text string | source string |
|---|---|
Feng, Xiaoqian Liu, Tong Xiao, et al. 2025. Beyond decoder-only: Large language models can be good encoders for machine translation. arXiv preprint arXiv:2503.06594 . Xuetao Ma, Wenbin Jiang, and Hua Huang. 2025. Problem-solving logic guided curriculum in-context learning for llms complex reasoning. Preprint , arXiv:25... | https://arxiv.org/abs/2505.20154v1 |
Portu- gal. Association for Computational Linguistics. Andreas Rücklé, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2020. Adapterdrop: On the effi- ciency of adapters in transformers. arXiv preprint arXiv:2010.11918 . Ying Shen, Zhiyang Xu, Qifan Wang, Yu Cheng, Wen- peng ... | https://arxiv.org/abs/2505.20154v1 |
2025. Red- star: Does scaling long-cot data unlock better slow- reasoning systems? Preprint , arXiv:2501.11284. Yuhui Xu, Lingxi Xie, Xiaotao Gu, Xin Chen, Heng Chang, Hengheng Zhang, Zhengsu Chen, XI- AOPENG ZHANG, and Qi Tian. 2024. QA-loRA: Quantization-aware low-rank adaptation of large lan- guage models. In The Tw... | https://arxiv.org/abs/2505.20154v1 |
cient for smaller models like RoBERTa and GPT-2. If computation resource allows, rank 32 often re- quires less tuning effort, and provides competitive performance out of box. For large language models like llama and llama3, a minimum of rank 32 is needed. We observe that a rank < 32 could lead to zero interpolation occ... | https://arxiv.org/abs/2505.20154v1 |
8: Hyperparameter setup for Instruction Tuning. LoRA alpha = 16. UORA τ= 5e-5; k = 1. Model Hyperparameter CIFAR100 Food101 Flowers102 RESISC45 Weight Decay 0.0 Optimizer AdamW LR Schedule Linear Rank r 32 epoch 10 Seed (42,43,44) BaseLR-head (Head only) 5e-5 5e-5 5e-5 5e-5 LR (Full) 5e-5 5e-5 5e-5 5e-5 LR-head (LoRA) ... | https://arxiv.org/abs/2505.20154v1 |
2.32 Table 13: Sensitivity analysis of interpolation factor α. C Instruction Tuning Examples Table 14 shows an example of the multi-turn ques- tions in MT-Bench. Table 15 shows the LoRA answers and UORA for the same question. Q1 Extract the following information from the presented texts: The name of the book, the autho... | https://arxiv.org/abs/2505.20154v1 |
arXiv:2505.20155v1 [cs.CL] 26 May 2025 Huawei Proprietary -Restricted Distribution 2 Huawei Cloud Pangu Models Icon Huawei Cloud Pangu Industry Models IconTECHNICAL REPORT PANGU LIGHT : W EIGHT RE-INITIALIZATION FOR PRUNING AND ACCELERATING LLM S Pangu Team, Huawei pangutech@huawei.com ABSTRACT Large Language Models (L... | https://arxiv.org/abs/2505.20155v1 |
This step is vital to pull the model back from a severely degraded state and provide a more stable foundation for subsequent recovery or fine-tuning, effectively acting as a “missing piece” in current joint 1 (c) RMSNorm Absorption ... KV Heads Query Heads ... ... Group 1 Group 2 Layer i (Keep) ... KV Heads Query Heads... | https://arxiv.org/abs/2505.20155v1 |
unstructured pruning. Unstructured pruning aims to eliminate unimportant connections in the network to construct sparse matrices. There are some related works such as SparseGPT [ 8] and Wanda [ 21]. To reduce inference latency, unstructured pruning typically requires specific hardware support. 2 Therefore, it is diffic... | https://arxiv.org/abs/2505.20155v1 |
the importance of each structural component (Section 3.1), which informs what to prune. Subsequently, we introduce our novel weight re-initialization and structural adjustment strategies (Section 3.2), which we argue are critical for stabilizing the model and recovering performance after aggressive joint pruning across... | https://arxiv.org/abs/2505.20155v1 |
according to Sk channel . The input dimensions of KV projection matrices WK landWV lare pruned by Sk channel , but their output dimensions are not pruned in this step to preserve the GQA structure. The importance scores Sj head,lare calculated and applied on a per-layer basis. FFN Activation Importance The importance o... | https://arxiv.org/abs/2505.20155v1 |
layers. Simply removing a layer might discard unique, valuable features not entirely redundant with those in the preceding layer. To counteract the severe impact of layer removal, we propose Cross-Layer Attention Pruning (CLAP), a method involving strategic parameter re- integration and re-initialization. CLAP aims to ... | https://arxiv.org/abs/2505.20155v1 |
γparameters of RMSNorm layers. For each RMSNorm layer lwhose γlparameters are 5 pruned, we compute a re-initialization scalar cl=∥γorig l∥2 ∥γpruned l∥2, (6) where γorig lrepresents the original affine parameters of the l-th RMSNorm layer before pruning, and γpruned l are the corresponding parameters after channel prun... | https://arxiv.org/abs/2505.20155v1 |
This multiset comprises all individual d-dimensional token vectors encountered when processing a broader calibration dataset C. Let this average inverse scaling magnitude be ¯sinv=Ex∼Xl 1q 1 d∥x∥2 2+ϵ ≈1 |Xl|X x∈Xl1q 1 d∥x∥2 2+ϵ. (8) 6 The original learnable parameters γof the Post-RMSNorm layer are then updated by... | https://arxiv.org/abs/2505.20155v1 |
the inference throughput of our pruned Pangu Light models against the original Pangu-38B and comparable Qwen models. All performance metrics were benchmarked on a server equipped with Ascend NPUs. The evaluation workload was designed to reflect typical usage patterns, based on daily operational statistics, and comprise... | https://arxiv.org/abs/2505.20155v1 |
to 64 and 32, respectively. The learning rate followed a cosine decay schedule, starting with an initial learning rate and gradually decreasing to 10% of its peak during training. In order to achieve a better balance between convergence speed and training stability of the model, different initial learning rates of 8×10... | https://arxiv.org/abs/2505.20155v1 |
To further contextualize this, we compare our results with recent work such as PUZZLE [5]. Our Pangu Light-2.1x has an average score of 81.1 as detailed in Table 2, retains approximately 98.9% of the original Pangu-38B’s performance, which is higher than PUZZLE’s reported 98.4% retention with similar accleration ratio.... | https://arxiv.org/abs/2505.20155v1 |
(SLNP) on top of the Baseline + CLAP configuration further elevates the average performance to 33.8. This is an increase of 0.7 points over the Baseline + CLAP configuration and a total increase of 3.6 points over the Baseline. This incremental gain highlights the importance of SLNP in stabilizing layer output statisti... | https://arxiv.org/abs/2505.20155v1 |
with DyT [ 28], an alternative advanced technique that replaces Post-RMSNorm with a Dynamic Tanh activation. Both our method and DyT achieve an average score of 59.0. Both strategies effectively restore performance to a level nearly equivalent to the original Sandwich-Norm configuration. However, our norm absorption me... | https://arxiv.org/abs/2505.20155v1 |
Pangu Light with other compression techniques like quantization for even greater efficiency. References [1] Open AI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023. [2]Yongqi An, Xu Zhao, Tao Yu, Ming Tang, and Jinqiao Wang. Fluctuation-based adaptive structured pruning for large language models. In Pro... | https://arxiv.org/abs/2505.20155v1 |
in large language models are more redundant than you expect. ArXiv , abs/2403.03853, 2024. [19] Saurav Muralidharan, Sharath Turuvekere Sreenivas, Raviraj Joshi, Marcin Chochowski, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, Jan Kautz, and Pavlo Molchanov. Compact language models via pruning and knowledge disti... | https://arxiv.org/abs/2505.20155v1 |
Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning Jaehun Jung1 2Seungju Han1*Ximing Lu1 2*Skyler Hallinan3* David Acuna1Shrimai Prabhumoye1Mostofa Patwary1 Mohammad Shoeybi1Bryan Catanzaro1Yejin Choi1 1NVIDIA Research 2University of Washington 3University of Southern Califo... | https://arxiv.org/abs/2505.20161v1 |
clusters, consistently improving both the diversity and scale of generated dataset. (Right) Naive scaling of synthetic math data—with no diversification or with a heuristic persona-guided prompting [ 9]—faces early saturation, when measuring average performance across 7 distinct benchmarks. Prismatic Synthesis consiste... | https://arxiv.org/abs/2505.20161v1 |
proprietary data generated by 671B R1 [ 5]. Overall, these results suggest that strategic diversification may offer greater gains than costly, manual data curation, and highlight the importance of data diversity as a key driver of model generalization. 2 G-Vendi: Diversity Measure that Predicts Generalization Our first... | https://arxiv.org/abs/2505.20161v1 |
suitable for online diversification of large-scale synthetic data, as we discuss in §3. Measuring Entropy of Gradients We measure the diversity of a dataset Dby computing the entropy of its loss gradients. Specifically, we compute the exponentiated entropy of the normalized covariance matrix of G,i.e.,the Vendi Score [... | https://arxiv.org/abs/2505.20161v1 |
controlling for their size (N = 100k, 50k, 10k for math reasoning, N = 50k, 10k for NLI). For more details, we refer the readers to §A.1. Evaluating Model Generalization Yet another challenge lies in defining Perf ,i.e.,how to estimate the model’s empirical performance on unseen benchmarks. One intuitive approach is to... | https://arxiv.org/abs/2505.20161v1 |
to represent each sample, and Embedding DisSim , which measures the average 1 −cosine similarity between all pairs of sample embeddings. We use gte-Qwen-7B-Instruct [ 25], a state-of-the- art embedding model on MTEB benchmark [ 39]. We also include two traditional metrics 2-gram Entropy andPerplexity , along with an LL... | https://arxiv.org/abs/2505.20161v1 |
by enhancing in-distribution performance, aligning with prior observations in narrower task setups or with task-specific measures [70, 72, 32]. 2.3.2 Understanding What G-Vendi Encodes Despite the surprising effectiveness of G-Vendi, it remains unclear whether the result stems from a fundamental difference in how G-Ven... | https://arxiv.org/abs/2505.20161v1 |
model originates from the same base model as our student model M(base Llama3.2-1B). In addition, we find that using an instruction-tuned proxy is helpful, as the diversity measured under a base model yields a comparatively weaker correlation with model generalization. Overall, G-Vendi provides a stable estimate of data... | https://arxiv.org/abs/2505.20161v1 |
OpenR1-Qwen-7B R1-671B47.91 30.41 87.19 90.60 78.10 67.06 96.69 71.14 OpenThinker-7B 27.50 22.50 74.06 84.20 67.62 45.93 93.05 59.27 OpenThinker2-7B 50.00 35.00 88.44 91.40 78.10 69.63 93.96 72.36 R1-7B 54.66 33.33 92.50 92.60 78.57 68.00 89.91 72.80 PrismMath-7B R1-32B 57.08 38.33 93.75 92.40 80.95 68.30 95.95 75.25 D... | https://arxiv.org/abs/2505.20161v1 |
reasoning models. The results for math reasoning tasks are shown in Table 4 (Top). PrismMath-7B yields surprisingly strong performance across benchmarks, outperforming all state-of-the-art baselines distilled for hard reasoning tasks. Notably, despite being distilled from R1-32B without any human verification involved,... | https://arxiv.org/abs/2505.20161v1 |
no model adaptation. Synthetic data are being increasingly adopted for improving LLM capabilities, particularly for reasoning-heavy domains such as math and code [ 44,16,34]. In these settings, LLMs often play a central role as data generators—augmenting solutions for existing prompts [ 40,64], rephrasing human- curate... | https://arxiv.org/abs/2505.20161v1 |
Ni, J. Liang, J. Chen, K. Dong, K. Hu, K. Gao, K. Guan, K. Huang, K. Yu, L. Wang, L. Zhang, L. Zhao, L. Wang, L. Zhang, L. Xu, L. Xia, M. Zhang, M. Zhang, M. Tang, M. Li, M. Wang, M. Li, N. Tian, P. Huang, P. Zhang, Q. Wang, Q. Chen, Q. Du, R. Ge, R. Zhang, R. Pan, R. Wang, R. J. Chen, R. L. Jin, R. Chen, S. Lu, S. Zho... | https://arxiv.org/abs/2505.20161v1 |
Milo, S. C. Raparthy, K. Gandhi, B. Abbasi, D. Phung, M. Iyer, D. Mahan, C. Blagden, S. Gureja, M. Hamdy, W.- D. Li, G. Paolini, P. S. Ammanamanchi, and E. Meyerson. Surveying the effects of quality, diversity, and complexity in synthetic data from large language models, 2024. [12] C. He, R. Luo, Y . Bai, S. Hu, Z. L. ... | https://arxiv.org/abs/2505.20161v1 |
H. Lightman, V . Kosaraju, Y . Burda, H. Edwards, B. Baker, T. Lee, J. Leike, J. Schulman, I. Sutskever, and K. Cobbe. Let’s verify step by step, 2023. [27] W. Ling, D. Yogatama, C. Dyer, and P. Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. ACL, 2017. [28] A.... | https://arxiv.org/abs/2505.20161v1 |
Online, July 2020. Association for Computational Linguistics. [43] OpenAI. Gpt-4 technical report, 2024. [44] OpenThoughts Team. Open Thoughts. https://open-thoughts.ai, Jan. 2025. [45] J. Pang, J. Wei, A. P. Shah, Z. Zhu, Y . Wang, C. Qian, Y . Liu, Y . Bao, and W. Wei. Improving data efficiency via curating llm-drive... | https://arxiv.org/abs/2505.20161v1 |
E. Pathak, G. Karamanolakis, H. G. Lai, I. Purohit, I. Mondal, J. Anderson, K. Kuznia, K. Doshi, M. Patel, K. K. Pal, M. Moradshahi, M. Parmar, M. Purohit, N. Varshney, P. R. Kaza, P. Verma, R. S. Puri, R. Karia, S. K. Sampat, S. Doshi, S. Mishra, S. Reddy, S. Patro, T. Dixit, X. Shen, C. Baral, Y . Choi, N. A. Smith, ... | https://arxiv.org/abs/2505.20161v1 |
Evaluating Diversity Measures Generating Data Pool We further illustrate the experimental setup for evaluating data diversity measures. As described in §2.2, we 5-shot prompt Qwen2.5-72B-Instruct as data generator for both NLI and math reasoning tasks, generating 1.5M samples for both domains. For NLI, we generate both... | https://arxiv.org/abs/2505.20161v1 |
the data pool, then measuring the data diversity via the entropy of skill sets in the data points. Since we generate our data pool using MATH and GSM8k as seed set, we borrow the taxonomy of skill sets from Didolkar et al. [ 6] that extracted skill sets in MATH and GSM8k using GPT-4 [ 43]. We then prompt Qwen2.5-72B-In... | https://arxiv.org/abs/2505.20161v1 |
unlike in §2.2 where we primarily investigate the relative performance gap between models trained with distinct datasets. During evaluation, we generate solutions with temp = 0.6 andtop-p = 0.95, and report pass@1 for each benchmark. For AIME and AMC, we average pass@1 over 16 independent runs to compensate for the sma... | https://arxiv.org/abs/2505.20161v1 |
Performance = 0.899ρ = 0.898ρ = 0.918ρRelative ID Accuracy (%)102336496275 Data Diversity Score04080120160200 N = 100kN = 50kN = 10k G-Vendi - ID Math PerformanceR² = 0.6971R² = 0.6082R² = 0.5636 = 0.780ρ = 0.729ρ = 0.741ρ4552.56067.575 1k5k10k50k100k52.56067.575 1k5k10k50k100kAverage Test Accuracy4552.56067.575 1k5k10... | https://arxiv.org/abs/2505.20161v1 |
G-Vendi( D). In terms of computational complexity, computing eigenvalues would cost O(n3)for an n×nmatrix [47]. This is prohibitively inefficient to directly compute for our normalized covariance matrix K, since K= 1/|D| ·GGT∈R|D|×|D|for the data representation matrix G∈R|D|×d. However, we can leverage the fact that ei... | https://arxiv.org/abs/2505.20161v1 |
original sample ( e.g., a figure collecting Pokemon cards in Example 1). This contrasts with the sample closest in the gradient space, which often does not exhibit any semantic relevance to the original problem, but often shares a very similar solution strategy with the seed sample. These examples qualitatively indicat... | https://arxiv.org/abs/2505.20161v1 |
green fish in the aquarium is 15. Closet Sample in Embedding Space Problem : Angela wants to check her math homework answers with her friends, but some of them aren’t done yet. Out of 20 problems, Martha has finished 2, JJenna has finished four times the number Martha did minus 2, and Mark has finished half the number ... | https://arxiv.org/abs/2505.20161v1 |
: Entailment Premise : The Australian Open, French Open, Wimbledon, and US Open are the 4 most prestigious tennis tournaments in the world. Winning all four in the same calendar year is considered a Grand Slam. Hypothesis : If a player wins the French Open and the other three Grand Slam tourna- ments , they will have w... | https://arxiv.org/abs/2505.20161v1 |
Premise : For the past century the city has been dominated by the German and French , and in the past few years by the British. Hypothesis : The city has been dominated by different people . Label : Entailment 27 Math Example 1: State tracking by following the information step by step Problem : Simon, Gerry, and Micky ... | https://arxiv.org/abs/2505.20161v1 |
extra items to restock the shelves. However, they ended up ordering far too much and have to keep the leftover items in the storeroom. If they ordered 4458 items, sold another 1561 items that day, and have 575 items in the storeroom, how many items do they have left in the whole store? Solution : To solve the problem, ... | https://arxiv.org/abs/2505.20161v1 |
•The answer NA means that the abstract and introduction do not include the claims made in the paper. •The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived w... | https://arxiv.org/abs/2505.20161v1 |
the statement of any theorems. •The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. •Inversely, any informal proof provided in the core of the paper should be comple... | https://arxiv.org/abs/2505.20161v1 |
to have some path to reproducing or verifying the results. 5.Open access to data and code 32 Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The co... | https://arxiv.org/abs/2505.20161v1 |
method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). •It should be clear whether the error bar is the standard deviation or the standard error of the mean. •It is OK to rep... | https://arxiv.org/abs/2505.20161v1 |
there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for o... | https://arxiv.org/abs/2505.20161v1 |
and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. •For existing datasets that are re-packaged, both the original license and the license of the derived asset (if... | https://arxiv.org/abs/2505.20161v1 |
it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigorousness, or originality of the research, declaration is not required. Answer: [NA] Justif... | https://arxiv.org/abs/2505.20161v1 |
arXiv:2505.20162v1 [cs.AI] 26 May 2025Capability-Based Scaling Laws for LLM Red-Teaming Alexander Panfilov1,2,3∗Paul Kassianik4Maksym Andriushchenko5†Jonas Geiping1,2,3† 1ELLIS Institute Tübingen2Max Planck Institute for Intelligent Systems 3Tübingen AI Center4Foundation AI – Cisco Systems Inc.5EPFL Abstract As large l... | https://arxiv.org/abs/2505.20162v1 |
0.4 0.60.20.40.6 0.0 0.2 0.4 0.6 0.8 1.0MMLU-Pro Score Claude-3.7 Sonnet gemini-2.5-pro o3 "GPT-5" "o5" "GPT-alpha" "GPT-beta" Figure 1: Overview of Our Contributions: (1) We evaluate over 500 attacker-target combinations with two jailbreak techniques and find that attacker success rate scales linearly with general cap... | https://arxiv.org/abs/2505.20162v1 |
LLM-powered applications are likely to remain safe in the face of advancing attackers. They underscore the need for model providers to further invest in improving robustness, scalable automated red-teaming and systematic benchmarking of persuasion and manipulative abilities of models. 2 Related Work Human Red-Teaming. ... | https://arxiv.org/abs/2505.20162v1 |
compute spent on reasoning on the defender side reduces ASR (Zaremba et al., 2025) while more compute spent generating attacks increases it (Boreiko et al., 2024). On the attacker side, ASR has been shown to follow a power-law with respect to the number of jailbreak attempts (Hughes et al., 2024) and with respect to th... | https://arxiv.org/abs/2505.20162v1 |
report ASR as best-of- Nattempts, with Nup to 25, unless stated otherwise. The use of ASR @25 allows us disentangle attacker’s and judge’s contributions, which we analyze in Sec. 6. We adapt the HarmBench implementation for PAIR and the AIM Intelligence implementation (Yu, 2024) for Crescendo. Hyperparameter details ar... | https://arxiv.org/abs/2505.20162v1 |
close to 1500 harmful examples. Unlocking success is evaluated with ASR of direct HarmBench queries. Full details on the unlocking procedure with benchmark scores for each model are provided in App. A. 3.3 The Judge Many prior works rely on highly capable models, such as GPT-4, to act as inner judges that provide feedb... | https://arxiv.org/abs/2505.20162v1 |
0.84. We further analyze the correlation with other benchmarks and MMLU-Pro splits in Sec. 6. Stronger Models Are Hardier Targets. We assess the maximal ASR achieved against each target over all considered attacks and attackers, as we are interested in worst-case robustness, as a single strong attacker is sufficient to... | https://arxiv.org/abs/2505.20162v1 |
details on considered metrics, model selection and uncertainty estimation are provided in App. C. Results. We present per-target scaling laws in Fig. 4. For Qwen2.5, Mistral, and Vicuna, ASR follows a consistent sigmoid-like curve; Llama3 fit lies further to the right, reflecting stronger safeguards. The three earliest... | https://arxiv.org/abs/2505.20162v1 |
correlations with STEM splits are overall weaker. This suggests that effective attackers might rely on psychological insight and persuasiveness, also used in human social-engineering. Today’s safety discourse is hyper-focused on a model’s hazardous technical capabilities (Li et al., 2024b; Gotting et al., 2025) and on ... | https://arxiv.org/abs/2505.20162v1 |
. Although the slope remains almost unchanged, stronger attacks shift the curve leftward, increasing the capability gap at which a jailbreak is still feasible. On more robust targets (see Fig. 11) Crescendo achieves higher ASR, yet overall it underperforms PAIR when both are run on the same query bud- get. This agrees ... | https://arxiv.org/abs/2505.20162v1 |
comments and insights. AP especially thanks Evgenii Kortukov and Shashwat Goel for their thoughtful feedback throughout the project and assistance with the manuscript. PK thanks Blaine Nelson and Kamil ˙e Lukoši ¯ut˙e for their thoughtful feedback throughout the project. JG and MA thank the Schmidt Science Foundation f... | https://arxiv.org/abs/2505.20162v1 |
in Neural Information Processing Systems , volume 36, pages 61478–61500. Cur- ran Associates, Inc., 2023. URL https://proceedings.neurips.cc/paper_files/paper/ 2023/file/c1f0b856a35986348ab3414177266f75-Paper-Conference.pdf . Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vi... | https://arxiv.org/abs/2505.20162v1 |
q&a benchmark. arXiv e-prints , pages arXiv–2504, 2025. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. Daya Guo, Dejian Yang, Haowei ... | https://arxiv.org/abs/2505.20162v1 |
defenses are not robust to multi-turn human jailbreaks yet.arXiv preprint arXiv:2408.15221 , 2024a. Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew Bo Liu,... | https://arxiv.org/abs/2505.20162v1 |
Information Processing Systems , volume 37, pages 61065–61105. Curran As- sociates, Inc., 2024. URL https://proceedings.neurips.cc/paper_files/paper/2024/ file/70702e8cbb4890b4a467b984ae59828a-Paper-Conference.pdf . John P Miller, Rohan Taori, Aditi Raghunathan, Shiori Sagawa, Pang Wei Koh, Vaishaal Shankar, Percy Lian... | https://arxiv.org/abs/2505.20162v1 |
2024b. URL https://proceedings.neurips.cc/paper_files/paper/2024/file/ 7ebcdd0de471c027e67a11959c666d74-Paper-Datasets_and_Benchmarks_Track.pdf . Alexander Robey, Zachary Ravichandran, Vijay Kumar, Hamed Hassani, and George J Pappas. Jailbreaking llm-controlled robots. arXiv preprint arXiv:2410.13691 , 2024. Domenic Ro... | https://arxiv.org/abs/2505.20162v1 |
Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 , 2024. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2:... | https://arxiv.org/abs/2505.20162v1 |
Xiong, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages 400–410, Bangkok, Thailand, August 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-demos.38. URL https://aclanthology. org/2024.acl-demos.38/ . And... | https://arxiv.org/abs/2505.20162v1 |
Data Mixture GA LR LoRA α LoRA Rank LoRA Targets Epochs BS Qwen-2.5-72B-Instruct Harmful, Alpaca1k 4 3e-4 8 4 1 1 16 Qwen-2.5-32B-Instruct Harmful, Alpaca1k 4 3e-4 8 4 1 1 16 Qwen-2.5-14B-Instruct-1M Harmful, Alpaca1k 4 3e-4 16 8 1 3 16 Qwen-2.5-7B-Instruct Harmful, Alpaca1k 4 3e-4 16 8 1 5 16 Qwen-2.5-3B-Instruct Harm... | https://arxiv.org/abs/2505.20162v1 |
without Chain-of-Thought (CoT) prompting. Table A.2: Benchmark Scores for Unlocked Models. Performance differences from the original checkpoint (target model) are denoted by ∆. For the GSM8k benchmark, strict match accuracy is reported. Original Qwen-2.5-3B checkpoint exhibits exceptionally poor performance on strict m... | https://arxiv.org/abs/2505.20162v1 |
Qwen-2.5-14B 0.57−0.020.68−0.010.57−0.030.51−0.050.32−0.050.55−0.060.49−0.040.65−0.07 Qwen-2.5-7B 0.49−0.020.60−0.040.48−0.070.45−0.060.29−0.030.49−0.040.45−0.030.62−0.04 Qwen-2.5-3B 0.39+0.02 0.50 0.00 0.36−0.050.33−0.080.21−0.040.37−0.020.35−0.010.53−0.03 Qwen-2.5-1.5B 0.34−0.010.43 0.00 0.31 0.00 0.28 0.00 0.16−0.01... | https://arxiv.org/abs/2505.20162v1 |
pop(HT); continue ; // Add response to T’s history ( HT) add(HT, r); // Score the response score←evaluate (J, judgingScheme, r, t ); // Add evaluation results to A’s history add(HA, score ); success = 0; fori←1toN×Rdo r, q←Htrials[i]; success ←max(HJ(r, q), success ); return success ; 21 Algorithm 2: PAIR Input : Taskt... | https://arxiv.org/abs/2505.20162v1 |
/uni00000014/uni00000012/uni00000015/uni0000001c/uni00000014/uni00000012/uni00000014/uni0000001a /uni00000014/uni00000014/uni00000012/uni00000014/uni00000018 /uni00000014/uni00000012/uni00000014/uni00000018 /uni00000014/uni00000012/uni00000016/uni00000014/uni00000012/uni00000015/uni00000018 /uni00000014/uni00000012/uni... | https://arxiv.org/abs/2505.20162v1 |
/uni00000014/uni00000012/uni00000017/uni00000016 /uni00000014/uni00000012/uni00000015/uni0000001c /uni00000014/uni00000012/uni00000016/uni0000001c /uni00000014/uni00000012/uni00000017/uni0000001a /uni00000014/uni00000012/uni00000017/uni00000016 /uni00000014/uni00000012/uni00000017/uni00000018 /uni00000014/uni00000012/u... | https://arxiv.org/abs/2505.20162v1 |
is to predict the expected ASR, along with calibrated uncertainty estimates. To quantify the capability difference , we define the capability gap δa→t, as function of MMLU-Pro scores of attacker and target. We compare different capability gap definitions in the following sections. To model worst-case scenario, for the ... | https://arxiv.org/abs/2505.20162v1 |
implement the model using the PyMC python library with the HMC NUTS (No-U-Turn Sampler) algorithm for posterior approximation. Hyperparameters are selected separately for each model via 25 Type II Maximum Likelihood (Empirical Bayes) by maximizing the marginal log likelihood over the range [0.01,3.0]using Optuna (100 s... | https://arxiv.org/abs/2505.20162v1 |
α= 0.05, and (iv) Winkler interval score at α= 0.05, for predictive uncertanty calibration. We report each metric averaged over all per-target fits (including outliers) in Tab. A.5. Miscoverage is defined as the proportion of observed ASR values that fall outside the model’s predicted 95% confidence interval: Miscovera... | https://arxiv.org/abs/2505.20162v1 |
Exploring Generative Error Correction for Dysarthric Speech Recognition Moreno La Quatra*♦, Alkis Koudounas*♣, Valerio Mario Salerno♦, Sabato Marco Siniscalchi♠ ♦Kore University of Enna, Italy,♣Politecnico di Torino, Italy,♠Universit `a di Palermo, Italy moreno.laquatra@unikore.it, alkis.koudounas@polito.it, valerio.sa... | https://arxiv.org/abs/2505.20163v1 |
its understanding of language patterns and context to produce more accurate transcriptions. While previous approaches to dysarthric speech recognition have focused on adapting acous- tic models [19, 20], we also explore error correction through text-to-text mapping using language modeling. We evaluate this system as pa... | https://arxiv.org/abs/2505.20163v1 |
distances between all remaining hypotheses 3. Iteratively select hypotheses that maximize the minimum dis- tance to previously selected ones This approach is designed to select hypotheses that represent genuinely different interpretations rather than minor variations of the same transcription. 2.2. Generative Error Cor... | https://arxiv.org/abs/2505.20163v1 |
SemScore WER SemScore WER SemScore Lv3 ✗ ✗ 11.60 83.91 11.39 84.82 14.49 78.83 Lv3 ✗ 3B 7.91 89.42 10.99 86.45 13.87 79.88 Lv3 ✗ 11B 7.34 90.32 10.63 86.96 13.64 80.4 Lv2 ✓ ✗ 7.17 91.94 10.90 87.24 13.04 81.89 Lv2 ✓ 3B 6.40 92.47 10.67 87.51 12.89 82.16Table 2: Ablation study on N-best list size. ASR ModelNDev WER SemS... | https://arxiv.org/abs/2505.20163v1 |
decay. GER Stage. The GER stage uses two sizes of the FlanT5 model family: XL (3B parameters) and XXL (11B parameters). Given the large model sizes, we employ LoRA adaptation with rank r= 16 ,α= 32 , targeting all linear layers. This introduces ap- proximately 1% additional trainable parameters while keeping the base m... | https://arxiv.org/abs/2505.20163v1 |
development set, indicating effective acoustic model adaptation for dysarthric speech. Combining fine-tuning with GER achieves the best overall performance (6.40% WER, 92.47 SemScore on devel- opment; 12.89% WER on TEST -2), demonstrating complemen- tary benefits from acoustic and linguistic components. Our ex- perimen... | https://arxiv.org/abs/2505.20163v1 |
where our diversity-based hypothesis selection preserves the correct transcription. While standard beam search ranks hypotheses by confidence score alone, our approach explicitly selects diverse alternatives. This diversity is evident in the example, where despite errors in the top hypoth- esis ( play,set,Monday ), the... | https://arxiv.org/abs/2505.20163v1 |
Interna- tional Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW) . IEEE, 2024, pp. 505–509. [5] A. Koudounas, E. Pastor, G. Attanasio, V . Mazzia, M. Giollo, T. Gueudre, L. Cagliero, L. de Alfaro, E. Baralis, and D. Amberti, “Exploring subgroup performance in end-to-end speech models,” inICASS... | https://arxiv.org/abs/2505.20163v1 |
of speech accessibility,” Journal of Speech, Language, and Hearing Research , 2024. [19] H. Wang, Z. Jin, M. Geng, S. Hu, G. Li, T. Wang, H. Xu, and X. Liu, “Enhancing pre-trained asr system fine-tuning for dysarthric speech recognition using adversarial data augmenta- tion,” in ICASSP , 2024. [20] I.-T. Hsieh and C.-H... | https://arxiv.org/abs/2505.20163v1 |
arXiv:2505.20164v1 [cs.CL] 26 May 2025Visual Abstract Thinking Empowers Multimodal Reasoning Dairu Liu3∗, Ziyue Wang1♠∗, Minyuan Ruan1∗, Fuwen Luo1, Chi Chen1, Peng Li2 ,Yang Liu1,2 1Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China 2Institute for AI Industry Research (AIR), Tsinghua... | https://arxiv.org/abs/2505.20164v1 |
Standard CoT SketchPad VAT (Ours)Odd One Out Illusion B* Counting B* Sem. Corr. Figure 2: Time and budget costs of V AT compared to standard prompting, CoT, and the tool-using method Visual SketchPad (SketchPad). B*: BLINK.Recent advances in Multimodal Large Language Models (MLLMs) [ 29,20,35] demonstrate impressive co... | https://arxiv.org/abs/2505.20164v1 |
without step-by-step thinking. V AT thus serves as a visual counterpart to CoT, emphasizing the potential of visual abstracts over accumulations. Based on V AT, we investigate whether current MLLMs elicit similar ability to thinking in an abstract way as humans do. Specifically, we discuss: (1) whether visual abstracts... | https://arxiv.org/abs/2505.20164v1 |
learning-based sketch representation [ 13, 6,25] as a notable paradigm for visual understanding. It offers the abstraction and compression of visual context while retaining essential semantic and structural cues. Consequently, research has long focused on uncovering the implicit semantics conveyed through sketches, the... | https://arxiv.org/abs/2505.20164v1 |
transformer function T:I → V ato converts input images Iinto various forms of visual abstracts Va, preserving essential visual concepts, geometric information, silhouettes, and relational features. This transformer function Tcan take diverse forms. In this paper, we employ both code-based conversion approaches, includi... | https://arxiv.org/abs/2505.20164v1 |
175.00 - - - - 32.80 40.40 50.00 56.00 38.50 - Current proprietary and open-source models Qwen2.5-VL-32B 37.25 60.42 167.56 175.00 180.00 83.92 74.17 47.48 77.91 36.33 48.12 55.56 68.84 34.54 - + Scaffold7.84 -29.4161.81 +1.39158.37 -9.19171.53 -3.47190.00 +10.0083.22 -0.7066.67 -7.5045.32 -2.1677.91 +0.0034.80 -1.5341... | https://arxiv.org/abs/2505.20164v1 |
reasoning ability with extensive reliance on everyday knowledge, involving MME (commonsense reasoning) and BLINK (functional correspondence), where V AT can offer a condensed view for reasoning. Please refer to Appendix B for details of our created Odd-One-Out benchmark and other employed tasks, with corresponding eval... | https://arxiv.org/abs/2505.20164v1 |
OpenSketch style visual abstracts, outperforming CoT and Scaffold by +13.02% and+4.17%, respectively. On Qwen-2.5-VL and Gemini-2.0, V AT consistently surpasses both CoT and Scaffold. Notably, on Gemini-2.0, V AT achieves top performance on nine out of 13 tasks, further highlighting its effectiveness. These results col... | https://arxiv.org/abs/2505.20164v1 |
some results in Table 4. Image-only uses the original query image without any visual abstraction. Vis.Abs-only removes the query im- age and uses only the visual abstraction as input. VAT w/o Vis.Abs retains the two-image input format of V AT but replaces the visual abstract by the original input image. VAT (Best) repr... | https://arxiv.org/abs/2505.20164v1 |
from different viewpoints. Notably, for object-centric task requiring models to focus on specific objects attributes or features, visual abstract thinking larger assists in filtering out irrelevant background clutters. Though achieving improvement in spatial reasoning and commonsense reasoning compared to the GPT-4o ba... | https://arxiv.org/abs/2505.20164v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.