text string | source string |
|---|---|
the output probability. The attribution score is proportional to the magnitude of this change. We experimented with different perturbation strategies (zero embedding, [MASK] token, random token) and found padding tokens to yield the most stable results. •DeepLIFT Shrikumar et al. [2017]: A backpropagation-based attribu... | https://arxiv.org/abs/2505.18970v1 |
Measures whether removing just the single most important sentence changes the model’s decision: DFS(x, e) =⊮g(˜x(1) e)̸=g(x)(14) where ˜x(1) erepresents the input with the most important sentence removed, and ⊮is the indicator function that equals 1 when the condition is true and 0 otherwise. Across a dataset, its aver... | https://arxiv.org/abs/2505.18970v1 |
presents a detailed comparison of ProtoSurE and four token-level attribution baselines—SHAP, Integrated Gradients (IG), Occlusion, and DeepLIFT—across four datasets and four target LLMs. We evaluate using multiple faithfulness metrics: Accuracy (Acc), Comprehensiveness (Comp), Sufficiency (Suff), Decision Flip Fraction... | https://arxiv.org/abs/2505.18970v1 |
0.109 0.212 0.104 0.172 0.146 0.283 0.317 0.365 0.343 0.389 0.351 0.235 1.00 Suff↓SHAP 0.205 0.192 0.234 0.257 0.063 0.050 0.173 0.027 0.196 0.185 0.171 0.190 0.159 0.137 0.184 0.215 0.165 4.03 IG 0.236 0.114 0.218 0.244 0.104 0.094 0.160 0.022 0.138 0.126 0.166 0.186 0.098 0.086 0.164 0.201 0.141 2.09 Occl 0.208 0.119... | https://arxiv.org/abs/2505.18970v1 |
can substantially improve ProtoSurE’s alignment with the underlying LLM’s decision boundary. F Case Study We provide qualitative examples to illustrate how PROTO SUREdecomposes predictions into interpretable prototype- level contributions. Figures 5 and 6 present two hotel reviews—one positive and one negative—and show... | https://arxiv.org/abs/2505.18970v1 |
arXiv:2505.18971v1 [cs.CL] 25 May 2025Is Architectural Complexity Overrated? Competitive and Interpretable Knowledge Graph Completion with RELAT E Abhijit Chakraborty∗1Chahana Dahal†2Ashutosh Balasubramaniam†3 Tejas Anvekar1 Vivek Gupta∗1 1Arizona State University2Westminster University3IIT Guwahati {achakr40,vgupt140}... | https://arxiv.org/abs/2505.18971v1 |
that architectural minimalism when paired with principled training can achieve both efficiency and robustness. In this paper: •We introduce RELAT E, a fully real-valued, interpretable knowledge graph embedding model that uses dual-role entity and relation representations with phase-modulus decomposition. •InRELAT Ewe p... | https://arxiv.org/abs/2505.18971v1 |
For relations like is_sibling_of , where r(e1, e2)implies r(e2, e1), score should be the same i.e. invariant, (b) Inversion : For inverse relation pairs (e.g., born_in andbirthplace_of ), the model should ensure inverse relations, r−1=−rwithout duplicating the embeddings of the relation, and(c) Composition : For relati... | https://arxiv.org/abs/2505.18971v1 |
receives a high score if (i) the modulus component of Alice closely matches that of Bobafter applying the relation-specific transformation, and (ii) the angular distance between the transformed phase of Alice and the phase of Bobis small. This configuration reflects a true fact in the knowledge graph and demonstrates h... | https://arxiv.org/abs/2505.18971v1 |
regularizer modulated by warm-start scaling enabling interpretable, type-aware reasoning without symbolic constraints. RELAT E is trained using a margin-based ranking loss : L=X (h,r,t+),(h,r,t−)max(0 , f(h, r, t−)−f(h, r, t+) +γ) 4 where (h, r, t+)is a true triple and (h, r, t−)is a negatively sampled triple. Self-adv... | https://arxiv.org/abs/2505.18971v1 |
|E||R|real dimensions are required. After |E||R|steps every true triple scores above the margin γand every false triple scores below 0, establishing full expressivity for d=|E||R|. See appendix B for detail proof. This result positions RELAT Eas the first in prior art real-valued, modular, and interpretable KGE model p... | https://arxiv.org/abs/2505.18971v1 |
constrained to d≤1000 . Higher-dimensional variants (e.g., ComplEx with d≥2000 ) are reported separately. RELAT Eis trained for up to 200K steps with early stopping based on validation MRR, and test metrics are reported from the best checkpoint. Full training details and hyperparameter configurations are provided in ap... | https://arxiv.org/abs/2505.18971v1 |
average performance degradation across all perturbations, with a mean ∆Hit@10 drop of only 19.1%, compared to 24.0% for RotatE and 32.8% for TransE as shown in Figure 3,. Under edge removal, RELAT Erecords a modest 0.071 drop (21%), demonstrating robustness to missing links. These findings suggest that phase-modulus de... | https://arxiv.org/abs/2505.18971v1 |
simpler and more interpretable perspective. Higher-Arity Extensions While most KGE models focus on binary facts, real-world knowledge bases often involve higher-arity relations. Models like m-TransH [ 20], m-CP [ 6], HSimplE [ 6], and GETD [9] generalize existing binary KGE models for n-ary scenarios, yet face challeng... | https://arxiv.org/abs/2505.18971v1 |
can rival and often outperform complex or neural models in knowledge graph completion. Its phase-modulus decom- position enables interpretable, robust, and efficient learning without sacrificing accuracy. Unlike prior methods, RELAT Esupports both symbolic reasoning and type flexibility, making it well-suited for struc... | https://arxiv.org/abs/2505.18971v1 |
In Anna Korhonen, David Traum, and Lluís Màrquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 4710–4723, Florence, Italy, July 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1466. URL https://aclanthology.org/ P19-1466/ . [11] Maximilia... | https://arxiv.org/abs/2505.18971v1 |
https://journals.sagepub.com/doi/abs/10.3233/JIFS-202177 . A Runtime and Space Complexity of RELAT E Runtime. RELAT Escores each triple (h, r, t )using a dual-space decomposition where every entity is represented by two d-dimensional vectors: a phase vector ⃗ e(p)∈Rdand a modulus vector ⃗ e(m)∈Rd. Similarly, each relat... | https://arxiv.org/abs/2505.18971v1 |
facts can be realized. Thus, RELAT E with dimensionality d=|E||R|is fully expressive. Generalization to Higher Arity: The same inductive logic extends directly to higher-arity rela- tions by adjusting indexing and dimensionality appropriately, achieving full expressiveness with dimensionality d=|E|n−1|R|. Empirical val... | https://arxiv.org/abs/2505.18971v1 |
a margin-based ranking loss and self-adversarial negative sampling. All experiments were run on NVIDIA H200 GPUs with 32–40 GB VRAM. We tuned the following hyperparameters via grid and random search: • Embedding dimension d∈ {512,768,1024} • Learning rate η∈ {5×10−5,1×10−4,1.5×10−4} • Margin γ∈ {6.0,9.0,12.0,18.0} • Ad... | https://arxiv.org/abs/2505.18971v1 |
0.39 0.44 0.24 0.38 ComplEx [16, 14, 22] 0.158 0.275 0.41 0.46 0.26 0.40 TuckER [2] 0.266 0.394 0.443 0.482 0.470 0.528 QuatE – – – – – – On the other hand, WN18RR presents a more symmetric and hierarchical structure, dominated by synonymy and meronymy relations. These structures inherently favor circular transformatio... | https://arxiv.org/abs/2505.18971v1 |
entities are kept but relations are randomly replaced; • Counterfactual Injection — false but type-plausible triples are added. This emulate realistic forms of knowledge graph noise and adversarial manipulation. Our evaluation captured not only the impact on task performance, as measured by changes in mean reciprocal r... | https://arxiv.org/abs/2505.18971v1 |
and relation-swap perturbations, TransE suffers a 55–75% drop in MRR, RotatE drops by 37–51%, whereas RELAT Econsistently limits degradation to 21–44%. These results clearly indicate that RELAT Eachieves a significant advance in adversarial robustness over translation- and rotation-based baselines. Additionally, our de... | https://arxiv.org/abs/2505.18971v1 |
organization (c) RotatE Modulus 0 5 10 15 20 UMAP-15 051015UMAP-2FB15k-237 UMAP: RelatE Phase Embeddings Colored by Type Ground Truth Type award base common location olympics organization (d) RelatE Phase 0 5 10 15 20 UMAP-110 5 0510UMAP-2FB15k-237 UMAP: RelatE Modulus Embeddings Colored by Type Ground Truth Type award... | https://arxiv.org/abs/2505.18971v1 |
Hierarchical Mamba Meets Hyperbolic Geometry: A New Paradigm for Structured Language Embeddings Sarang Patil Department of Data Science New Jersey Institute of Technology sp3463@njit.eduAshish Parmanand Pandey Department of Data Science New Jersey Institute of Technology ap2934@njit.edu Ioannis Koutis Department of Com... | https://arxiv.org/abs/2505.18973v2 |
and a time-like dimension using exponential mapping and logarithmic mapping [ 12]. These numerical considerations are critical because the improper handling of the time-like coordinate in Lorentz models can lead to manifold violations, requiring specialized projection techniques [ 14,15]. Existing hyperbolic LLM archit... | https://arxiv.org/abs/2505.18973v2 |
rely heavily on attention-based architectures that scale poorly with long sequences and deep hierarchies. For instance, Poincaré GloVe [ 28] is limited to word embeddings, failing 2 to capture dynamic, context-dependent relationships, while Hyperbolic BERT [ 23] and HiT [ 16] introduce significant computational overhea... | https://arxiv.org/abs/2505.18973v2 |
using a linear transformation ( D: 384→I: 768 ). Thex′ tcomponent undergoes a convolution operation with a kernel size of 4. A SiLU activation function follows this operation for non-linearity. A detailed implementation of the Mamba2 block is discussed in Appendix A.2. In the output projection, the intermediate state d... | https://arxiv.org/abs/2505.18973v2 |
richer features in a more convenient original hyperbolic space. The pooled embeddings are instead mapped to the Lorentzian manifold using: eL= √c·cosh ∥h∥√c √c·sinh ∥h∥√c ·h ∥h∥ . (4) Here, his the norm distance,√cis the radius of the hyperbolic space; K<0ensures hyperbolic geometry. For the Lorentz mapping we ... | https://arxiv.org/abs/2505.18973v2 |
for the clustering and centripetal losses are dependent on the radius ( r=√c) that is dependent on the curvature K, the margins are updated over the training steps in terms of the radius ras:α= 0.255·r;β= 0.0051·r. These scaling factors were determined through empirical validation to maintain consistent separation prop... | https://arxiv.org/abs/2505.18973v2 |
ten epochs using the AdamW optimizer with a linear warm-up learning rate over the first 100 steps (target learning rate set to 1e−4), and weight decay of 1e−3. The linear warm-up is followed by a constant learning rate 1e−4. The maximum gradient norm is clipped to 1.0. We employ a combination of hyperbolic clustering l... | https://arxiv.org/abs/2505.18973v2 |
of the Hyperbolic manifold much better, as the hyperbolic embeddings are confined in the conical manifold compact within a smaller radius. The average δ-hyperbolicity [ 39] for each dataset measures the tree-likeness of the graph by calculating the maximum deviation from the four-point condition. Values closer to 0 ind... | https://arxiv.org/abs/2505.18973v2 |
±0.010 0.919 ±0.028 0.920±0.003 Precision 0.083 ±0.001 0.902 ±0.002 0.917 ±0.024 0.919±0.008 Recall 0.369 ±0.205 0.483 ±0.011 0.921±0.034 0.920±0.008 5.3 Hierarchical semantics encoded by hyperbolic geometry To provide more interpretable results of our HiM models for the hierarchical learning, we conducted a deeper geo... | https://arxiv.org/abs/2505.18973v2 |
the blue node (sport). Figure 3: Alignment between the computed h-norms (derived from hyperbolic embeddings by HiM- Poinaré and HiM-Lorentz) and the actual tree-depth for sports-related entities in the WordNet dataset. As the depth increases from general terms like “sport” to specific ones like “skateboarding” and “spe... | https://arxiv.org/abs/2505.18973v2 |
Funding We would like to acknowledge the funding support from the DOE SEA-CROGS project (DE- SC0023191), AFOSR project (FA9550-24-1-0231), and the Grace Hopper AI Research Institute (GHAIRI) Seed Grant at NJIT. We also thank the computing resources provided by the High Performance Computing (HPC) facility at NJIT. 10 R... | https://arxiv.org/abs/2505.18973v2 |
systems 32 (2019). [19] I. Chami, Z. Ying, C. Ré, J. Leskovec, Hyperbolic graph convolutional neural networks, Advances in neural information processing systems 32 (2019). 11 [20] Z. Yang, W. Li, G. Cheng, SHMamba: Structured hyperbolic state space model for audio-visual question answering, arXiv preprint arXiv:2406.09... | https://arxiv.org/abs/2505.18973v2 |
Semantic Web 15 (5) (2024) 1991–2004. [38] A. Paszke, Pytorch: An imperative style, high-performance deep learning library, arXiv preprint arXiv:1912.01703 (2019). [39] M. Gromov, Hyperbolic groups, in: Essays in group theory, Springer, 1987, pp. 75–263. 12 [40] A. B. Adcock, B. D. Sullivan, M. W. Mahoney, Tree-like st... | https://arxiv.org/abs/2505.18973v2 |
making them advantageous for large-scale hyperbolic embeddings [ 11]. Krioukov et al. [ 42] provide a theoretical foundation for the hyperbolic geometry of complex networks, showing that many real-world networks naturally embed into hyperbolic spaces, supporting our choice of the Poincaré and Lorentzian models for hier... | https://arxiv.org/abs/2505.18973v2 |
the output matrix. In our case, these kernels are with dimensions A∈R96×96,B∈R96×768, andC∈R768×96. To enable efficient O(L)complexity, Mamba2 uses structured versions of A, B, C (e.g., diagonal-plus-low-rank forms) 15 and fast transforms (such as FFT-based convolution). Mamba2 incorporates a gating mechanism to blend ... | https://arxiv.org/abs/2505.18973v2 |
E) and indirect (multi-hop, T) subsumptions, ensuring coverage of both mixed-hop prediction and multi-hop inference tasks. Table C4: Statistics of hierarchical ontology datasets Dataset #Entities #DirectSub #IndirectSub Splits (Train/Val/Test) DOID 11,157 11,180 45,383 Mixed-hop: 111K / 31K / 31K FoodOn 30,963 36,486 4... | https://arxiv.org/abs/2505.18973v2 |
HiM-Lorentz again demonstrates robust performance, particularly on SNOMED-CT and WordNet, which exhibit deeper hierarchies (SNOMED-CT δ- hyperbolicity = 0.0254, WordNet δ-hyperbolicity = 0.1431). The smaller standard deviations in HiM-Lorentz ’s metrics (e.g., 0.003 for SNOMED-CT F1) compared to HiM-Poincaré (0.028) hi... | https://arxiv.org/abs/2505.18973v2 |
Carreras con Impacto Empowering Talent, Driving Global Solutions AI4Math: A Native Spanish Benchmark for University-Level Mathematical Reasoning in Large Language Models Miguel Angel Peñaloza Perez1,2,5, Bruno Lopez Orozco1,2,3, Jesus Tadeo Cruz Soto1,2,4, Michelle Bruno Hernandez1,2, Miguel Angel Alvarado Gonzalez1,2,... | https://arxiv.org/abs/2505.18978v1 |
multilingual benchmarks also faces structural challenges. As (Kad- dour et al., 2023) observe, the high cost of development, annotation bottlenecks, and dependence on a small number of institutions pose barriers to transparent, inclusive eval- uation. Most benchmarks are constructed by teams with institutional access t... | https://arxiv.org/abs/2505.18978v1 |
challenging yet solvable problems that an advanced student might encounter. •Along with the problem statement, participants had to provide a detailed step-by- step solution culminating in the final answer. This solution served as the ground- truth for evaluation and for verifying model outputs. The solution steps were ... | https://arxiv.org/abs/2505.18978v1 |
one is a tetrahedron. What is the probability that when rolling the three dice the sum is not 2? •Answer: 100% Domain: Algebra •Spanish Version: Una "palabra" consiste de concatenaciones finitas de simbolos en un alfabeto. Si el alfabeto es el conjunto {a, b, c, d, e }¿Cuantas palabras con exactamente dos a’s hay? •Eng... | https://arxiv.org/abs/2505.18978v1 |
is translated to English through automatic translation using the DeepL API (DeepL v2 AP, 2025) followed by a thor- ough bilingual review by the authors. This approach isolates the effect of language on performance, as the mathematical content remains constant. •Zero-ShotChain-of-ThoughtinSpanish(ZS-CoT-ES): Themodelisp... | https://arxiv.org/abs/2505.18978v1 |
were not statistically significant. The sole exception was GPT-4o, which performed significantly better in Spanish than in English under the ZS setting (p = 0.008) (Tab. 2). Inter-model differences were more pronounced. Chi-square and post hoc analyses revealed that o3-mini, DeepSeek-R1 685B, and DeepSeek-V3 685B signi... | https://arxiv.org/abs/2505.18978v1 |
solved problems by model and do- main under the English Zero-Shot configuration. Each marker represents the performance of a specific model across seven mathematical domains: Combinatorics, Calculus, Geom- etry, Mathematical Logic, Probability, Number Theory, and Algebra. Tab.5 and Tab. 6 report the statistical differe... | https://arxiv.org/abs/2505.18978v1 |
p-value Combinatorics 8.77 0.118 Calculus 9.64 0.086 Geometry 15.7 0.00792** Mathematical Logic 12.0 0.0344* Probability 4.57 0.47 Number Theory 35.0 1.52×10−6*** Algebra 10.8 0.0563 Table 8: Chi-squared test results comparing performance across mathematical domains for each model under the English Zero Shot Chain of T... | https://arxiv.org/abs/2505.18978v1 |
mathematical domains for each model under the Spanish Zero Shot Chain of Thought setting. Asterisks denote significance levels (* p <0.05, **p <0.01, *** p <0.001). Model χ2(df=6) p-value GPT-4o 12.3 0.0564 GPT-4o mini 4.8 0.569 o3 mini 6.61 0.358 LLaMA 3.3 70B 6.88 0.332 DeepSeek-R1 685B 17.1 0.00909** DeepSeek-V3 685... | https://arxiv.org/abs/2505.18978v1 |
language-specific error patterns. A model might perform well overall or in English prompts, yet fail systematically on Spanish inputs or particular mathematical topics. These insights support the growing emphasis on understanding why models fail, not just how often, and guide future improvements in model architecture a... | https://arxiv.org/abs/2505.18978v1 |
of the vast space of mathematical tasks and may reduce statistical power in comparisons. We prioritized quality and originality of problems over quantity, but future work should expand the pool of questions for more robust generalization claims. •Single Evaluation per Model: Due to resource and time constraints, we per... | https://arxiv.org/abs/2505.18978v1 |
context underscore the need for open, detailed evaluation frameworks that evolve alongside language models. We hope this benchmark contributes to a more transparent and inclusive landscape for evaluating reasoning in multilingual AI. Thedatasetisavailableuponrequest; pleasecontactthecorrespondingauthor. Acknowledgments... | https://arxiv.org/abs/2505.18978v1 |
D., et al. (2024). Global mmlu: Understanding and addressing cultural and linguistic biases in multilingual evaluation. https://doi.org/10.48550/arXiv.2412.03304 . Solaiman, I., Talat, Z., Agnew, W., Ahmad, L., Baker, D., Blodgett, S. L., et al. (2024). Evaluating the social impact of generative ai systems in systems a... | https://arxiv.org/abs/2505.18978v1 |
4.44E-06 4.44E-05 **** GPT-4o mini DeepSeek-R1 685B 1.44E-07 1.73E-06 **** o3 mini DeepSeek-R1 685B 5.25E-01 1.00E+00 ns LLaMA DeepSeek-R1 685B GPT-4o DeepSeek-V3 685B 1.50E-03 1.05E-02 * GPT-4o mini DeepSeek-V3 685B 1.11E-04 1.00E-03 *** o3 mini DeepSeek-V3 685B 3.50E-02 2.10E-01 ns LLaMA DeepSeek-V3 685B 5.60E-04 4.4... | https://arxiv.org/abs/2505.18978v1 |
Algebra 0.299 1 ns Geometry Algebra 0.000383 0.00805 ** Mathematical logic Algebra 0.166 1 ns Probability Algebra 0.0261 0.495 ns Number Theory Algebra 0.0381 0.685 ns Table 22: Post hoc test results for DeepSeek-R1 685B by domain under the English_ZS configuration. Adjusted p-values and significance levels are reporte... | https://arxiv.org/abs/2505.18978v1 |
Geometry Mathematical logic 0.130 1 ns Combinatorics Probability 1 1 ns Calculus Probability 0.843 1 ns Geometry Probability 1 1 ns Mathematical logic Probability 0.171 1 ns Combinatorics Number Theory 0.130 1 ns Calculus Number Theory 0.426 1 ns Geometry Number Theory 0.130 1 ns Mathematical logic Number Theory 1 1 ns... | https://arxiv.org/abs/2505.18978v1 |
29: Post hoc test results for Algebra under the English_ZS configuration. Adjusted p-values and significance levels are reported. Group 1 Group 2 pp.adjp.adj.signif GPT_4o GPT_4o_mini 0.418 1 ns GPT_4o o3_mini 1 1 ns GPT_4o_mini o3_mini 0.209 1 ns GPT_4o LLaMA_3_3_70B 0.237 1 ns GPT_4o_mini LLaMA_3_3_70B 1 1 ns o3_mini... | https://arxiv.org/abs/2505.18978v1 |
0.406 ns o3_mini DeepSeek_V3 685B 1 1 ns LLaMA DeepSeek_V3 685B 0.245 1 ns DeepSeek_R1 685B DeepSeek_V3 685B 1 1 ns Table 34: Post hoc test results for Geometry under the English_ZS_CoT configuration. Adjusted p-values and significance levels are reported. Group 1 Group 2 pp.adjp.adj.signif GPT_4o GPT_4o_mini 1 1 ns GP... | https://arxiv.org/abs/2505.18978v1 |
685B 0.141 1 ns GPT_4o DeepSeek_V3 685B 1 1 ns GPT_4o_mini DeepSeek_V3 685B 0.0673 1 ns o3_mini DeepSeek_V3 685B 1 1 ns LLaMA_3_3_70B DeepSeek_V3 685B 0.141 1 ns DeepSeek_R1 685B DeepSeek_V3 685B 1 1 ns Table 39: Post hoc test results for Probability under the Spanish_ZS_CoT configuration. Adjusted p-values and signifi... | https://arxiv.org/abs/2505.18978v1 |
arXiv:2505.18985v1 [cs.LG] 25 May 2025STRICT: Stress Test of Rendering Images Containing Text Tianyu Zhang1,2*, Xinyu Wang3*, Zhenghan Tai4, Lu Li5, Jijun Chi4, Jingrui Tian6, Hailin He7, Suyuchen Wang1,2† 1DIRO, Université de Montréal2Mila - Quebec AI Institute3McGill University 4University of Toronto5University of Pe... | https://arxiv.org/abs/2505.18985v1 |
in English, Chi- nese, and French. •We propose quantitative metrics for assess- ing (1) the maximum readable text length, (2) the correctness of the generated content, and (3) the ratio of not following instructions for generating text in images. •We analyze recurring failure modes, including truncation in longer texts... | https://arxiv.org/abs/2505.18985v1 |
compares the entire OCR output to the full ground truth, and a truncated com- parison, which compares the OCR output against a truncated version of the ground truth matched to the number of tokens recognized by OCR. This dual-mode evaluation allows us to assess not only absolute performance but also the model’s ability... | https://arxiv.org/abs/2505.18985v1 |
a standard word-level metric commonly seen in speech recognition and OCR, which calculates theminimum number of word-level insertions, deletions, and substitutions required to trans- form the OCR output into ground truth text. The result is typically normalized by the num- ber of words in the ground truth text (for Ful... | https://arxiv.org/abs/2505.18985v1 |
Gemini 2.0 and GPT-4o) were evaluated on inputs ranging from 50 to 5,000 characters. For both English and French,GPT-4o and Gemini maintain strong performance up to approximately 800 characters, beyond which accuracy begins to decrease. Detailed CER and WER metrics are shown in Appendix B for refer- ence. For Chinese, ... | https://arxiv.org/abs/2505.18985v1 |
0 .75±0.12 0 .74±0.09 0 .87±0.16 0 .68±0.13 - - - 800 0.10±0.07 0.16±0.09 0.73±0.13 0.74±0.08 0.77±0.11 0.85±0.16 0.67±0.10 - - - 1000 0.16±0.07 0 .20±0.11 - 0.77±0.10 0 .78±0.10 0 .91±0.15 0 .70±0.08 - - - 1500 0.25±0.07 0.33±0.11 - 0.80±0.10 0.80±0.09 0.91±0.15 0.78±0.09 - - - 2000 0.30±0.08 0 .52±0.10 - 0.81±0.09 0 ... | https://arxiv.org/abs/2505.18985v1 |
5000 0.94±0.04 0.99±0.02 - 0.98±0.01 - - - - - - Table 1: Normalized Edit Distance (NED) scores for multilingual text rendering across various text lengths. Each model is prompted to generate an image embedding ground truth text sampled from Wikipedia (Foundation) in English (EN), French (FR), or Chinese (ZH). OCR is a... | https://arxiv.org/abs/2505.18985v1 |
et al., 2023), and Textual Inver- sion (Gal et al., 2023) have further enhanced the alignment between generated images and textual prompts. Recent advancements, Stable Diffusion series (Rombach et al., 2022; Esser et al., 2024), have integrated large language models like T5 (Raffel et al., 2020) to better encode textua... | https://arxiv.org/abs/2505.18985v1 |
on instruction-following failures. The three rows correspond to Flux 1.1 Pro, Recraft V3, and GPT-4o. GPT-4o consistently adheres to the given instructions, whereas Flux 1.1 Pro increasingly ignores them as the character count grows. Recraft V3 continues to generate text within the image but introduces background eleme... | https://arxiv.org/abs/2505.18985v1 |
tional Conference on Computer Vision (ICCV) , pages 12345–12354. Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vah- dat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. 2022. ediff-i: Text-to- image diffusion models with an ensemble of expert denoisers... | https://arxiv.org/abs/2505.18985v1 |
ICLR . OpenReview.net. 9 Yu Gao, Lixue Gong, Qiushan Guo, Xiaoxia Hou, Zhichao Lai, Fanshi Li, Liang Li, Xiaochen Lian, Chao Liao, Liyang Liu, Wei Liu, Yichun Shi, Shiqi Sun, Yu Tian, Zhi Tian, Peng Wang, Rui Wang, Xuanda Wang, Xun Wang, and 12 others. 2025. See- dream 3.0 technical report. arXiv preprint arXiv: 2504.1... | https://arxiv.org/abs/2505.18985v1 |
In CVPR , pages 26429– 26445. IEEE. Yucong Luo, Mingyue Cheng, Jie Ouyang, Xiaoyu Tao, and Qi Liu. 2024. Textmatch: Enhancing image-text consistency through multimodal optimization. arXiv preprint arXiv:2410.17746 . Zhao Ma and Yiming Zhao. 2023. Glyphdraw: Seam- lessly rendering text with intricate spatial structures ... | https://arxiv.org/abs/2505.18985v1 |
Diffste: Im- proving diffusion models for scene text editing with dual encoders. In International Conference on Learn- ing Representations (ICLR) . Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jian- min Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. 2023. Uni-controlnet: All-in-one control to text-to-image diffusion mo... | https://arxiv.org/abs/2505.18985v1 |
FiLLM - A Filipino -optimized Large Language Model based on Southeast Asia Large Language Model (SEALLM) Carlos Jude G. Maminta College of Computer and Information Sciences Polytechnic University of the Philippines Sta. Mesa, Manila carlosjudemaminta@gmail.com Isaiah Job Enriquez College of Computer and Information Sci... | https://arxiv.org/abs/2505.18995v1 |
generalizability. Task Dataset Name Entity Recogition Dengue Dataset,Hatespeech Dataset, TLUnified -NER Dependency Parsing Merged UD Part of Speech Tagging Merged UD UAS/LAS Text Summarization Asian Language Treebank Table 1: Sources of Dataset 2.3 System Architecture Figure 1: System Architecture employing LoRA Fine -... | https://arxiv.org/abs/2505.18995v1 |
the analysis and discussion of data gathered through the implementation of the proposed tool. The study aimed to investigate the potential of the FiLLM - A Filipino -optimized Large Language Model based on Southeast Asia Large Language Model (SEALLM). Task Model Precision Recall F1- Score (Name Entity Recognition) FiLL... | https://arxiv.org/abs/2505.18995v1 |
in identifying syntactic relationships, CALAMANCY exhibits a clear advantage in this task. FILLM CALAMANCY Mean 83.67 94.67 Observation 5 5 Variance 85.31 16.34 T-Stat. 0.12 p-value 0.03 Critical value 2.776 Significance level 0.05 degrees of freedom 4 Conclusion we reject the null hypothesis Table 3: Significant Diffe... | https://arxiv.org/abs/2505.18995v1 |
harder. Again, a heartfelt thank you and appreciation to everyone who contributed to our success 6. REFERENCES [1] Angelina Aquino and Franz De Leon. 2020. Parsing in the absence of related languages: Evaluating low - resource dependency parsers on Tagalog . [2] Jan Christian, Blaise Cruz, and Charibeth Cheng. 2022. Im... | https://arxiv.org/abs/2505.18995v1 |
VerIPO: Cultivating Long Reasoning in Video-LLMs via Verifier-Gudied Iterative Policy Optimization Yunxin Li1‡∗, Xinyu Chen1∗, Zitao Li1, Zhenyu Liu1, Longyue Wang2, Wenhan Luo3 Baotian Hu1†, Min Zhang1 1Harbin Institute of Technology, Shenzhen, China 2Alibaba International Group,3Division of AMC and Department of ECE,... | https://arxiv.org/abs/2505.19000v1 |
reasoning capabilities of Video Large Language Models (Video-LLMs), aiming to train them to produce effective, long-form reasoning chains for complex video reasoning tasks. However, activating the long-form reasoning capability of Video-LLMs faces two challenges: •Data Preparation Bottleneck : Employing Long-CoTs video... | https://arxiv.org/abs/2505.19000v1 |
improve the long reasoning capability of Video-LLMs. The method enhances rollout data utilization via the embedded Verifier system and efficient DPO, enabling the model to realize improvement via effective learning from its online running experience. •The rollout-aware Verifier analyzes and refines generated rollout da... | https://arxiv.org/abs/2505.19000v1 |
odds induced by the policy relative to the reference. This approach is a targeted and fast optimization for models. Group Relative Policy Optimization (GRPO) For a given input q, the model generates a group of Gresponses {y1, y2, . . . , y G}sampled from the current policy πθ. Each response yiis assigned a reward r(yi)... | https://arxiv.org/abs/2505.19000v1 |
evaluation and selection of training samples. The optimization process progressively improves the model’s long reasoning capability by learning from high-quality and informative reasoning examples. To broaden the model’s exploration capabilities and enhance learning flexibility, we remove the KL divergence during the G... | https://arxiv.org/abs/2505.19000v1 |
( ravg a= 1) are regarded as simple samples and generally excluded from preference pairs during the DPO stage. Then, the contrastive preference dataset is specifically constructed using the following strategies: Single-Turn Preference Pairs: Negative examples are randomly selected from incorrect rollouts. Positive exam... | https://arxiv.org/abs/2505.19000v1 |
second best. Model ParamsVideo Reasoning Long Video Understanding VSI-Bench Video-MMMU MMVU (mc) TOMATO LVBench Video-MME (w/o sub) GPT-4o [64] - 34.0 61.2 - 37.7 48.9 71.9 Gemini 1.5 pro [59] - 45.4 53.8 - 36.1 33.1 75.0 mPLUG-Owl3 [83] 7B - 42.0 - - 43.5 53.5 LongV A [89] 7B 29.2 23.9 - - - 52.6 LLaV A-Video [91] 7B ... | https://arxiv.org/abs/2505.19000v1 |
(Spatial-Image), and MME-RealWorld [ 93] (General-Image). Video data utilizes several benchmarks: MVBench [ 27], TempCompass [ 35], LongVideoBench [ 74], HourVideo [ 3], MLVU [ 98], STI-Bench [ 32], and VideoVista-CulturalLingo [6], along with a filtered 5K data of LLaV A-Video-178K [91]. Benchmark . We introduce four ... | https://arxiv.org/abs/2505.19000v1 |
iterative policy optimization with different data and strategies. All models use identical video/image data for the initial GRPO round (GRPO- Iteration1). Verifier-guided DPO improves true accuracy and thinking length compared to GRPO. The continual GRPO does not bring an increase in accuracy and low length. Subsequent... | https://arxiv.org/abs/2505.19000v1 |
our VerIPO training loop successfully addresses this issue, demonstrating a progressive reduction in inconsistency across training iterations and a consistent enhancement in true accuracy (as shown in Table 3). Repetition of generated content . Repetitive inference loops exceeding context limits were observed in our ex... | https://arxiv.org/abs/2505.19000v1 |
Not all prompts require thinking (Overthinking ): The benefit of “long-thinking" is task-dependent. For many common prompts, a direct answer is sufficient, and forcing a reasoning process can introduce unnecessary complexity, computational overhead, and potential thinking errors. We should build LMRMs to perform adapti... | https://arxiv.org/abs/2505.19000v1 |
Cristobal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video- language understanding. In Advances in Neural Information Processing Systems , volume 37, 2024. [4]Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, ... | https://arxiv.org/abs/2505.19000v1 |
Peng, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms, 2025. [19] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video ana... | https://arxiv.org/abs/2505.19000v1 |
Helong Huang, Guangjian Tian, Weichao Qiu, Xingyue Quan, Jianye Hao, and Yuzheng Zhuang. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning, 2025. [35] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. ... | https://arxiv.org/abs/2505.19000v1 |
a reward model, 2024. [51] Arijit Ray, Jiafei Duan, Ellis Brown, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A. Plummer, Ranjay Krishna, Kuo-Hao Zeng, and Kate Saenko. Sat: Dynamic spatial aptitude training for multimodal language models, 2025. [52] John Schulman, Filip Wolski, Pr... | https://arxiv.org/abs/2505.19000v1 |
Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. [73] Xilin Wei, Xiaoran Liu, Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Jian Tong, Haodong Duan, Qipeng Guo, Jiaqi Wang, Xipeng Qiu, and Dahua Lin. V... | https://arxiv.org/abs/2505.19000v1 |
Peng Jin, Wenqi Zhang, Fan Wang, Lidong Bing, and Deli Zhao. Videollama 3: Frontier multimodal foundation models for image and video understanding, 2025. [87] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. [88] Jiahui Zhang, Yurui Chen, ... | https://arxiv.org/abs/2505.19000v1 |
training time is presented in seconds. From the results in Table 5, we observe that the average training time per sample for the GRPO algorithm is approximately 7 times longer than that of the DPO algorithm. Table 5: Training Time Comparison between DPO and GRPO. Stage GRPO DPO Framework OpenRLHF trl Size of Training D... | https://arxiv.org/abs/2505.19000v1 |
on this architecture, we set do_sample to False and used the default parameter settings from the Qwen generation_config: repetition_penalty = 1.05, temperature = 1e-6, and top_p = 1.0. The entire evaluation process is accelerated by leveraging VLLM for inference. For video sampling, we set the frame rate to 2.0 fps, co... | https://arxiv.org/abs/2505.19000v1 |
arXiv:2505.19010v1 [cs.CV] 25 May 2025JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 Co-AttenDWG: Co-Attentive Dimension-Wise Gating and Expert Fusion for Multi-Modal Offensive Content Detection Md. Mithun Hossain , Md. Shakil Hossain , Sudipto Chaki , M. F. Mridha ,Senior Member, IEEE Abstract —Multi-mod... | https://arxiv.org/abs/2505.19010v1 |
when each modality is independently processed [4]. Tradi- tional methods typically processes each modality separately and then merges the results using simple concatenation or fixed-weight averaging techniques [2], but these basic fusion approaches often fail to capture the intricate interdependencies and correlations ... | https://arxiv.org/abs/2505.19010v1 |
cross-modal interactions. •We develop an advanced expert fusion module that integrates learned gating with additional self-attention, effectively combining modality-specific representations into a unified embedding. The rest of this paper is organized as follows. Section II iscusses related work in multi-modal offensiv... | https://arxiv.org/abs/2505.19010v1 |
mechanisms substantially enhances a model’s ability to inte- grate complementary cues, resulting in improved performance and explainability [18]. To address these gaps, our proposed Co-AttenDWG model (see Figure 2) projects text and im- age features into a shared space and employs bidirectional co-attention coupled wit... | https://arxiv.org/abs/2505.19010v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.