.18853v1 |
Accounting for Discreteness ✓ ✗ ✓ ✓ Accounting for Semantics ✗ ✓ ✗ ✓ semantic relationships between tokens during the noising process, resulting in a more erratic and less meaningful degradation of information. In this paper, we propose SMOOTHIE , a smoothing diffusion framework that satisfies both properties. We repre... | https://arxiv.org/abs/2505.18853v1 |
the vocabulary and sw (i)=+k, i =w −k,otherwise(1) Then the latent is represented as a sequence S0= (swy 1, . . . ,swy m). Corruption is performed with the Gaussian diffusion process with noise variance multiplied by k2(k= 5 by default), St=√¯αtS0+k√1−¯αtε. The model input is calculated by first producing a probabilit... | https://arxiv.org/abs/2505.18853v1 |
1 σ2 tD0, δ2I (3) The noise scheduler σt(1< σ 1<···< σT) controls the amount of noise added at each timestep. The hyperparameter δcontrols the stochasticity of the diffusion process and makes it non-deterministic. Following [40], we keep δindependent of the timestep t. To construct the model input, we convert Dtinto a... | https://arxiv.org/abs/2505.18853v1 |
in the following loss function: LD(θ) =Ewy,t,pt ∥D0(Ewy)−gθ(pt, t)∥2 (10) However, this loss function is challenging to optimize due to the high variance and dimensionality of D0. To address this issue, we introduce the following theorem: Theorem 4.1. Letg∗(pt, t)be an optimal prediction for Eq. 10. Then g∗(pt, t) =D... | https://arxiv.org/abs/2505.18853v1 |
proposed in [ 42], consisting of Transformer decoder layers [ 44] augmented with UNet-style skip connections. Specifically, the output of the first layer is added to the input of the last, the second to the second-last, and so on. The full model has 12 layers and approximately 100M parameters. For conditional generatio... | https://arxiv.org/abs/2505.18853v1 |
204K BBC articles and their corresponding summaries. More detailed information about each dataset is provided in the Appendix G. Metrics Following the evaluation protocol from prior work [ 13,25], we employ a combina- tion of n-gram-based, diversity and semantic similarity metrics. Specifically, we report BLEU [35] and... | https://arxiv.org/abs/2505.18853v1 |
DiffuSeq and FLAN-T5 results from [ 32], and AR-Diffusion, GENIE, and Transformer results from [ 46]. We re-implement and train the embedding- and simplex-based diffusion baselines within our framework for a fair comparison. For clarity we mark autoregressive methods with †, previous diffusion approaches with ⋄and our ... | https://arxiv.org/abs/2505.18853v1 |
3.33 2 Mean rank To assess generalization across tasks, Table 5 reports the mean rank of diffusion models across all datasets except XSum, which is excluded due to limited overlap in evaluated methods. Rankings are computed based on the average of BERTScore and BLEU metrics. SMOOTHIE achieves the best mean rank, while ... | https://arxiv.org/abs/2505.18853v1 |
Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. [4]Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel. Extracting training data from large langua... | https://arxiv.org/abs/2505.18853v1 |
Word embeddings as metric recovery in semantic spaces. Transactions of the Association for Computational Linguistics , 4:273–286, 2016. doi: 10.1162/tacl_a_00098. URL https://aclanthology. org/Q16-1020/ . [18] Zhengfu He, Tianxiang Sun, Qiong Tang, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. DiffusionBERT: Improving... | https://arxiv.org/abs/2505.18853v1 |
for automatic evaluation of summaries. In Text summarization branches out , pages 74–81, 2004. [29] Zhenghao Lin, Yeyun Gong, Yelong Shen, Tong Wu, Zhihao Fan, Chen Lin, Nan Duan, and Weizhu Chen. Text generation with diffusion language models: a pre-training approach with continuous paragraph denoise. In Proceedings o... | https://arxiv.org/abs/2505.18853v1 |
Eleventh International Conference on Learning Representations , 2023. URL https://openreview.net/forum?id=4PJUBT9f2Ol . [41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Compu... | https://arxiv.org/abs/2505.18853v1 |
more advanced approaches need to be developed. B Societal Impact Language models have been shown to produce harmful outputs [ 45], spread disinformation [ 43], hallucinate [ 23], and potentially violate user privacy [ 4]. Although our study focuses on tasks that differ from those typically used in prior harmfulness eva... | https://arxiv.org/abs/2505.18853v1 |
7. We trained our models using two 80 GB NVIDIA A100 GPUs for 15 hours on average. For all the tasks, we save checkpoints every 25,000 steps. We select the best checkpoint by the quality on the development set. During generation we do not apply the clamping trick [ 27], since it does not improve quality in our experime... | https://arxiv.org/abs/2505.18853v1 |
large-scale dataset for the question generation task. It requires models to comprehend natural language queries and extract answers from a large corpus. The dataset consists of open-domain trivia questions and their corresponding answers, collected from various internet sources. We use the version preprocessed by [ 13]... | https://arxiv.org/abs/2505.18853v1 |
arXiv:2505.18855v1 [cs.CV] 24 May 2025Inference Compute-Optimal Video Vision Language Models Peiqi Wang1, ShengYun Peng2, Xuewen Zhang3, Hanchao Yu3, Yibo Yang3, Lifu Huang4, Fujun Liu3, Qifan Wang3 1MIT,2Georgia Tech,3Meta,4UC Davis Email: wpq@mit.edu Code: github/vvlm_inference_scaling Abstract This work investigates... | https://arxiv.org/abs/2505.18855v1 |
al., 2024b; Wu et al., 2024b; Snell et al., 2024; Brown et al., 2024) don’t investigate the pos- sible interaction between scaling factors xand fine- tuning data size n. Scaling factors influence how effectively finetuning data is used: (1) larger mod- els are more sample-efficient, achieving lower er- ror with the sam... | https://arxiv.org/abs/2505.18855v1 |
of data size on the compute-optimal frontier. 2 Relationship to Scaling Law Studies In contrast to previous studies on training compute- optimal model that balances model size xNand pretraining data size npt(Kaplan et al., 2020; Hoff- mann et al., 2022), we aim to optimize scaling factors for video VLMs such that the m... | https://arxiv.org/abs/2505.18855v1 |
model components. We assume each ex- ample consists of a single video, where the length of visual tokens dominate that of the input instruc- tions or output generations. Thus, we disregard the compute cost of the latter. Using the standard approximation of 2xNFLOPs per token for a trans- former model with xNparameters ... | https://arxiv.org/abs/2505.18855v1 |
referred to as the “(inference) compute-optimal frontier”. Unlike training compute optimization, inference compute optimization does not treat nas an op- timization variable, as finetuning data size does not affect inference compute cost. However, for certain parametric functions (e.g., add-interact ), ncan still influ... | https://arxiv.org/abs/2505.18855v1 |
for video chat capabilities, LongVideoBench (LVB Wu et al., 2024a) for long video understanding, and PerceptionTest (PT Patraucean et al., 2023) for fine- grained perception. Additionally, MVBench (MV Li et al., 2024c), Video-MME (VMME Fu et al., 2024), and Next-QA (NQA Xiao et al., 2021) pro- vide broad evaluations ac... | https://arxiv.org/abs/2505.18855v1 |
kαkx−ak k +ξn−d+ϵ 0.56 1.11 0.94 2.04 2.15 0.83 add-interact sP kαkx−ak k +P kβkxbk kn−d+ϵ 0.24 0.8 0.97 0.94 1.32 0.92 add-interactP kαkx−ak k +P kβkxbk kn−d+ξn−d+ϵ0.2 0.77 0.98 0.95 1.33 0.92 Table 2: Comparison of Parametric Models of Task Performance. Evaluation of different parametric functions for modeling the av... | https://arxiv.org/abs/2505.18855v1 |
8. For exam- ple, in PerceptionTest, the frontier prioritizes xV (with higher marginal benefits) over xT(with lower marginal benefits). This highlights that optimal in- ference compute allocation strategies should adapt to the task of interest, focusing on factors that yield the highest marginal returns on performance.... | https://arxiv.org/abs/2505.18855v1 |
subfigure in Figure 3 reveals consider- able variability in the model’s extrapolation per- formance across tasks. While average task perfor- mance is easier to model, add-interact struggles with tasks like LongVideoBench (LVB) and Next- QA (NQA). For these tasks, E%≥5%corresponds to an average deviation exceeding 3poin... | https://arxiv.org/abs/2505.18855v1 |
shown in Figure 5, though with task-specific vari- ations. The average trend suggests decreasing xN and increasing xT, xVas data size ngrows. Figure 5: Elasticity Across Tasks. Bar plot showing the elasticity (defined in Equation 11) for scaling fac- torsk∈ {N, T, V }across video tasks. This measures the sensitivity of... | https://arxiv.org/abs/2505.18855v1 |
data size af- fects the compute-optimal frontier. Figure 4 and 5 suggest that as more data becomes available, it is optimal to allocate less compute to LM size xN and more to video visual representations xTand xV. We hypothesize this is because detailed vi- sual representations are more complex to learn andrequire more... | https://arxiv.org/abs/2505.18855v1 |
includes runs with lower inference compute than the star sweep), it remains uncertain how well the predicted compute-optimal frontier generalizesto significantly higher inference compute, such as 10x the current FLOPs. Addressing these issues is challenging due to the limited computational re- sources allocated to this... | https://arxiv.org/abs/2505.18855v1 |
tasks. Preprint , arXiv:2403.08540. Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. 2017. The “Something Something” Video Databa... | https://arxiv.org/abs/2505.18855v1 |
Detailed Video Understanding via Large Vision and Language Models. In Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. 2024b. VideoGPT+: Integrating Image and Video Encoders for Enhanced Video... | https://arxiv.org/abs/2505.18855v1 |
Compute-Optimal Infer- ence for LLM Problem-Solving. In The Thirteenth International Conference on Learning Representa- tions . Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. 2021. NExT-QA: Next Phase of Question- Answering to Explaining Temporal Actions. In 2021 IEEE/CVF Conference on Computer Vision and Pat... | https://arxiv.org/abs/2505.18855v1 |
course of a month (30 days), the total number of videos processed for inference becomes ninf= 34×30 = 1020 million videos. Substituting these values into the cost ratio gives: Cost ratio =ninf 3n=1020 3×1= 340 . Thus, in this scenario, the inference cost is approximately 340times higher than the finetuning cost over a ... | https://arxiv.org/abs/2505.18855v1 |
et al., 2024). Here, we emphasize the importance of accounting for the vision model’s compute cost in video VLMs. Using the standard approximation of 2xNFLOPs per token for a transformer model with xNparam- eters (Kaplan et al., 2020), the per-example inference compute costs for both the vision and language components ... | https://arxiv.org/abs/2505.18855v1 |
additional details. To assess the dataset’s effectiveness, we conduct a coarse ablation study by removing subsets such as LLaV A-Hound, the VideoChatGPT suite, and the QA datasets. Although some tasks show improved performance in these ablated setups, the complete dataset demonstrates the best average performance acros... | https://arxiv.org/abs/2505.18855v1 |
In this paper, we evaluate the video VLM on a diverse set of 8 downstream video tasks to ensure comprehensive assessment across various tasks and domains. For reproducibility, we use LMMs-Eval (Zhang et al., 2024a). The evaluation includes: (1) VideoDetailedCaption (VDC Zhang et al., 2024a) for detailed video descripti... | https://arxiv.org/abs/2505.18855v1 |
from sweeps is used for visualization and parametric fitting of scaling curves. In this paper, we conduct two types of sweeps on scaling factors xand finetuning data size n: (1) the star sweep, proposed in SoViT (Alabdulmohsin et al., 2023), and (2) the isoFLOP sweep, used by Chinchilla (Hoffmann et al., 2022). We fine... | https://arxiv.org/abs/2505.18855v1 |
x∗(c)ind.n Multiple Factors ( K > 1) Single Factor ( K= 1) mult α(Q kx−ak k)n−d+ε αx−an−d+ε ✓ addP kαkx−ak k +ξn−d+ε αx−a+ξn−d+ε ✓ add-interact sP kαkx−ak k +P kβkxbk kn−d+ε αx−a+βxbn−d+ε × add-interactP kαkx−ak k +P kβkxbk kn−d+ξn−d+ε αx−a+βxbn−d+ξn−d+ε × Table 7: Parametric Models of Task Performance. This table summ... | https://arxiv.org/abs/2505.18855v1 |
xRequires More Data to Match the Same Error: When b >0, the coefficient (βxb+ξ) increases with x. For two values xS< x L, we have (βxb S+ξ)<(βxb L+ξ). To achieve the same error f(x, n), a larger xLrequires a larger finetuning data size nLcompared to nS. Specifically, the equality (βxb S+ξ)n−d S= (βxb L+ξ)n−d Lholds onl... | https://arxiv.org/abs/2505.18855v1 |
squared error (MSE) and Huber loss. MSE consistently outper- forms Huber loss and is therefore used as the objective. •Parameter Initialization : The optimization problem is nonconvex and nonlinear, so we mitigate the risk of poor local minima by running 500 trials with random initializations and selecting the best fit... | https://arxiv.org/abs/2505.18855v1 |
0.7 10 -0.22 0.17 0.79△n e N eT eV 0.1 0.32 0.2 1.13 0.5 -0.08 0.2 0.99 1.0 -0.15 0.2 0.96 2.0 -0.2 0.19 0.89 3.0 -0.23 0.18 0.83 4.0 -0.23 0.17 0.8 5.0 -0.22 0.17 0.79 6.0 -0.22 0.17 0.78 7.0 -0.22 0.16 0.75 8.0 -0.21 0.16 0.72 9.0 -0.2 0.15 0.68 10.0 -0.19 0.15 0.64 Table 10: Ablation Study on Elasticity Computation.... | https://arxiv.org/abs/2505.18855v1 |
and data sizes as ek=1 |C||N|X c∈CX n∈Nek(c, n), (11) withNdenoting the set of finetuning data sizes of interest. These aggregated metrics provide a concise summary of how data size influences optimal scaling factors. Table 10 presents an ablation study analyzing key design choices for computing elasticity. The finding... | https://arxiv.org/abs/2505.18855v1 |
arXiv:2505.18859v1 [cs.CL] 24 May 2025Writing Like the Best : Exemplar-Based Expository Text Generation Yuxiang Liu Kevin Chen-Chuan Chang University of Illinois at Urbana-Champaign, USA {yuxiang, kcchang}@illinois.edu Abstract We introduce the Exemplar-Based Expository Text Generation task, aiming to generate an ex- p... | https://arxiv.org/abs/2505.18859v1 |
ensuring cross-topic con- sistency , while maintaining accuracy and relevance to the new topic by addressing cross-topic variabil- ity. However, existing methods that rely on exten- sive exemplar data are impractical (Balepur et al., 2023; Shao et al., 2024b), making it difficult to achieve cross-topic consistency due ... | https://arxiv.org/abs/2505.18859v1 |
of imitativeness and adaptiveness. Additionally, our analysis reveals that each module within RePA contributes effectively to its overall performance. Our contributions can be summarized as follows: (1) We are the first to study Exemplar-Based Expos- itory Text Generation task, addressing a practical yet under-explored... | https://arxiv.org/abs/2505.18859v1 |
scores indicating the correctness of generated text–is crucial for the trustworthiness of real-world systems. Previous methods requiredwhite-box access to model architectures or fine- tuning (Kadavath et al., 2022; Jiang et al., 2021; Lin et al., 2022; Yang et al., 2023b; Zhang et al., 2023; Slobodkin et al., 2023), wh... | https://arxiv.org/abs/2505.18859v1 |
for returnY As showed in Figure 2, REPAmirrors the recur- rence structure in LSTMs, but a closer comparison highlights several distinct features: 1) it employs a text-based representation for input xt, output yt, short-term memory ht, and long-term memory ct; 2) it leverages prompting of general-purpose LLMs Msuch as G... | https://arxiv.org/abs/2505.18859v1 |
source topic to target topic with simply topic token substitution. Thus, we have: qt=LLM Outline (x′ t, tx, ty) 3.2.2 The A DAPT Module Previous plan-then-generate frameworks often as- sume plans will seamlessly translate into effective outputs. However, in our task, plans for the target topic are derived from those fo... | https://arxiv.org/abs/2505.18859v1 |
belong to the same domain. Such "correspondence" demands a high level of similarity, though "variations" usually exist. To this end, we collect three diverse datasets5to cover both open-domain and domain-specific scenarios, including Wikipedia, RoleEE, and USNews. Wikipedia The Wikipedia dataset is domain- agnostic. We... | https://arxiv.org/abs/2505.18859v1 |
comparable inference steps with R EPA. Default is to simply replace source topic in the source text with target topic, serving as an approxi- mation for understanding cross-topic variability. We also developed retrieval-augmented genera- tion (RAG) variants of the aforementioned LLM- based baselines to integrate retrie... | https://arxiv.org/abs/2505.18859v1 |
) and contradiction ( NLI-C ) sentences in outputs. Our human evaluation of the effectiveness of NLI-based metrics are included in Appendix E. We also compute the percentage of hallucinated tokens compared to both inputs and ground truths, named Halluc . 5 Experimental Results 5.1 Comparison with Baselines We present t... | https://arxiv.org/abs/2505.18859v1 |
0.6137 0.7019 0.6973 0.8126 0.5495 19.5238 0.3527 0.4562 SR+Retr 0.8458 0.7239 0.8182 0.8139 0.8639 0.7236 11.5269 0.5338 0.1757 Default 0.7245 0.5808 0.6999 0.7013 0.7768 0.5494 18.1090 0.1667 0.5700 Table 1: Evaluation results on basic and factuality metrics. LLM denotes GPT-4 specifically. Datasets Models I.↑ A.↑ A.... | https://arxiv.org/abs/2505.18859v1 |
also re- moved, and retrieval is based solely on the clarified input segment without any generated questions. 3) w/o Refusal removes confidence calibration in QA, resulting in no refusal in answering questions. 4) w/o Revise-LTM removes the Revise compo- nent in Write, as well as the long-term memory for storing histor... | https://arxiv.org/abs/2505.18859v1 |
fundamental premises of our approach, and our study is scoped around these assumptions. How- ever, scenarios involving low-quality or dissimilar exemplars pose challenges to the model’s perfor- mance, and future work might explore these scenar- ios. Additionally, although our recurrent prompting pipeline enhances perfo... | https://arxiv.org/abs/2505.18859v1 |
Halcomb. 2018. Students use of exem- plars to support academic writing in higher education: An integrative review. Nurse education today , 65:87– 93. Jun Chen. 2024. Exploring imitative learning in a blended efl writing class. Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language gener... | https://arxiv.org/abs/2505.18859v1 |
know? on the calibration of language models for question answering. Transactions of the Associa- tion for Computational Linguistics , 9:962–977. Yizhu Jiao, Sha Li, Yiqing Xie, Ming Zhong, Heng Ji, and Jiawei Han. 2022. Open-vocabulary argumentrole prediction for event extraction. In Findings of the Association for Com... | https://arxiv.org/abs/2505.18859v1 |
1687–1698, Dublin, Ireland. Association for Compu- tational Linguistics. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettle- moyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings o... | https://arxiv.org/abs/2505.18859v1 |
in the hidden states of over-confident large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 3607–3625, Singapore. Association for Com- putational Linguistics. Xiaofei Sun, Zijun Sun, Yuxian Meng, Jiwei Li, and Chun Fan. 2022. Summarize, outline, and el... | https://arxiv.org/abs/2505.18859v1 |
Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2023. R-tuning: Teaching large language models to refuse unknown questions. arXiv preprint arXiv:2311.09677 . Ruqing Zhang, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi Cheng. 2019. Outline generation: Under- standing the inherent content structure of document... | https://arxiv.org/abs/2505.18859v1 |
of an F1 score. D.2 Discussion on Known Limitations There are known limitations of LLM-as-a-judge such as verbosity and self-enhancement biases, and we’d like to clarify in this section. Verbosity bias means LLM judges favor longer, verbose responses compared to shorter alternatives. However, our proposed REPA, despite... | https://arxiv.org/abs/2505.18859v1 |
heartfelt lyrics and catchy melodies. She rose to fame following the release of her self-titled debut album in October 2006. Starting her career as a solo artist, Taylor Alison Swift has achieved global superstardom, including winning the Grammy Award for album of the year for Midnights (2022), suggesting she is among ... | https://arxiv.org/abs/2505.18859v1 |
Midnights in 2024, showcase her strong and dedicated fan base that likely contributes to her high sales figures. Taylor Swift’s success has made her a cultural icon, captivating audiences with heartfelt lyrics and catchy melodies. She has become one of the most influential artists in contemporary music. She was the hig... | https://arxiv.org/abs/2505.18859v1 |
her songs in English. Beyoncé’s early career focused on R&B and pop music, highlighting her storytelling skills and personal experiences in her songs. She transitioned to pop music, broad- ening her appeal to a global audience. She is tied with American lyricist Diane Warren at third with nine songwriting credits on nu... | https://arxiv.org/abs/2505.18859v1 |
helps mitigate potential errors from earlier stages. Furthermore, as demonstrated in our ablation study (Table 3, 10), removing any step in the pipeline degrades performance. This finding underscores that every stage contributes positively to the over- all efficacy of REPA. Therefore, while the risk of error propagatio... | https://arxiv.org/abs/2505.18859v1 |
0.3344 13.6244 0.7509 0.0963 LLM 0.7329 0.5870 0.6916 0.6899 0.7098 0.5729 15.3616 0.3950 0.4766 LLM+Retr 0.7073 0.6079 0.6849 0.6809 0.6560 0.5606 13.0367 0.5580 0.3434 RoM 0.5711 0.4280 0.5316 0.5309 0.4617 0.3540 12.7861 0.3190 0.5181 RoM+Retr 0.6286 0.5119 0.5915 0.5830 0.5624 0.4634 15.6738 0.5835 0.2497 RoleEELLa... | https://arxiv.org/abs/2505.18859v1 |
and A.-I. denotes Adaptive- Imitativeness . R1↑ R2↑ RL↑ RLsum ↑Meteor ↑BLEU↑Halluc ↓ ↑ NLI-E ↓NLI-C I.↑ A.↑ A.-I.↑ Full 0.8164 0.6937 0.7727 0.7736 0.7888 0.6598 7.5361 0.7992 0.0741 4.14 3.70 3.7996 - C 0.7945 0.6806 0.7409 0.7411 0.6997 0.6057 8.6106 0.7416 0.0974 4.18 3.56 3.7403 - O 0.7489 0.6375 0.6932 0.6969 0.71... | https://arxiv.org/abs/2505.18859v1 |
, romanized: Davlekanovskiy rayon; Bashkir and Tatar: Дәүләкән районы, D¨ awl¨ ak¨ an rayonı ; Chuvash: Давлекан район˘ e, Davlekan rayon˘ e ) is an administrative and municipal district (raion), one of the fifty-four in the Republic of Bashkortostan, Russia. It is situated in the southeast of the republic and shares b... | https://arxiv.org/abs/2505.18859v1 |
of a Wikipedia example comparing proposed model with GPT-4 related baselines. Dataset I.- w/ tie I.- w/o tie A.- w/ tie A.- w/o tie Wikipedia 57.3% 78.0% 62.1% 82.7% RoleEE 58.7% 78.3% 59.7% 81.3% USNews 58.6% 80.7% 62.7% 84.8% Mean 58.2% 79.0% 61.5% 82.9% Table 12: Agreement between LLM-judge and human- judge on Imita... | https://arxiv.org/abs/2505.18859v1 |
arXiv:2505.18864v1 [cs.CL] 24 May 2025Audio Jailbreak Attacks: Exposing Vulnerabilities in SpeechGPT in a White-Box Framework Binhao Ma Department of Computer Science University of Missouri-Kansas City Kansas City, United States binhaoma@umkc.edu Zhengping Jay Luo Department of Computer Science and Physics Rider Univer... | https://arxiv.org/abs/2505.18864v1 |
is finalizing plans to incorporate LLM-powered voice capabilities into future versions of iOS [8]. As these models become embedded into mainstream platforms and devices, they are poised to redefine user expectations around voice interfaces, ushering in a new era of AI-powered, conversational agents that feel more intui... | https://arxiv.org/abs/2505.18864v1 |
for com- plex visual-linguistic tasks, while LLaV A [1] integrates visual encoders with language models to enable rich multimodal dialogue.Building upon these developments, SpeechGPT [3] and NExT-GPT [4] further extend the frontiers of multimodal interaction. SpeechGPT unifies speech and text into discrete token repres... | https://arxiv.org/abs/2505.18864v1 |
to generate a sequence of adver- sarial audio tokens that, when synthesized into audio and fed into the model, leads to harmful or policy-violating outputs that would typically be suppressed by alignment mechanisms. III. M ETHODOLOGY We propose a speech token-level adversarial attack pipeline for speech language models... | https://arxiv.org/abs/2505.18864v1 |
minimize L; ifˆy=ythen break C. Audio Reconstruction After performing greedy Adversarial Token Search and defining the target speech cluster token sequence, we first convert it into a waveform using a vocoder [33]. A global noise perturbation is then applied to the synthesized audio. The perturbed audio is passed throu... | https://arxiv.org/abs/2505.18864v1 |
purely random noise method directly optimizes entire speech token sequences as adversarial inputs. These sequences are then converted into audio waveforms using only random noise, without incorpo- rating or relying on any harmful speech content. B. Attack Performance We evaluate five audio attack methods: Random Noise,... | https://arxiv.org/abs/2505.18864v1 |
a range of question types. As illustrated in Q1Q2Q3Q4Q5Q6Q7Q8Q9Q10 Illegal Activity1.01.52.02.5NISQA score Q1Q2Q3Q4Q5Q6Q7Q8Q9Q10 Hate Speech1.52.02.5NISQA score Q1Q2Q3Q4Q5Q6Q7Q8Q9Q10 Physical Harm1.52.02.5NISQA score Q1Q2Q3Q4Q5Q6Q7Q8Q9Q10 Fraud1.52.02.5NISQA score Q1Q2Q3Q4Q5Q6Q7Q8Q9Q10 Pornography1.52.02.5NISQA score Q... | https://arxiv.org/abs/2505.18864v1 |
characteristics. Number of Iterations Required for Adversarial Token Optimization Across Different Jailbreak Scenarios. We set the number of adversarial tokens to 200 and tested the number of optimization iterations required to attack various types of jailbreak problems. The results are shown in Table IV. We observe th... | https://arxiv.org/abs/2505.18864v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.