text string | source string |
|---|---|
driven discourse ranging from armed conflict (e.g., the Tigray War) to social justice movements (e.g., abortion rights, migration crises). Unlike prior work, POLAR supports three com- plementary tasks: 1.Binary Polarization Detection: Is a message polarized or not? 2.Polarization Type Classification: What so- cial dime... | https://arxiv.org/abs/2505.20624v1 |
and multievent online polarization, en- abling robust cross-lingual and context-aware mod- eling. 3 POLAR Dataset Construction 3.1 Data Collection We collected data from various online platforms, including X (formerly Twitter), Facebook, Reddit, Bluesky, Threads, and news/commentary forums. Language Source(s) Train Dev... | https://arxiv.org/abs/2505.20624v1 |
via sur- veys and underwent pilot testing. Final anno- tation used a locally hosted POTATO interface adapted for multilabel inputs. Inter-annotator agreement was moderate, with notable consis- tency on polarized cases. •Hausa: We used trained annotators and Label Studio with iterative feedback and internal consistency ... | https://arxiv.org/abs/2505.20624v1 |
shapes type distribution. Spanish and Urdu show diverse forms of identity-based polarization, while political conflict dominates Amharic, Urdu, and German.Polarization Manifestations: Figure 4 show types of Manifestation. Stereotyping and vilification are most frequent in Urdu, Amharic, German, and Spanish. Urdu’s vili... | https://arxiv.org/abs/2505.20624v1 |
Setup Results To benchmark the performance of state-of-the- art multilingual language models (MLMs) for po- larization detection, we fine-tuned six prominent pretrained models: InfoXLM (Chi et al., 2021), LaBSE (Feng et al., 2022), RemBERT (Chung et al., 2021), XLM-R (Conneau et al., 2020), mBERT (Devlin et al., 2019),... | https://arxiv.org/abs/2505.20624v1 |
0.00 43.58 Arabic 40.05 51.55 52.52 55.61 42.18 47.73 16.56 30.68 22.14 0.00 19.57 17.24 English 14.40 15.01 19.39 18.61 18.60 15.15 7.16 10.16 10.05 5.62 10.69 8.85 German 38.49 49.88 52.74 51.70 46.85 51.91 2.38 36.12 27.05 0.00 23.38 12.93 Hausa 19.23 20.04 19.18 18.89 18.74 18.93 5.74 5.86 3.77 3.12 6.19 6.43 Spani... | https://arxiv.org/abs/2505.20624v1 |
GPT 4o GPT 4o-mini Mistral-7B Llama-3.1-8B Qwen3-8B Zero-shot Few-shot Zero-shot Few-shot Zero-shot Zero-shot Zero-shot Amharic 71.30 68.74 55.40 62.59 71.92 36.32 58.09 Arabic 71.51 79.65 58.21 75.95 33.65 34.82 54.54 English 75.46 79.56 64.78 80.93 50.34 56.75 72.07 German 72.04 71.12 63.03 72.44 52.63 56.22 61.67 Ha... | https://arxiv.org/abs/2505.20624v1 |
we provided detailed instructions and support resources to reduce stress and clarify expectations, but some emotional burden may have remained. Third, our choice of models is not exhaustive. Although we included several leading multilingual models and both open and closed LLMs. Adding more language-specific models in t... | https://arxiv.org/abs/2505.20624v1 |
sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers) , pages 878–891, Dublin, Ireland. Kiran Garimella. 2018. Polarization on Social Media . Ph.D. thesis, Aalto University, Finland. Todd Gitlin. 2016. The Outrage Industry: Political Opi... | https://arxiv.org/abs/2505.20624v1 |
Long Context Scaling: Divide and Conquer via Multi-Agent Question-driven Collaboration Sibo Xiao♣, Zixin Lin♣, Wenyang Gao♣,♢, Yue Zhang♢ ♣Zhejiang University,♢Westlake University {sibodotxiao}@gmail.com Abstract Processing long contexts has become a critical capability for modern large language models (LLMs). Existing... | https://arxiv.org/abs/2505.20625v1 |
Decider D1 decides the action of the next step is to Replay or Conclude. If Replay is proposed, E[i,2]will be called in a reverse direction with a start point according to the question state tracker. the multi-agent system [ 9]. Furthermore, the most concerning aspect of agent-based methods is the frequent model invoca... | https://arxiv.org/abs/2505.20625v1 |
8;34], or agent-based reasoning [ 11;45;100], to further improve the long context capability. 2.2 LLM Reasoning with Search and Planning Human problem-solving relies on decomposing complex tasks into manageable steps [ 19], a paradigm successfully adapted to LLMs. Early work by [ 85] and [ 36] introduced Chain-of-Thoug... | https://arxiv.org/abs/2505.20625v1 |
+ 1, l)otherwise rev← −rev; replay times ←replay times + 1; end else Break ; end A←Parse (ˆA); return A;The Dynamic Partition method uses a satura- tion function: for shorter inputs, it increases unit chunk length; for longer inputs, the length sat- urates and chunk count grows. Let wbe the input token length, nthe tar... | https://arxiv.org/abs/2505.20625v1 |
where chunks are vertices and dependencies are edges. When MRT (max replay times) > x−1, XpandA guarantees to complete dependency resolution A if LLM can finish information extraction in a chunk unit. 4 Experiment 4.1 Experiment Setup Datasets. We conduct comprehensive experiments on long context datasets from LongBenc... | https://arxiv.org/abs/2505.20625v1 |
Progress Score ffrom AgentBoard [ 54] for quantification of the multi-turn LLM agents’ performance which is denoted by: rt= max i,0≤i≤t|G ∩ P i| |G|(9) where an overall goal can be done by conjunction g1∧g2, . . . ,∧gmof atomic subgoal G= {g1, g2, . . . , g m}. While Pi={p1, p2, . . . , p m}express processed subgoals a... | https://arxiv.org/abs/2505.20625v1 |
XpandA, possibly due to Llama’s weaker ability to follow the multi-agent workflow instructions in XpandA, leading to a performance decrease. Additionally, RAG performs poorly in multi-hop tasks, likely because it cannot iteratively retrieve and reason about information step-by-step, a core mechanism enabling XpandA to ... | https://arxiv.org/abs/2505.20625v1 |
rate by18% over the best static variant. 5.2 LLM Reasoning Scheme Ablation (a) Reasoning scheme (F1 Score). (b) Replay strategy (Time). (c) Replay strategy (F1 Score). Figure 4: Ablation studies on NarrativeQA. (a) Comparison of XpandA with alternative planning schemes. (b) Impact of different replay strategies on aver... | https://arxiv.org/abs/2505.20625v1 |
static strategies (2k/8k) used in CoA [ 100] and LongAgent [ 101], which suffer from accumulated per-agent latency—especially critical for commercial API calls. 6 Conclusion In this paper, we propose XpandA, a novel paradigm for multi-agent LLM workflow in long context language modeling through continuous questioning-a... | https://arxiv.org/abs/2505.20625v1 |
benchmarks for long-context code models, 2024. [8]Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Ma... | https://arxiv.org/abs/2505.20625v1 |
Zhang, Shaoting Zhang, and Tong Ruan. MedOdyssey: A Medical Domain Benchmark for Long Context Evaluation Up to 200K Tokens, June 2024. arXiv:2406.15019 [cs]. [22] Tianyu Fu, Haofeng Huang, Xuefei Ning, Genghan Zhang, Boju Chen, Tianqi Wu, Hongyi Wang, Zixiao Huang, Shiyao Li, Shengen Yan, Guohao Dai, Huazhong Yang, and... | https://arxiv.org/abs/2505.20625v1 |
the Association for Computational Linguistics , 6:317–328, 2017. [36] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems , 35:22199–22213, 2022. [37] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Y... | https://arxiv.org/abs/2505.20625v1 |
Shilong Li, Ziqiang Liu, Yong Shan, Yifan Song, Jiayi Tian, Wenhao Wu, Zhejian Zhou, Ruijie Zhu, Junlan Feng, Yang Gao, Shizhu He, Zhoujun Li, Tianyu Liu, Fanyu Meng, Wenbo Su, Ying Tan, Zili Wang, Jian Yang, Wei Ye, Bo Zheng, Wangchunshu Zhou, Wenhao Huang, Sujian Li, and Zhaoxiang Zhang. A comprehensive survey on lon... | https://arxiv.org/abs/2505.20625v1 |
Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Ting... | https://arxiv.org/abs/2505.20625v1 |
agent through multi-persona self-collaboration. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , Jun. 2024. [84] Zhenyi Wang et al. Large language model enabled semantic communication systems, 2024... | https://arxiv.org/abs/2505.20625v1 |
mechanisms for LLM agents: A social psychology view. In Proceedings of the Annual Meeting of the Association for Computational Linguistics , Aug. 2024. [98] Xinrong Zhang, Yingfa Chen, Shengding Hu, Zihang Xu, Junhao Chen, Moo Hao, Xu Han, Zhen Thai, Shuo Wang, Zhiyuan Liu, and Maosong Sun. ∞Bench: Extending long conte... | https://arxiv.org/abs/2505.20625v1 |
Zhang, and Zheng Hu. Pose: Suppressing perceptual noise in embodied agents for enhanced semantic navigation. IEEE Robotics and Automation Letters , 9:963–970, 2024. A Proof of Completeness Hypothetical Premise : the LLM( I,θ) acting an agent can complete the problem decomposition and reasoning of a specified context ch... | https://arxiv.org/abs/2505.20625v1 |
Additionally, two distinct "needles" were synthetically generated for each trial, whose depth distribution is represented on the Y-axis of the heatmaps. For a given trial, the two needles were semantically independent but both were required to answer the query. The "Depth Percent of Needles", shown on the Y-axis of the... | https://arxiv.org/abs/2505.20625v1 |
arXiv:2505.20630v1 [cs.SE] 27 May 2025SV-TrustEval-C: Evaluating Structure and Semantic Reasoning in Large Language Models for Source Code Vulnerability Analysis Yansong Li∗, Paula Branco∗, Alexander M. Hoole†, Manish Marwah†, Hari Manassery Koduvely†, Guy-Vincent Jourdan∗, Stephan Jou† ∗University of Ottawa,†OpenText ... | https://arxiv.org/abs/2505.20630v1 |
variations in code within real-world scenarios. To overcome these limitations, we formally define the core capabilities required for vulnerability analysis and introduce SV-T RUST EVAL-C, a new question-answering benchmark designed for the trustworthy evaluation of LLMs in source code vulnerability analysis. This bench... | https://arxiv.org/abs/2505.20630v1 |
de- tection accuracy and efficiency [43], [44]. Advances in ma- chine learning and deep learning have significantly enhanced vulnerability detection capabilities [43], [45], [46], [47], [48], [49]. More recently, LLMs [1] have been explored for their potential to understand complex code patterns, predict vulnerabilitie... | https://arxiv.org/abs/2505.20630v1 |
/childfor manually labeled real-world data, /cogfor automatically labeled real-world data, and /editfor synthetic data. Support levels are shown as ✓(full),✓⃝(partial), and ✘(none). Key features of our benchmark—CWE coverage, function count, and reasoning capability—are highlighted relative to S ECLLMH OLMES . Evaluati... | https://arxiv.org/abs/2505.20630v1 |
robustness against introducing vulnerabilities. 2.1.4. Question-Answering Tasks. Question-Answering (QA) tasks tests the LLM model’s knowledge and under- standing of vulnerability concepts by requiring it to ac- curately and thoroughly answer security related questions with contextually appropriate explanations. Benchm... | https://arxiv.org/abs/2505.20630v1 |
reliable ground truth information, although resulting in smaller datasets. Third, synthetic sources such as the Juliet Test Suite [32] and Romeo [57] systematically generate code with high 2. https://cwe.mitre.org/top25/label accuracy, though risks of oversimplification remain. Our benchmark dataset is built on synthet... | https://arxiv.org/abs/2505.20630v1 |
C =A: OrgFunc (Var B)If (0) { } Var B= VulFunc (Var B) Var C =B: OrgFunc (Var B)If (0) { } Var B= VulFunc( Var B) If (1) {Bypass } Else { } Var C =C: OrgFunc (Var B)If (1) { } Var B= VulFunc (Var B)D: No idea.. Var C = OrgFunc (Var B)[Mask1] { } Var B= VulFunc (Var B) [Mask2] {Bypass } [Mask3] { } A: [Mask1] = If (0), ... | https://arxiv.org/abs/2505.20630v1 |
If code {X} is: If code {X} is: If code {Y} is:If code {X} is:If code {X} is: If code {Y} is:If code {Y} is:If code {X} is:If code {X} is: Counterfactual Goal-Driven Predictive Question: Examine the following Code Snippet . Which set of [Mask] values will not trigger `CWE-x ` weakness while maintaining the original fun... | https://arxiv.org/abs/2505.20630v1 |
no new vulnerabilities are introduced. •Code templates emphasize control statements that can either bypass, trigger, or prevent vulnerabilities. By ensuring the resulting code remains secure and func- tional, these scenarios evaluate the model’s proficiency in context-aware vulnerability reasoning and code refinement. ... | https://arxiv.org/abs/2505.20630v1 |
within Gc. This step tests if LLMs can maintain consistency with code variants that have similar syntax but different structures. As illustrated in Figure 2, the simulator then introduces Outer, Inner, and Outer&Inner structures to manage the code’s behavior, infusing Csafe,Cunsafe and ourCimpaired with masked control ... | https://arxiv.org/abs/2505.20630v1 |
We generate code variants us- ing the Structure-Oriented Variants Generator with masked control statements as conditions. LLMs are then tasked with DataFlow-wise (2430) Contr olFlow-wise (1345) Counterfactual (3748) Goal-driven (1159) Predictive (719)None (1058) 1-Hop (993)2-Hop (307) 3-Hop (62)4-Hop (10) None (652) 1-... | https://arxiv.org/abs/2505.20630v1 |
ICL, while the ↑symbol signifies an improvement in performance. ModelsStructure Reasoning Semantic Reasoning Base Scenario DataFlow ControlFlow Average Counterfactual Goal-driven Predictive AverageUnsafe Safe Average FPRSafe Zero ICL Zero ICL Zero ICL Zero ICL Zero ICL Zero ICL Zero ICL >15B Param. Models GPT-4-turbo 6... | https://arxiv.org/abs/2505.20630v1 |
understanding and reasoning abilities. In contrast, the in-context learning approach provides models with specific prompts that demonstrate question-and-answer patterns along with corresponding explanations for each answer. This contextual information guides their responses, enhancing their ability to learn from contex... | https://arxiv.org/abs/2505.20630v1 |
scenarios demanding extensive domain knowledge. We also conducted a baseline assessment by classifying code from the original Juliet Test Suite as safe orunsafe using a QA format with the options: A) Vulnerable, B) Non-Vulnerable, C) Do Not Know. The results presented in Table 2 indicate that most models perform unsati... | https://arxiv.org/abs/2505.20630v1 |
condition inside is true, and 0 otherwise. •Ci safeandCi unsafe denote the correctness indicators for safe andunsafe classifications in the base scenario for thei-th case. •Ci base=I Ci safe= 1∧Ci unsafe= 1 is the correctness indicator for the base scenario for the i-th case. •Ci DFL andCi CFL are the correctness ind... | https://arxiv.org/abs/2505.20630v1 |
“None” Level, but Improved with Increased Complexity: Conversely, models like CodeQwen ,CodeLlama , and CodeGemma exhibit aver- age performance below 30% at the “ None ” level. Surpris- ingly, these models perform better in scenarios requiring connections of ≥3hops. This counterintuitive outcome arises because these LL... | https://arxiv.org/abs/2505.20630v1 |
↑ 32.54 31.56 ↓ 33.51 36.07 ↑ 1131 -\w CI 41.46 50.00↑ 35.77 32.32 ↓ 38.62 55.28↑ 36.99 34.15 ↓ 31.10 32.32 ↑ 32.32 32.93 ↑ 33.13 32.93 ↓ 32.93 34.55 ↑ 35.57 32.72 ↓ 32.93 32.52 ↓ 36.59 31.91 ↓ 492 Inner 45.73 53.38↑ 38.79 40.04 ↑ 44.48 62.81↑ 40.04 39.32 ↓ 37.19 37.90 ↑ 37.54 37.54 ↕ 37.54 37.54 ↕ 32.38 31.67 ↓ 38.43 ... | https://arxiv.org/abs/2505.20630v1 |
17.46 ↓ 33.33 30.16 ↓ 63 O.&I. 19.12 32.35↑ 19.12 25.74 ↑ 30.88 46.32↑ 24.26 27.94 ↑ 26.47 28.68 ↑ 24.26 27.21 ↑ 17.65 30.88 ↑ 23.53 18.38 ↓ 19.12 21.32 ↑ 25.74 22.79 ↓ 27.94 27.94 ↕ 136 -\w CI 14.29 33.67 ↑ 22.45 24.49 ↑ 27.55 44.90↑ 24.49 29.59 ↑ 27.55 37.76↑ 21.43 29.59 ↑ 20.41 19.39 ↓ 19.39 37.76↑ 18.37 19.39 ↑ 22.... | https://arxiv.org/abs/2505.20630v1 |
-series models demon- /uni00000027/uni0000004c/uni00000055/uni00000048/uni00000046/uni00000057/uni00000003/uni00000026/uni00000011 /uni00000031/uni00000052/uni00000057/uni00000003/uni00000026/uni00000011 /uni0000002c/uni00000051/uni00000047/uni0000004c/uni00000055/uni00000048/uni00000046/uni00000057/uni00000003/uni0000... | https://arxiv.org/abs/2505.20630v1 |
/uni00000017/uni00000014/uni00000011/uni00000014/uni0000001b /uni00000016/uni00000018/uni00000011/uni00000015/uni0000001c /uni00000017/uni00000014/uni00000011/uni00000014/uni0000001b /uni00000015/uni0000001c/uni00000011/uni00000017/uni00000014 /uni00000015/uni0000001c/uni00000011/uni00000017/uni00000014 /uni00000015/un... | https://arxiv.org/abs/2505.20630v1 |
/uni00000013/uni00000011/uni00000013/uni00000013 /uni00000016/uni00000013/uni00000011/uni00000013/uni00000013 /uni00000013/uni00000011/uni00000013/uni00000013 /uni00000017/uni00000013/uni00000011/uni00000013/uni00000013 /uni00000013/uni00000011/uni00000013/uni00000013 /uni00000017/uni00000013/uni00000011/uni00000013/un... | https://arxiv.org/abs/2505.20630v1 |
/uni00000014/uni0000001b/uni00000011/uni00000014/uni0000001b /uni00000014/uni0000001b/uni00000011/uni00000014/uni0000001b /uni00000018/uni00000017/uni00000011/uni00000018/uni00000018 /uni00000014/uni00000019/uni00000011/uni00000019/uni0000001a /uni00000014/uni00000019/uni00000011/uni00000019/uni0000001a /uni00000015/un... | https://arxiv.org/abs/2505.20630v1 |
/uni00000017/uni00000018/uni00000011/uni0000001b/uni00000016 /uni00000017/uni00000016/uni00000011/uni00000013/uni00000019 /uni00000016/uni00000016/uni00000011/uni00000016/uni00000016 /uni00000016/uni00000016/uni00000011/uni00000016/uni00000016 /uni00000016/uni00000016/uni00000011/uni00000016/uni00000016 /uni00000016/un... | https://arxiv.org/abs/2505.20630v1 |
/uni00000015/uni00000017/uni00000011/uni00000016/uni00000015 /uni00000016/uni0000001a/uni00000011/uni0000001b/uni00000017 /uni00000016/uni00000018/uni00000011/uni00000014/uni00000017 /uni00000015/uni0000001c/uni00000011/uni0000001a/uni00000016 /uni00000014/uni0000001b/uni00000011/uni0000001c/uni00000015 /uni00000013/un... | https://arxiv.org/abs/2505.20630v1 |
/uni00000015/uni00000013/uni00000011/uni0000001b/uni00000016 /uni0000001a/uni00000011/uni0000001b/uni00000017 /uni00000015/uni00000014/uni00000011/uni00000018/uni0000001a /uni00000018/uni0000001b/uni00000011/uni0000001b/uni00000015 /uni00000014/uni0000001a/uni00000011/uni00000019/uni00000018 /uni00000017/uni00000014/un... | https://arxiv.org/abs/2505.20630v1 |
26.84 29.60 55.97 53.05 54.51 46.95 research [74], [75], [76], [77]. Introducing in-context learn- ing slightly mitigates this tendency, though a substantial gap remains. In the Counterfactual scenario, some LLMs cannot differentiate cases of “Bypass” code, struggling to correctly determine the code’s execution path. T... | https://arxiv.org/abs/2505.20630v1 |
semantically distinct safe&unsafe label pairs required for meaningful pairwise comparison. As shown in Figure 8, performance in the Base scenario increases from 60.7% to 86.2% under pairwise evaluation, but drops from 36.9% to 12.3% in the Counterfactual scenario, consistently across all difficulty levels. We also exam... | https://arxiv.org/abs/2505.20630v1 |
however, models like GPT-4 may underperform in certain Goal-driven cases. Overall, these findings underscore the importance of tailoring prompt designs to each CWE for effective vulnerability analysis. 5.5. Effects of Temperature We conducted an ablation study on inference tempera- ture (ranging from 0.0 to 1.0 in 0.2 ... | https://arxiv.org/abs/2505.20630v1 |
foundation language models,” arXiv preprint arXiv:2302.13971 , 2023. 5. https://github.com/Jackline97/SV-TrustEval-C [4] B. Roziere, J. Gehring, F. Gloeckle, S. Sootla, I. Gat, X. E. Tan, Y . Adi, J. Liu, T. Remez, J. Rapin et al. , “Code llama: Open foundation models for code,” arXiv preprint arXiv:2308.12950 , 2023. ... | https://arxiv.org/abs/2505.20630v1 |
Engineering , 2021, pp. 1565–1569. [21] J. Fan, Y . Li, S. Wang, and T. N. Nguyen, “Ac/c++ code vulnerability dataset with code changes and cve summaries,” in Proceedings of the 17th International Conference on Mining Software Repositories , 2020, pp. 508–512.[22] Y . Chen, Z. Ding, L. Alowain, X. Chen, and D. Wagner, ... | https://arxiv.org/abs/2505.20630v1 |
A benchmark for evaluating llms in cyber threat intelligence,” arXiv preprint arXiv:2406.07599 , 2024. [36] M. Bhatt, S. Chennabasappa, Y . Li, C. Nikolaidis, D. Song, S. Wan, F. Ahmad, C. Aschermann, Y . Chen, D. Kapil et al. , “Cyberseceval 2: A wide-ranging cybersecurity evaluation suite for large language models,” ... | https://arxiv.org/abs/2505.20630v1 |
Y . Wu, Y . Li et al. , “Deepseek-coder: When the large language model meets programming–the rise of code intelligence,” arXiv preprint arXiv:2401.14196 , 2024. [52] Y . Nong, R. Fang, G. Yi, K. Zhao, X. Luo, F. Chen, and H. Cai, “Vgx: Large-scale sample generation for boosting learning-based software vulnerability ana... | https://arxiv.org/abs/2505.20630v1 |
preprint arXiv:2406.11409 , 2024. [67] A. Q. Jiang, A. Sablayrolles, A. Roux, A. Mensch, B. Savary, C. Bamford, D. S. Chaplot, D. d. l. Casas, E. B. Hanna, F. Bressand et al. , “Mixtral of experts,” arXiv preprint arXiv:2401.04088 , 2024. [68] OpenAI, “Openai o1 system card,” https://openai.com/index/ openai-o1-system-... | https://arxiv.org/abs/2505.20630v1 |
L)` and located within `if (hModule != NULL)`. B:Modifying `if (hModule != NULL)` has no impact on the behavior or output of `FreeLibrary`. C:The change of if (hModule != NULL) indirectly affects the execution of `FreeLibrary`, as the modifications in `if (hModule != NULL)` influence the ar guments or conditions within... | https://arxiv.org/abs/2505.20630v1 |
does it af fect the functionality of the code? Options: A:No, Function Preserved: The vulnerability CWE1 14 will not be triggered, and the original functionality is fully preserved. B:No, Function Impaired: The vulnerability CWE1 14 will not be triggered, but the functionality of the original code is impaired. C:Yes: T... | https://arxiv.org/abs/2505.20630v1 |
clear evidence of one, so I choose option D. There is no CWE-1 14 issue in the current code snippet; therefore, the answer is D.The provided options do not include CWE- 114; therefore, the answer is D. Counterfactual Goal-driven PredictiveCWE-1 14, "Process Control," refers to vulnerabilities in managing system process... | https://arxiv.org/abs/2505.20630v1 |
strengthen the benchmark’s generalizability. 2) Generalization Beyond Juliet Dataset: The benchmark’s reliance on the Juliet Test Suite raises questions about its applicability to real-world vulnerabilities. The au- thors have demonstrated extensibility by applying their approach to the PrimeVul dataset, but further va... | https://arxiv.org/abs/2505.20630v1 |
Test-Time Learning for Large Language Models Jinwu Hu1 2 *Zitian Zhang1 *Guohao Chen1 2 *Xutao Wen1Chao Shuai3Wei Luo2 4 * Bin Xiao5†Yuanqing Li2†Mingkui Tan1 6† Abstract While Large Language Models (LLMs) have ex- hibited remarkable emergent capabilities through extensive pre-training, they still face critical limi- t... | https://arxiv.org/abs/2505.20633v1 |
environments. Most existing methods can be broadly categorized into four types, as shown in Table 1. Fine- tuning (Hu et al., 2022; Thirunavukarasu et al., 2023) adapts pre-trained models to specific tasks by updating their pa- rameters with labeled data, but it is constrained by the need for extensive labeled datasets... | https://arxiv.org/abs/2505.20633v1 |
(see Observation 2 ). Building on this observation, we propose a Sample Efficient Learning Strategy that employs a perplexity-based weighting scheme to actively select and emphasize high-perplexity test samples for backpropagation, thereby facilitating efficient parameter updates during Test-Time Learning. Moreover, we... | https://arxiv.org/abs/2505.20633v1 |
model parameters during inference by utilizing unlabeled test data (Wang et al., 2021; Niu et al., 2022a; 2023; Chen et al., 2024b;a; Liang et al., 2024; Yi et al., 2024). This approach enables real-time adaptation to distributional shifts, making it suitable for scenarios where labeled data is unavailable or the test ... | https://arxiv.org/abs/2505.20633v1 |
been explicitly trained to handle such distribution shifts, resulting in less coherent or contextually appropriate text generation on OOD test samples x∼Q(x), where Q(x)̸=P(x). Test-Time Learning (TTL) seeks to improve the perfor- mance of LLMs in the target domain by adjusting the model using only test data. Specifica... | https://arxiv.org/abs/2505.20633v1 |
predicts a sequence of tokens (Devlin et al., 2019; Brown et al., 2020). Given a sequence of tokens {x1, x2, ..., x T}, the perplexity Pis defined as the exponentiation of the average negative log-likelihood of the predicted tokens: P({x1, x2, ..., x T}) =e(−1 TPT t=1logp(xt|x1:t−1;Θ)),(2) where logp(xt|x1:t−1; Θ)is th... | https://arxiv.org/abs/2505.20633v1 |
logPΘ′(y|x)≥logPΘ(y|x)for small η(We compute the gradient inner product using 400 batches (batch size = 50) of QA pairs from the Domain- Bench on LLaMA3.1-8B. Results show 98.75% of batch- samples satisfy the non-negativity condition, with average ⟨∇x,∇y⟩= +5 .60). This form is consistent with the autoregressive proper... | https://arxiv.org/abs/2505.20633v1 |
Test-Time Learning for Large Language Models Table 2. Comparison of experimental results on the DomainBench and InstructionBench of the AdaptEval (see Supp. B). We mark the better scores in bold for better visualization and easier interpretation. MethodDomainBench InstructionBench Geography Agriculture Medicine Finance... | https://arxiv.org/abs/2505.20633v1 |
at the beginning of training, with Ausing random Gaussian initialization and Bset to zero, and we update only ∆Θduring the Test-Time Learning. 5. Experiments 5.1. Experimental Settings Datasets. To evaluate the effectiveness of our TLM, we construct a comprehensive benchmark named AdaptEval , designed to cover diverse ... | https://arxiv.org/abs/2505.20633v1 |
the stability of outputs produced by LLMs, we apply greedy decoding with a temperature of 0 across all experiments. More details in Supp. C.2. The source code is available at https://github.com/Fhujinwu/TLM 5.2. Comparison Experiments We compare our proposed TLM, the original LLM, Tent, EATA, and COME to demonstrate th... | https://arxiv.org/abs/2505.20633v1 |
30% compared to the original Llama3-8B-Instruct model. No- tably, on the Medicine dataset, the improvement reaches 83.9%. The effectiveness of input perplexity minimization P(x; Θ) lies in its ability to enhance the LLM’s understand- ing and representation of the input, which helps improve the model’s adaptation to the... | https://arxiv.org/abs/2505.20633v1 |
excluded from TTL in Eqn. (6) Experiments on Quantized LLM. To evaluate the perfor- mance of our method on quantized LLMs, we conduct exper- iments on a 4-bit quantized version of Llama3-8B-Instruct, following the settings of QLoRA (Dettmers et al., 2024). From Table 5, our method also demonstrates strong perfor- mance... | https://arxiv.org/abs/2505.20633v1 |
test-time training for self-supervised test- time adaption. In International Conference on Artificial Intelligence and Statistics , pp. 3080–3090. PMLR, 2022. Bella, G., Helm, P., Koch, G., and Giunchiglia, F. Tackling language modelling bias in support of linguistic diversity. InThe 2024 ACM Conference on Fairness, Ac... | https://arxiv.org/abs/2505.20633v1 |
model pretrain- ing with autoregressive blank infilling. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.), Proceed- ings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pp. 320–335, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10... | https://arxiv.org/abs/2505.20633v1 |
Y ., Callan, J., and Neubig, G. Active retrieval augmented generation. In Bouamor, H., Pino, J., and Bali, K. (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pp. 7969–7992, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023. emnlp-... | https://arxiv.org/abs/2505.20633v1 |
id=22OTbutug9 . Liu, Y . Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 , 364, 2019. Liu, Y ., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., and Alahi, A. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing ... | https://arxiv.org/abs/2505.20633v1 |
org/CorpusID:201646309 . Ren, Y ., Cao, Y ., Guo, P., Fang, F., Ma, W., and Lin, Z. Retrieve-and-sample: Document-level event argument extraction via hybrid retrieval augmentation. In Rogers, A., Boyd-Graber, J., and Okazaki, N. (eds.), Proceed- ings of the 61st Annual Meeting of the Association for Computational Lingu... | https://arxiv.org/abs/2505.20633v1 |
Darrell, T. Tent: Fully test-time adaptation by entropy minimiza- tion. International Conference on Learning Representa- tions , 2021. Wang, Q., Hu, J., Li, Z., Wang, Y ., Hu, Y ., Tan, M., et al. Generating long-form story using dynamic hierarchical outlining with memory-enhancement. The 2025 Annual Conference of the ... | https://arxiv.org/abs/2505.20633v1 |
sun, R., Torralba, A., and Fidler, S. Aligning books and movies: Towards story-like visual explanations by watch- ing movies and reading books. In Proceedings of the IEEE international conference on computer vision , pp. 19–27, 2015. 13 Supplementary Materials for “Test-Time Learning for Large Language Models” In the S... | https://arxiv.org/abs/2505.20633v1 |
(Raffel et al., 2020), WebText2, Books 1, Books 2, and Wikipedia datasets. GPT-3 has demonstrated extraordinary few-shot and zero-shot learning capabilities on many language tasks. In addition to GPT series, many decoder-only models have been developed, such as OPT, LLaMA, Llama2, Llama3 from Meta (Zhang et al., 2022b;... | https://arxiv.org/abs/2505.20633v1 |
generation process by prepending retrieved documents to the original prompt without altering LLM parameters. IRCoT (Trivedi et al., 2023) enhances reasoning by interleaving chain-of-thought (CoT) generation and retrieval, ensuring access to more relevant information across iterative reasoning steps. SKR (Wang et al., 2... | https://arxiv.org/abs/2505.20633v1 |
(Sun et al., 2020), contrastive learning (Liu et al., 2021; Bartler et al., 2022), or reconstruction learning (Gandelsman et al., 2022). To address scenarios where modifying the training process or accessing source data is not feasible, Fully TTA methods directly update pre-trained models during testing. These methods ... | https://arxiv.org/abs/2505.20633v1 |
domain. Looking forward, there is great potential to extend our method to the realm of efficient BP-Free TTA, thereby further broadening the practical applicability of our approach in diverse real-world scenarios. B. AdaptEval Benchmark To the best of our knowledge, no existing benchmark is specifically designed to eva... | https://arxiv.org/abs/2505.20633v1 |
GeoSignal, Agriculture-QA, GenMedGPT-5k, and Wealth-Alpaca Lora. Each dataset is sourced from a broad range of specialized domains, enabling the measurement of large model performance on complex domain-specific knowledge and task execution. The distribution of sequence lengths for the dataset samples and an example tab... | https://arxiv.org/abs/2505.20633v1 |
a deep red scarf and a black fedora. Finish off the look with a classic cane and a golden pocket watch. Medicine : GenMedGPT-5k3with a total of 5.45k samples is a medical dialogue dataset generated by ChatGPT, and is designed to emulate real-life conversations between patients and doctors. It provides rich, scenario-sp... | https://arxiv.org/abs/2505.20633v1 |
large dataset focused on real-world user instructions, consisting of over 50k high- quality bilingual (Zh-En) instructions. The data is derived from real user-shared scenarios, offering diversity and broad applicability. Its format aligns with that of the Alpaca dataset, enabling seamless integration and usage. The dat... | https://arxiv.org/abs/2505.20633v1 |
It \u2019 s home to some of the clearest skies on the planet and offers incredible stargazing opportunities. \n Antarctica: For the adventurous traveler, Antarctica is a wild, unforgiving landscape like no other. It\u2019 s home to breathtaking glaciers, unspoiled coastlines, and an abundance of wildlife. With few visi... | https://arxiv.org/abs/2505.20633v1 |
the following article: \nArticle: Researchers believe that if mothers are exposed to more pesticides in the first few months of pregnancy, the babies born may be less intelligent. They believe that the embryonic brain begins to develop shortly after pregnancy, so the pre-pregnancy is the baby’s brain In the critical pe... | https://arxiv.org/abs/2505.20633v1 |
•Tent 0.4541 -1.2410 0.0011 0.0321 0.0000 0.0320 •EATA 0.6999 -0.7661 0.0703 0.3132 0.1441 0.2421 •TLM (Ours) 0.6902 -0.6738 0.0722 0.2996 0.1319 0.2225 Qwen2.5-7B-Instruct 0.7003 -0.5802 0.0823 0.2911 0.1125 0.2128 •Tent 0.6925 -0.9422 0.0810 0.2703 0.1293 0.2242 •EATA 0.6856 -0.9806 0.0723 0.2444 0.1182 0.2063 •TLM (... | https://arxiv.org/abs/2505.20633v1 |
on the Alpaca-GPT4 dataset. Specifically, compared to Llama3-8B-Instruct, the proposed method achieves a relative improvement of 4.30% in BERTScore. As shown in Table 15, the proposed method 23 Test-Time Learning for Large Language Models Table 11. Comparison of experimental results on the Agriculture dataset of Domain... | https://arxiv.org/abs/2505.20633v1 |
0.0911 •Tent 0.5525 -1.3487 0.0014 0.0113 0.0006 0.0089 •EATA 0.6072 -1.1079 0.0011 0.0139 0.0011 0.0100 •TLM (Ours) 0.7095 -0.4781 0.0486 0.2646 0.0836 0.1889 Llama2-13B-chat 0.6559 -0.5971 0.0158 0.1439 0.0397 0.0956 •Tent 0.6235 -0.5867 0.0100 0.1250 0.0261 0.0874 •EATA 0.6543 -0.4349 0.0145 0.1465 0.0410 0.1007 •TL... | https://arxiv.org/abs/2505.20633v1 |
Test-Time Learning for Large Language Models Table 16. Comparison of experimental results on the InstructionWild dataset of InstructionBench. Method BERTScore ↑ BLEURT ↑ BLEU ↑ Rouge-1 ↑ Rouge-2 ↑ Rouge-L ↑ Llama3.2-3B-Instruct 0.7019 -0.4837 0.0337 0.2796 0.0942 0.1699 •Tent 0.5409 -1.2307 0.0018 0.0341 0.0008 0.0341 ... | https://arxiv.org/abs/2505.20633v1 |
arXiv:2505.20645v1 [cs.CL] 27 May 2025STEER -BENCH : A Benchmark for Evaluating the Steerability of Large Language Models Kai Chen1,2, Zihao He1,2, Taiwei Shi1, Kristina Lerman2 1Department of Computer Science, University of Southern California 2Information Sciences Institute, University of Southern California {kchen03... | https://arxiv.org/abs/2505.20645v1 |
social, cultural, or ideological contexts. To address this gap, we introduce STEER - BENCH , a benchmark designed to evaluate population-specific steering in LLMs. STEER - BENCH draws on Reddit, a platform comprising thousands of topic-specific communities, many of which represent contrasting perspectives on shared iss... | https://arxiv.org/abs/2505.20645v1 |
et al., 2022; Wang et al., 2023) and in-context learning (Zhao et al., 2025; Edwards and Camacho-Collados, 2024). Recent research has proposed comprehensive benchmarks and frame- works to systematically evaluate the proficiency of LLMs in executing instructed tasks (Zeng et al., 2024; Zhou et al., 2023). These benchmar... | https://arxiv.org/abs/2505.20645v1 |
evaluation challenges. He et al. (2024c) and Santurkar et al. (2023) measure the opinionand affective alignment of LLMs to different social groups when being steered to mimic them. Several frameworks have been proposed to enhance role- playing capabilities. Wang et al. (2024a) introduce RoleLLM with the RoleBench datas... | https://arxiv.org/abs/2505.20645v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.