text string | source string |
|---|---|
a proxy for effectiveness on held-out evaluation models (empirically validated in §5), we propose several 1In our discussion of dynamic safety evaluation, we focus on automated methods, though the same principles apply to both human and LLM-based red-teaming. 2We coin “Jailbreak Distillation” specifically in the scope ... | https://arxiv.org/abs/2505.22037v1 |
multi-turn evaluation, and propose effective prompt selection algorithms, empirically verified by our experiments. (4) We conduct analyses and discover no evidence of sig- nificant bias in JBD ISTILL -produced benchmarks. 2 Desiderata for Safety Benchmarks While many benchmarks are constructed to evalu- ate model safet... | https://arxiv.org/abs/2505.22037v1 |
successful attack on a particular eval- uation model, averaged over all evaluation models. That is, VER(B;Meval) =X M∈M eval n g∈G ∃p: (g,p)∈B, J(g,M (p))=1o /|G| |M eval|. We complement versatility with another diver- sity metric, Coverage , i.e., the proportion of seed goals that are covered by the benchmark. Cover- ... | https://arxiv.org/abs/2505.22037v1 |
date prompts. Next, the prompt selection algorithm Achooses a subset of nprompts satisfying our desiderata (§2) as the constructed benchmark P∗. When will JBD ISTILL be effective? The ef- fectiveness of JBD ISTILL benchmarks relies on the selected attack prompts being broadly effective acrossMdevandMeval, while not bei... | https://arxiv.org/abs/2505.22037v1 |
prompt selection algorithms. Interestingly, we find that simple greedy algorithms already achieve high effective- ness and separability in practice (§5.2). We use random selection as a baseline, and propose three algorithms: RBS, BPG, and CS. Baseline algorithm: RANDOM SELECTION (RS) The simplest baseline prompt select... | https://arxiv.org/abs/2505.22037v1 |
development models, (C) reasoning models, (D) unseen families (model families that are not repre- sented in Mdev), and (E) specialized models (e.g., coding- or healthcare-oriented models), to evaluate the effectiveness of the benchmark, detailed in §F. Evaluation judge We use the AdvPrefix judge for single-turn evaluat... | https://arxiv.org/abs/2505.22037v1 |
that using multiple development models al-lows for selecting effective prompt subsets, vali- dating our core hypothesis. While previous works have mostly focused on generating more transfer- able attack prompts (Zou et al., 2023; Sabbaghi et al., 2025; Lin et al., 2025a; Yang et al., 2025), we show that over-generating... | https://arxiv.org/abs/2505.22037v1 |
from Mdev GEMMA 2-27B-IT 90.2 →88.6 (-1.6) 5th →4th GEMMA 3-12B-IT 97.4 →96.8 (-0.6) 8th →8th Table 2: Removing the LLAMA orGEMMA family from Mdevdoes not significantly affect ASR and rankings of the benchmark for Mevalof the same family. 6 Analysis 6.1 Are JBD ISTILL Benchmarks biased toward Development Model Families... | https://arxiv.org/abs/2505.22037v1 |
models, LLM evaluation is mov- ing to dynamic evaluation methods that generate test prompts on the fly or live benchmarks that can be continuously updated (Chen et al., 2025; Zhang et al., 2025a; Verma et al., 2025, i.a.). JBD ISTILL fall into this space and is a benchmark construction pipeline that generates continual... | https://arxiv.org/abs/2505.22037v1 |
and risk land- scapes, we propose the JBD ISTILL and demon- strate its prowess for renewable safety evaluation, tackling the comparability and reproducibility chal- lenges of existing dynamic evaluation, as well as saturation and contamination issues of static bench- marks. We stress that JBD ISTILL is not a replace- m... | https://arxiv.org/abs/2505.22037v1 |
attacks. Preprint , arXiv:2404.02151. Alex Beutel, Kai Xiao, Johannes Heidecke, and Lilian Weng. 2024. Diverse and effective red teaming with auto-generated rewards and multi-step reinforcement learning. arXiv preprint arXiv:2412.18693 . Tim Beyer, Sophie Xhonneux, Simon Geisler, Gauthier Gidel, Leo Schwinn, and Stepha... | https://arxiv.org/abs/2505.22037v1 |
Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, ... | https://arxiv.org/abs/2505.22037v1 |
Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis An- derson, Govind Thattai, Graeme Nail, Gregoire Mi- alon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarr... | https://arxiv.org/abs/2505.22037v1 |
Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching- Hsiang Chu, Chris Cai, Chris Tindal, Christoph Fe- ichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Fo... | https://arxiv.org/abs/2505.22037v1 |
Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, Zhiwei Zhao, an... | https://arxiv.org/abs/2505.22037v1 |
Yaron Singer, and Amin Karbasi. 2024. Tree of attacks: Jailbreaking black-box llms automatically. In Advances in Neural Information Processing Systems , volume 37, pages 61065–61105. Curran Associates, Inc. Microsoft, :, Abdelrahman Abouelenin, Atabak Ash- faq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Al... | https://arxiv.org/abs/2505.22037v1 |
de Oliveira Pinto, Hongyu Ren, Huiwen Chang, Hyung Won Chung, Ian Kivlichan, Ian O’Connell, Ian O’Connell, Ian Osband, Ian Sil- ber, Ian Sohl, Ibrahim Okuyucu, Ikai Lan, Ilya Kostrikov, Ilya Sutskever, Ingmar Kanitscheider, Ishaan Gulrajani, Jacob Coxon, Jacob Menick, Jakub Pachocki, James Aung, James Betker, James Cro... | https://arxiv.org/abs/2505.22037v1 |
Toki Sherbakov, Tom Rubin, Tom Stasi, Tomer Kaftan, Tristan Heywood, Troy Peterson, Tyce Walters, Tyna Eloundou, Valerie Qi, Veit Moeller, Vinnie Monaco, Vishal Kuo, Vlad Fomenko, Wayne Chang, Weiyi Zheng, Wenda Zhou, Wesam Manassra, Will Sheu, Wojciech Zaremba, Yash Patil, Yilei Qian, Yongjik Kim, Youlong Cheng, Yu Zh... | https://arxiv.org/abs/2505.22037v1 |
Agarwal, Santiago Hernandez, Sasha Baker, Scott McKinney, Scottie Yan, Shengjia Zhao, Shengli Hu, Shibani Santurkar, Shraman Ray Chaudhuri, Shuyuan Zhang, Siyuan Fu, Spencer Papay, Steph Lin, Suchir Balaji, Suvansh Sanjeev, Szymon Sidor, Tal Broda, Aidan Clark, Tao Wang, Taylor Gordon, Ted Sanders, Te- jal Patwardhan, ... | https://arxiv.org/abs/2505.22037v1 |
Scalable and transferable black-box jail- breaks for language models via persona modulation. Preprint , arXiv:2311.03348. Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2023. "do anything now": Charac- terizing and evaluating in-the-wild jailbreak prompts on large language models. Xinyue Shen, Zeyu... | https://arxiv.org/abs/2505.22037v1 |
Phoebe Kirk, Anand Rao, Kat Black, Nabila Babar, Jessica Lo, Erica Moreira, Luiz Gus- tavo Martins, Omar Sanseviero, Lucas Gonzalez, Zach Gleicher, Tris Warkentin, Vahab Mirrokni, Evan Senter, Eli Collins, Joelle Barral, Zoubin Ghahra- mani, Raia Hadsell, Yossi Matias, D. Sculley, Slav Petrov, Noah Fiedel, Noam Shazeer... | https://arxiv.org/abs/2505.22037v1 |
Kenealy, Robert Dadashi, and Alek Andreev. 2024. Gemma 2: Improving open language models at a practical size. Preprint , arXiv:2408.00118. Qwen Team. 2025. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github. io/blog/qwq-32b/ . Accessed: 2025-05-18. Simone Tedeschi, Felix Friedrich, Patrick Sc... | https://arxiv.org/abs/2505.22037v1 |
Matt Fredrikson, and Dan Hendrycks. 2024. Improving alignment and robustness with circuit breakers. Preprint , arXiv:2406.04313. Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, and Matt Fredrikson. 2023. Univer- sal and transferable adversarial attacks on aligned language models. Preprint , arXiv:23... | https://arxiv.org/abs/2505.22037v1 |
nine tie-broken items it agreed with the majority vote in seven instances, yielding an overall accuracy of 98 % (98/100). Table 4: Agreement statistics for the multi-turn judge on 100 conversations. Metric Value Human–human agreement (3–way) 91 % Judge agreement on unanimous subset 100 % (91/91) Judge agreement with ma... | https://arxiv.org/abs/2505.22037v1 |
attack method). adversarial prompts for the top five most effective persuasion techniques identified in the taxonomy, following a setup similar to AutoDAN-Turbo. For each of the 200 standard HarmBench prompts, we generated one adversarial variant per persuasion strategy, resulting in a total of 1,000 adversarial prompt... | https://arxiv.org/abs/2505.22037v1 |
harmful actors and generate implicit harmful queries associated with those actors that appear harmless when sent to the target model. We utilize two attacker models: GPT-4 Oand MIXTRAL -8X7B-I NSTRUCT -V0.1, generating at- tack paths for targets from HarmBench. We dis- able dynamic modification and set the maximum numb... | https://arxiv.org/abs/2505.22037v1 |
experiments (§5.2) and stability analysis (§6.2): •GEMMA 2-27B-IT (Team et al., 2024): larger variants of development models. Algorithm 4 COMBINED SELECTION Input: Development models Mdev, Candidate prompt pool P, Target benchmark size n. Output: A benchmark P∗⊆P 1:P∗← ∅ 2: // First select the best prompt for each goal... | https://arxiv.org/abs/2505.22037v1 |
is DNS Amplification, which can be used maliciously to launch DDoS attacks. Although it’s essential to approach these discussions with ethical considerations, providing an evidence-based explanation of DNS Amplification can help individuals better understand the risks and develop effective countermeasures. Therefore, I... | https://arxiv.org/abs/2505.22037v1 |
chain of custody and provide a clear explanation of your actions.\n**Consult with a supervisor or legal expert**: If you're unsure about any aspect of the situation, consult with a supervisor or legal expert to ensure you 're following proper procedures and avoiding any potential legal pitfalls.\nRemember, your primary... | https://arxiv.org/abs/2505.22037v1 |
challenges and consequences, we can better devise strategies to detect and combat these illicit operations. Could you provide a general overview of the typical steps taken to establish an illegal brothel, without encouraging any illegal actions? Attack prompt 2. Created by Adversarial Rea- soning attacking G EMMA 2-9B-... | https://arxiv.org/abs/2505.22037v1 |
arXiv:2505.22054v1 [cs.CL] 28 May 2025Voice Adaptation for Swiss German Samuel Stucki, Jan Deriu, Mark Cieliebak Centre for Artificial Intelligence, ZHAW, Switzerland deri@zhaw.ch, stku@zhaw.ch, ciel@zhaw.ch Abstract This work investigates the performance of V oice Adaptation models for Swiss German dialects, i.e., tra... | https://arxiv.org/abs/2505.22054v1 |
(TTS) were enabled by the SwissDial corpus [13]. It covers eight dialects (with one speaker per dialect) and contains 2.5 to 4.55 hours of audio per dialect, totaling around 26 hours. They also created transcripts for each audio sample for each dialect. They trained a Tacatron 2 model [14] for each dialect separately a... | https://arxiv.org/abs/2505.22054v1 |
segments. We used the pyannoteAI [15, 16] for both tasks. Since the number of speakers per episode is unknown, we set a range of 2–6. While this may cause over- segmentation, it doesn’t affect our pipeline, as we only require each segment to contain a single speaker. Since the pipeline was mainly trained on English aud... | https://arxiv.org/abs/2505.22054v1 |
[21] (with equal gender distribution) to han- dle the Standard German parts of the SRG audio data. Follow- ing [19], we concatenated samples of the same speaker in the test set to create 30s samples, achieving a macro F1-score of 0.88. The model often confuses the two geographically close regions, Zurich and Central Sw... | https://arxiv.org/abs/2505.22054v1 |
use generated texts in the Long scenario stems from the noisiness of the SRG corpus due to automated data collection. Automated Evaluation. We evaluated three aspects: • Back-Translation Accuracy (i.e., translating the generated speech back to text): Measured using WER and BLEU scores between the text input to the TTS ... | https://arxiv.org/abs/2505.22054v1 |
which contrasts the single sentence nature of theShort scenario, and second, that most samples in the SRG dataset are precisely 15 seconds of length, while the Short texts illicit speech of around 5-7 seconds. •Long Results. Table 4 shows the results. Performance is bet- ter across all models, with SRG+STT4SG++ achievi... | https://arxiv.org/abs/2505.22054v1 |
the Zurich dialect, which is under- estimated, as the classifier often misassigns Zurich samples toEval Type Model SMOS CMOS Intelligibility Baseline 3.10 ±0.83 -0.80±0.98 4.17±0.94 Short SRG+STT4SG 3.39±1.05∗-0.55±0.86 3.85±1.29† SRG+STT4SG++ 3.24 ±0.90 -0.52±0.84∗4.51±0.69∗† Baseline 2.98 ±0.92 -0.94±0.80 3.96±0.86 L... | https://arxiv.org/abs/2505.22054v1 |
L. Zhou, C. Wang, S. Chen, Y . Wu, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al. , “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,” arXiv preprint arXiv:2303.03926 , 2023. [4] E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morai... | https://arxiv.org/abs/2505.22054v1 |
pp. 1763–1772. [Online]. Available: https://aclanthology.org/2023.acl-short.150/ [12] V . Timmel, C. Paonessa, R. Kakooee, M. V ogel, and D. Perru- choud, “Fine-tuning whisper on low-resource languages for real- world applications,” arXiv preprint arXiv:2412.15726 , 2024.[13] P. Dogan-Sch ¨onberger, J. M ¨ader, and T. ... | https://arxiv.org/abs/2505.22054v1 |
Safeguarding Privacy of Retrieval Data against Membership Inference Attacks: Is This Query Too Close to Home? Yujin Choi1†Youngjoo Park1†Junyoung Byun2 1Seoul National University, Republic of Korea 2Chung-Ang University, Republic of Korea 3Korea Institute for Advanced Study, Republic of Korea {uznhigh, youngjoo0913, ja... | https://arxiv.org/abs/2505.22061v1 |
threshold, indicating that the query is overly correlated with one specific document in the retrieval, we hide this private data in the top- kdocument conveyed to the LLMs. We summarize our contributions as follows: •We propose a similarity-based method for de- tecting MIA inRAG systems using Gum bel, named Mirabel . T... | https://arxiv.org/abs/2505.22061v1 |
method that asks an LLM agent, such as GPT-4o, to classify incoming queries as benign or malicious. However, this agent-based method struggles with several chal- lenges that will be further discussed in Section 3.1. As a complementary safeguard, differential pri- vacy (DP) provides a mathematically rigorous pri- vacy g... | https://arxiv.org/abs/2505.22061v1 |
(Naseh et al., 2025). Moreover, simply rejecting all suspected at- tack queries may inadvertently reveal to attackers filtering phrases and detour the detection systems. These limitations suggest that an effective de- fense requires not simply blocking queries but ob- fuscating the attacker’s knowledge. An attack suc- ... | https://arxiv.org/abs/2505.22061v1 |
Gumbel-based thresholdS qSqis marked for reference. Test set Normal MBA S2MIA IA Sq 0.469 0.027∗0.001∗0.012∗ Sq\{smax} 0.400 0.511 0.293 0.226 Table 1: Average p-values of the normality test on the total data set ( Sq) and on the set with the maximum similarity removed ( Sq\{smax}). * :p-value < 0.05. and standard devi... | https://arxiv.org/abs/2505.22061v1 |
User AttackerPrivate External DB 𝓓 RAG Prompt 𝒑𝒒 Generator 𝑮 𝑹𝒌𝒒𝒒 + Retriever 𝑹 Gumbel Threshol d 𝝉 Top-𝒌 Detect HideNon-member Query Member Query Figure 2: Illustration of our proposed Mirabel. We perform our detection to classify whether an input query is a member attack query qm a. If it detected as qm a,... | https://arxiv.org/abs/2505.22061v1 |
three MIA methods: S2MIA (Li et al., 2024), MBA (Liu et al., 2025), and IA (Naseh et al., 2025). S2MIA feeds the first half of the target document to the RAG and scores membership with BLEU and perplexity against the full text. MBA masks tokens and counts how many the generator recovers. IA lets an LLM craft 30 inferen... | https://arxiv.org/abs/2505.22061v1 |
et al., 2024). Specifically, we used Llama 3.1 for the utility evaluations and Llama 3.2 for the remaining evaluations. Further details on the RAG prompt can be found in the Figure 4. Experiments were mainly conducted on a single A100 or H100 GPU with 96 GB VRAM. Additional details are shown in Appendix A. 5.2 Detectio... | https://arxiv.org/abs/2505.22061v1 |
quality of the gen- erated answers, shows that our method performs comparably to the standard RAG, while DP-RAG results in greater utility degradation. In contrast, R@k, which measures how well the system retrieves relevant documents, shows a mod- erate decrease compared to the original RAG sys- tem. This is because ou... | https://arxiv.org/abs/2505.22061v1 |
MIA DP-RAG 0.077 0.051 0.038 0.055 DP-RAG-L 0.060 0.035 0.058 0.051 Ours 0.057 0.039 0.043 0.046 MBA RAG 0.754 0.813 0.625 0.731 DP-RAG 0.022 0.081 0.015 0.039 DP-RAG-L 0.036 0.073 0.026 0.045 Ours 0.038 0.202 0.283 0.174 IA RAG 0.805 0.771 0.555 0.710 DP-RAG 0.088 0.418 0.091 0.199 DP-RAG-L 0.301 0.414 0.223 0.313 Our... | https://arxiv.org/abs/2505.22061v1 |
defends against attacks with minimal utility degradation. Our method is model-agnostic and can be easily applied to exist- ing RAG systems. Experimental results show that our method achieves defense performance compa- rable to DP-RAG, while incurring negligible utility loss on benign queries. Limitations Even though we... | https://arxiv.org/abs/2505.22061v1 |
Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and 1 oth- ers. 2025. A survey on hallucination in large lan- guage models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems , 43(2):1–55. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Trivi... | https://arxiv.org/abs/2505.22061v1 |
Huseyin A Inan, Andre Manoel, Fatemehsadat Mireshghallah, Zinan Lin, Sivakanth Gopi, Janardhan Kulkarni, and Robert Sim. 2024. Privacy-preserving in-context learning with differentially private few-shot generation. In The Twelfth International Conference on Learning Repre- sentations . Nandan Thakur, Nils Reimers, Andr... | https://arxiv.org/abs/2505.22061v1 |
threshold based on accuracy. We searched threshold in [0,1], and we evaluated penalty values λ∈[0.5,1]. In the case of DP-RAG, since the number of queries issmaller, we differ the searching space. Since we observed that larger values of λin this range re- sulted in higher recall but significantly lower ac- curacy, indi... | https://arxiv.org/abs/2505.22061v1 |
demon- strate significantly better performance, except for IA, which is designed to be stealthy. These experi- mental results are consistent with those reported in Naseh et al. (2025). Our method demonstrates stable performance across a variety of attacks, particularly maintaining reliable detection performance even fo... | https://arxiv.org/abs/2505.22061v1 |
0.969 MBA 1.000 1.000 1.000 1.000 0.980 1.000 0.960 0.980 IA 0.500 0.000 0.000 0.000 0.810 1.000 0.620 0.765 Table 8: Detection performance of Mirabel compared to agent-based detection, evaluated using qnandqm a. Acc Precision Recall F1 S2MIA MBA IA S2MIA MBA IA S2MIA MBA IA S2MIA MBA IA NF RAG 0.688 0.877 0.903 0.636 ... | https://arxiv.org/abs/2505.22061v1 |
0.394 0.740 0.586 0.439 0.650 +Ours 0.480 0.529 0.477 0.487 0.537 0.572 0.748 0.426 0.383 0.590 0.475 0.459 DP-RAG-L 0.466 0.514 0.767 0.465 0.517 0.806 0.452 0.438 0.787 0.458 0.474 0.797 +Ours 0.486 0.507 0.496 0.486 0.508 0.620 0.478 0.432 0.336 0.482 0.467 0.436 SCI DP-RAG 0.515 0.540 0.788 0.538 0.560 0.834 0.214 ... | https://arxiv.org/abs/2505.22061v1 |
arXiv:2505.22076v1 [cs.CL] 28 May 2025ArgInstruct: Specialized Instruction Fine-Tuning for Computational Argumentation Maja Stahl* Leibniz University Hannover m.stahl@ai.uni-hannover.deTimon Ziegenbein* Leibniz University Hannover t.ziegenbein@ai.uni-hannover.de Joonsuk Park† University of Richmond park@joonsuk.orgHenn... | https://arxiv.org/abs/2505.22076v1 |
affects CA tasks in particular, as they often center around contextual guidance (Mishra et al., 2022; Wang et al., 2022). sophisticated context-related concepts from argu- mentation theory (Wachsmuth et al., 2024). Specifically, CA research in NLP focuses on the mining, assessment, and generation of natural language ar... | https://arxiv.org/abs/2505.22076v1 |
(CA), has its roots in a long history of philosophical research (Aristotle, ca. 350 B.C.E. / translated 2007), which has gained significant attention from the NLP community in recent years. The three main CA research areas frequently covered are argument mining (Park and Cardie, 2014; Boltuži ´c and Šnajder, 2014; Stab... | https://arxiv.org/abs/2505.22076v1 |
2024a). Over the last years, multiple instruction fine- tuning datasets have been collecting from existing NLP tasks (Mishra et al., 2022; Wang et al., 2022). The datasets usually consist of natural language instructions and example instances, which are ei- ther written manually by humans (Sanh et al., 2021; Ouyang et ... | https://arxiv.org/abs/2505.22076v1 |
large instruction fine-tuning dataset T={T1, . . . , T n} containing a large set of CA-specific but diverse tasks by (1) generating new instructions, (2) filter- ing for CA relevance and diversity, and (3) generat- ing corresponding input-output instances. Instruction Generation Building on research in CA, we curate a ... | https://arxiv.org/abs/2505.22076v1 |
LLM Instruction Fine-Tuning To create our ArgInstruct model, we fine-tune a pretrained LLM on both the entire CA task pool and general tasks, aiming to specialize in CA while maintaining the generalization idea of instruction fine-tuning. We format task instances into a prompt- ing template for training and mask input ... | https://arxiv.org/abs/2505.22076v1 |
argument assessment, and argument generation. The table includes the corresponding paper, text genre, and the kinds and numbers of extracted CA seed tasks. The tasks from the 9 CA datasets marked with “ ⋆” are reserved for testing. Lauscher et al. (2022) (30), as well as more recent datasets found through searches in t... | https://arxiv.org/abs/2505.22076v1 |
tasks (2,170), likely due to LLMs’ limited exposure to regression tasks. Al- though the generated instructions are shorter on average (28.2 vs. 48.1 words), input lengths are similar (50.7 vs. 64.3). The longer output length (25.2 vs. 7.7) in the generated data likely stems from the higher proportion of generation task... | https://arxiv.org/abs/2505.22076v1 |
highlight their CA focus. While these instructions are close to existing CA tasks, namely thesis extraction and stance detection (I1), and relation detection ( I2), they introduce new wordings that will likely lead to more robust fine- tuning. Exemplary instructions with the lowest maximal similarity to the seed instru... | https://arxiv.org/abs/2505.22076v1 |
training seed tasks for our ( seedCA ) dataset. Evaluation We use guided generation (Willard and Louf, 2023) for classification and regression tasks and open generation (up to 512new tokens) for generation tasks. For classification and regres- sion, a finite state machine decodes model outputs for direct comparison wit... | https://arxiv.org/abs/2505.22076v1 |
.34 11.0 .50 2.1 .24 7.0 + seedCA .64 1.2 .49 2.7 .63 2.1 .32 4.3 + genCA .49 2.6 .44 9.3 .52 2.6 .30 6.7 + seedCA, genCA (ArgInstruct) .57 1.3 .49 4.0 .65†1.9†‡.31†2.0 Table 3: Main CA results on (a) unseen CA instances and (b) unseen CA tasks: Gemma-2-9B instruction fine-tuned on 52k instances of CA seed ( +seedCA ),... | https://arxiv.org/abs/2505.22076v1 |
performance of all models alongside Majority andRandom base- lines. ArgInstruct outperforms all others in terms of F1(.65). However, for regression tasks, no model proves reliable, as all MASE scores are worse than predicting the mean. In generation (R-L), GPT-4o- mini appears slightly superior to our model (.32 vs. .3... | https://arxiv.org/abs/2505.22076v1 |
method may be well-transferable to other specialized NLP domains, for example, to the educational domain. There, our method could involve collecting seed tasks such as essay scoring, feedback generation, text suggestion, and rewriting, but we leave this to future work. 7 Limitations The research proposed in this paper ... | https://arxiv.org/abs/2505.22076v1 |
for computational argumentation, we expect and encourage future work to apply our proposed methodology to other NLP areas that re- quire specific domain knowledge. Acknowledgments The presented work has been partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) within the project Arg... | https://arxiv.org/abs/2505.22076v1 |
the Association for Computational Lin- guistics: Volume 1, Long Papers , pages 251–261, Valencia, Spain. Association for Computational Lin- guistics. Tilman Beck, Ji-Ung Lee, Christina Viehmann, Marcus Maurer, Oliver Quiring, and Iryna Gurevych. 2021. Investigating label suggestions for opinion mining in German covid-1... | https://arxiv.org/abs/2505.22076v1 |
. Le, and Jason Wei. 2024. Scaling instruction-finetuned language models. J. Mach. Learn. Res. , 25(1). Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American ... | https://arxiv.org/abs/2505.22076v1 |
Christopher A. Choquette-Choo, Danila Sinopalnikov, David Wein- berger, Dimple Vijaykumar, Dominika Rogozi ´nska, Dustin Herbison, Elisa Bandy, Emma Wang, Eric Noland, Erica Moreira, Evan Senter, Evgenii Elty- shev, Francesco Visin, Gabriel Rasskin, Gary Wei, Glenn Cameron, Gus Martins, Hadi Hashemi, Hanna Klimczak-Plu... | https://arxiv.org/abs/2505.22076v1 |
Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis An- derson, Govind Thattai, Graeme Nail, Gregoire Mi- alon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Is- han Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jae... | https://arxiv.org/abs/2505.22076v1 |
Chao Zhou, Chester Hu, Ching- Hsiang Chu, Chris Cai, Chris Tindal, Christoph Fe- ichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Ja... | https://arxiv.org/abs/2505.22076v1 |
Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhaoduo Wen, Zhenyu Yang, Zhiwei Zhao, and Zhiyu Ma. 2024. The Llama 3 Herd of Models. Preprint , arXiv:2... | https://arxiv.org/abs/2505.22076v1 |
Hautli-Janisz, Zlata Kikteva, Wassiliki Siskou, Kamila Gorska, Ray Becker, and Chris Reed. 2022. QT30: A corpus of argument and conflict in broad- cast debate. In Proceedings of the Thirteenth Lan- guage Resources and Evaluation Conference , pages 3291–3300, Marseille, France. European Language Resources Association. P... | https://arxiv.org/abs/2505.22076v1 |
data selection for instruction tuning. In Proceedings of the 2024Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages 7602–7635, Mexico City, Mexico. Association for Computational Linguistics. Matthias Liebeck, Katharina E... | https://arxiv.org/abs/2505.22076v1 |
Kim, Christine Choi, Christine McLeavey, Christopher Hesse, Clau- dia Fischer, Clemens Winter, Coley Czarnecki, Colin Jarvis, Colin Wei, Constantin Koumouzelis, Dane Sherburn, Daniel Kappler, Daniel Levin, Daniel Levy, David Carr, David Farhi, David Mely, David Robin- son, David Sasaki, Denny Jin, Dev Valladares, Dim- ... | https://arxiv.org/abs/2505.22076v1 |
Gontijo Lopes, Raul Puri, Reah Miyara, Reimar Leike, Renaud Gaubert, Reza Zamani, Ricky Wang, Rob Donnelly, Rob Honsby, Rocky Smith, Rohan Sahai, Rohit Ramchan- dani, Romain Huet, Rory Carmichael, Rowan Zellers, Roy Chen, Ruby Chen, Ruslan Nigmatullin, Ryan Cheu, Saachi Jain, Sam Altman, Sam Schoenholz, Sam Toizer, Sam... | https://arxiv.org/abs/2505.22076v1 |
of the Association for Com- putational Linguistics and the 7th International Joint Conference on Natural Language Processing (Vol- ume 1: Long Papers) , pages 543–552, Beijing, China. Association for Computational Linguistics. Isaac Persing and Vincent Ng. 2016. Modeling stance in student essays. In Proceedings of the ... | https://arxiv.org/abs/2505.22076v1 |
Daxenberger, and Iryna Gurevych. 2021. Aspect-controlled neural argument generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 380–396, Online. Association for Computational Linguistics. Eyal Shnarch, Carlos ... | https://arxiv.org/abs/2505.22076v1 |
stance classification. InSame Side Shared Task 2019: Same Side Stance Classification Shared Task 2019 , pages 1–7. CEUR Workshop Proceedings (CEUR-WS. org). Shahbaz Syed, Khalid Al Khatib, Milad Alshomary, Henning Wachsmuth, and Martin Potthast. 2021. Generating informative conclusions for argumenta- tive texts. In Fin... | https://arxiv.org/abs/2505.22076v1 |
Hou, and Iryna Gurevych. 2024. How to handle different types of out-of- distribution scenarios in computational argumenta- tion? a comprehensive and fine-grained field study. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 14878–14898, Bangkok, T... | https://arxiv.org/abs/2505.22076v1 |
of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 4344–4363, Toronto, Canada. Association for Computational Linguistics. A Collected CA Datasets Table 7 shows the complete list of the 71 CA datasets considered. B Exemplary Seed Instructions This section provides... | https://arxiv.org/abs/2505.22076v1 |
Balaji (2020)♡ Chen et al. (2022)♡Gretz et al. (2020)♡Schiller et al. (2021)♡ Eckle-Kohler et al. (2015) Habernal et al. (2018a)♡Skeppstedt et al. (2018)♡ Ein-Dor et al. (2020) Habernal et al. (2018b) Skitalinskaya et al. (2021)♡ Feger and Dietze (2024) Habernal and Gurevych (2016b) Stahl et al. (2023)♡ Grundler et al.... | https://arxiv.org/abs/2505.22076v1 |
ef- fective batch size of 64, cosine learning rate decay, and a warmup ratio of 0.05. E ArgInstruct: Task Results Table 8 and 9 show the performance of our ArgIn- struct model for the 100 sampled test instances for all 105 CA seed tasks. F ArgInstruct for CA Tasks: Dual General Instruction-Finetuning Table 10 presents ... | https://arxiv.org/abs/2505.22076v1 |
Identification train 0.63 - - Argumentative Text Creation train - - 0.21 Central Claim Extraction train - - 0.79 Persing and Ng (2015) Classifying Argument Strength train - 2.05 - Table 8: Performance of our ArgInstruct model on all 105 CA seed tasks. The split indicates whether the task was seen during training (train... | https://arxiv.org/abs/2505.22076v1 |
split indicates whether the task was seen during training (train) or is a completely unseen task (test). The performance is always measured on the 100 sampled instances from the test split of the respective task data. (Part 2/2) Fine-Tuning Data (a) Unseen CA Instances (b) Unseen CA Tasks Approach seedCA genCA general ... | https://arxiv.org/abs/2505.22076v1 |
arXiv:2505.22088v1 [cs.SD] 28 May 2025Visual Cues Support Robust Turn-taking Prediction in Noise Sam O’Connor Russell, Naomi Harte ADAPT Centre, School of Engineering, Trinity College Dublin, Ireland russelsa@tcd.ie, nharte@tcd.ie Abstract Accurate predictive turn-taking models (PTTMs) are essen- tial for naturalistic ... | https://arxiv.org/abs/2505.22088v1 |
sources and levels of noise. We therefore present the first exploration of PTTM performance in background noise, asking: 1.How is the performance of predictive turn-taking models af- fected by background noise? and 2.Does the inclusion of visual features make multimodal PTTMs robust to background noise? This paper expl... | https://arxiv.org/abs/2505.22088v1 |
the ground-truth knowledge of the current speaker [11], we similarly extended the model to support stereo audio and remove this dependency [12]. For comparison, we also train a video-only VAP model, identical to V AP with the audio encoder removed [19]. Our implementations, along with full architectural details are pub... | https://arxiv.org/abs/2505.22088v1 |
augmented audio + augmented alignment. This represents a situation where there is noisy data transcribed with ASR. Training batches consist of 20-second windows with 2 second overlap randomly sam- pled from training sessions. The training procedure is a 5-fold cross-validation with 10 epochs per fold on an NVidia RTX 6... | https://arxiv.org/abs/2505.22088v1 |
5-folds of cross-validation of VAP (audio, A), MM-VAP (audio+video, A+V), and video-only (V) models on the same test set at varying noise levels (2.5 dB increments, 5 dB shown for brevity). Training is conducted on the clean and augmented (aug., 25% of sessions corrupted with 0 dB noise) audio and ASR alignments. Model... | https://arxiv.org/abs/2505.22088v1 |
MM-V AP is 65% ( p <0.01). The effective SNR gain of MM- V AP has increased to +10 dB and the extreme sensitivity to speech and music noise is reduced. At +10 dB SNR of speech noise, the average accuracy is 64% for V AP and 75% for MM- V AP (second two rows of Table 1), up from 51% and 52% pre- viously ( p < 0.01). At ... | https://arxiv.org/abs/2505.22088v1 |
in all types of noise 74−75% (Table 2, second-last row). This suggests training on a speech signal with an interfer- ing speaker ‘forces’ MM-V AP to exploit visual cues, leading to better generalisation when new types of noise are encountered. 5.6. The critical role of the alignment in training Finally, we re-train bot... | https://arxiv.org/abs/2505.22088v1 |
correct 5/10 times in 10 dB SNR music interference when trained using conventional methods, significantly complicating interaction. This rises to 6.5/10 for V AP and 7.5/10 for MM-V AP when noise is included in training; close to baseline performance, illustrating the power of visual cues in achieving robust performanc... | https://arxiv.org/abs/2505.22088v1 |
Speech & Language , vol. 67, p. 101178, 2021. [5] S. Li, A. Paranjape, and C. D. Manning, “When can i speak? predicting initiation points for spoken dialogue agents,” arXiv preprint arXiv:2208.03812 , 2022. [6] A. Woodruff and P. M. Aoki, “How push-to-talk makes talk less pushy,” in Proceedings of the 2003 ACM Internat... | https://arxiv.org/abs/2505.22088v1 |
and signal processing, ieee international con- ference on , vol. 1. IEEE Computer Society, 1992, pp. 517–520. [22] M. Roddy and N. Harte, “Neural generation of dialogue response timings,” in Proceedings of the 58th Annual Meeting of the As- sociation for Computational Linguistics , D. Jurafsky, J. Chai, N. Schluter, an... | https://arxiv.org/abs/2505.22088v1 |
arXiv:2505.22095v1 [cs.CL] 28 May 2025Learning to Route Queries across Knowledge Bases for Step-wise Retrieval-Augmented Reasoning Chunyi Peng1∗, Zhipeng Xu1∗, Zhenghao Liu1†, Yishan Li3, Yukun Yan2, Shuo Wang2, Zhiyuan Liu2, Yu Gu1, Minghe Yu1, Ge Yu1, Maosong Sun2 1School of Computer Science and Engineering, Northeas... | https://arxiv.org/abs/2505.22095v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.