text string | source string |
|---|---|
Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Lu... | https://arxiv.org/abs/2505.19236v1 |
Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe ... | https://arxiv.org/abs/2505.19236v1 |
AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Madry, Alex Baker- Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino, Alex Renzin, Alex Tachard Passos, Alexander Kirillov, Alexi Christakis, Alexis Conneau, Ali Kamali, Allan Jabri, Allison Moyer, Alliso... | https://arxiv.org/abs/2505.19236v1 |
Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean-Bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Eti- enne Pot, Ivo Penchev, Gaël Liu, Francesco Visin, Kathleen Kenealy, Lucas Beyer, Xiaohai Zhai, Anton Tsitsulin, Róbert B... | https://arxiv.org/abs/2505.19236v1 |
Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, (EMNLP) , 2023. 3, 6 [40] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In The International Conference on Learning Representations (ICLR) , 2019. 18 [41] Guillermo Mar... | https://arxiv.org/abs/2505.19236v1 |
Charline Le Lan, Sammy Jerome, Anton Tsitsulin, Nino Vieillard, Piotr Stanczyk, Sertan Girgin, Nikola Momchev, Matt Hoffman, Shantanu Thakoor, Jean-Bastien Grill, Behnam Neyshabur, Olivier Bachem, Alanna Walton, Aliaksei Severyn, Alicia Parrish, Aliya Ahmad, Allen Hutchison, Alvin Abdagic, Amanda Carl, Amy Shen, Andy B... | https://arxiv.org/abs/2505.19236v1 |
Majed Ceh, Christoph Meinel, and Mathias Benedek. What’s creative about sentences? a computational approach to assessing creativity in a sentence generation task. Creativity Research Journal , 34(4):419–430, 2022. 6 [63] Yang Wu, Yao Wan, Zhaoyang Chu, Wenting Zhao, Ye Liu, Hongyu Zhang, Xuanhua Shi, and Philip S. Yu. ... | https://arxiv.org/abs/2505.19236v1 |
Information Processing Systems (NeurIPS) , 2023. 2, 3 [71] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let’s think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In IEEE/CVF Conference on Computer Vision and Pat... | https://arxiv.org/abs/2505.19236v1 |
CrEval to facilitate fu- ture research on creativity evaluation. A.3 Length Distribution of Different Sources We randomly sample 800 samples from each source in the dataset and present the distribution of the response lengths in Figure 11. For better visualization of their KDE curves, we applied a log transformation to... | https://arxiv.org/abs/2505.19236v1 |
66.25 66.6267.8768.7567.25 67.25 63.50 63.75 63.87 59.63 51.1260.87 61.25 61.0062.12 59.12 54.3752.50 52.37 51.62 44.37 37.37CrEval GPT-4o Figure 13: Win rate curves of incorporating different ratios of hard reject samples in DPO training, evaluated by CrEval and GPT-4o. B.3 How CrEval Enhance Model Creativity by Selec... | https://arxiv.org/abs/2505.19236v1 |
the overseeing critic will impose severe penalties. ### Given Text: {{Text}} ### Your reply: ###任务描述： 你是一个敏锐严厉的文艺评论家，你需要对{{category}} 的质量和创意程度进行判断。 ###具体要求： 1.创意的核心内涵是否新颖且有意义，判定的时候可以从用词、词序、句法、象征意 义、修辞手法、整体意象等方面综合判定。 2.如果一段文本包含了较多的创意要素，例如新奇的句法和表达，应该得到高分；如果 是简单的陈述，或是不适合作为创意回答，具有低创意潜力，则应该得到低分。 3.请先给出对文本的简要分析鉴赏，然后从1到6分给出你的... | https://arxiv.org/abs/2505.19236v1 |
data are listed in the upper part of the dashed line, and our constructed components are in the lower part. 24 Type: Creativity-Dense Texts Source: Short Texts Response: Sunflowers don’t cry; even when down, they face the sun. 向日葵没有眼泪即使沮丧也朝着阳光。 Constructed Instruction: Please share your thoughts on maintaining a positi... | https://arxiv.org/abs/2505.19236v1 |
upper part of the dashed line, and our constructed components are in the lower part. Owing to length constraints, the middle part of each response is omitted. 26 Type: Creativity-Dense Texts Synthetic Responses (continued): ❹[Song of the Wind Against Us] \n \n (Prelude) \n In the dark of night, we see the first light o... | https://arxiv.org/abs/2505.19236v1 |
and I apart. \n In endless darkness, we seek a glimmer of light. \n Distance is not just space—it’s a bridge between hearts. (By Qwen2.5-14B- Promptc) 宇宙广袤无垠，星河间微尘般漂泊，\n孤独的光年旅程，你我相隔，\n在无尽黑暗 中，寻找微光，\n距离，不仅是长度，更是心灵的桥梁。 ... Figure 19: An example from type Creativity-Dense Texts and source Modern Poetry . We present texts ... | https://arxiv.org/abs/2505.19236v1 |
youths, then into mature adults, and finally, elderly ones with gray hair. Each stage is a gift from time. ...... It teaches us that every embrace, every farewell, is an essential part of life. Time reminds us that life is like a journey, where the value lies not in the destination, but in the scenery along the way and... | https://arxiv.org/abs/2505.19236v1 |
arXiv:2505.19240v1 [cs.CL] 25 May 2025LLLMs: A Data-Driven Survey of Evolving Research on Limitations of Large Language Models AIDA KOSTIKOVA, University of Bielefeld, Germany ZHIPIN WANG, University of Technology Nuremberg, Germany DEIDAMEA BAJRI, University of Mannheim, Germany OLE PÜTZ, University of Bielefeld, Germ... | https://arxiv.org/abs/2505.19240v1 |
University of Bielefeld, Bielefeld, Germany, aida.kostikova@uni-bielefeld.de; Zhipin Wang, University of Technology Nuremberg, Nuremberg, Germany; Deidamea Bajri, University of Mannheim, Mannheim, Germany; Ole Pütz, University of Bielefeld, Bielefeld, Germany; Benjamin Paaßen, University of Bielefeld, Bielefeld, German... | https://arxiv.org/abs/2505.19240v1 |
robotics, and system efficiency [ 33,77], as well as reasoning and planning capabilities in large-scale models [ 69]. In parallel to cross-domain surveys, a number of studies have investigated how LLMs are being adopted and evaluated in specific fields. In the medical domain, surveys examine the effectiveness of LLMs i... | https://arxiv.org/abs/2505.19240v1 |
being misled because of the very limitations this survey is supposed to study. In developing our methodology, we rely on a growing literature of LLMs being used as instruments for analyzing scientific literature [ 22]. Several recent approaches employ LLMs for topic modeling, semantic clustering, and concept induction,... | https://arxiv.org/abs/2505.19240v1 |
was chosen to capture the year preceding the release of ChatGPT as well as all subsequent research on LLMs [125]. For ACL Anthology, we scrape conference pages for AACL 2022–2023, ACL 2022–2024, EACL 2023–2024, EMNLP 2022–2024, ICLR 2022–2024, NAACL 2022 and 2024, and TACL 2022–2024 as the premier NLP venues.1For arXiv... | https://arxiv.org/abs/2505.19240v1 |
keyword list focused on strongly distinctive terms as the process converges. This results in a list of 90 keywords (19 unigrams, 44 bigrams, 16 trigrams, and 11 four-grams). Overall, the final keyword set covers key aspects of LLM research (see the full list in Section A supplementary material): •terms related to LLMs,... | https://arxiv.org/abs/2505.19240v1 |
source (ACL or arXiv, ensuring conference representation within ACL) and publication year. Papers are manually annotated based on their titles and abstracts to assess whether they discuss LLLMs . The human annotators rated each paper on a scale from 0-5, reaching from no relation to LLMs (0) to exclusive focus on LLLMs... | https://arxiv.org/abs/2505.19240v1 |
limitations (papers with ratings 3–5 as a final label, 48 jointly annotated papers), F1 score increases to 0.71. This score suggests reliable consistency, given the known difficulty of span-level annotation [19]. 8 Last Name et al. 3.4 Models and Prompting Evaluation We evaluate models on the human-annotated dataset to... | https://arxiv.org/abs/2505.19240v1 |
use not the full abstract but only the passages that explicitly describe the LLM limitation the paper is concerned with, i.e. the evidence statements of papers rated 3-5 as extracted in Section 3.3. To enrich the text representation for clustering, we follow the approach of Viswanathan et al . [105] and generate keyphr... | https://arxiv.org/abs/2505.19240v1 |
LLM then generates a concept (a short, human-readable label that describes the theme of the cluster) and inclusion prompt for each cluster (synthesize step), which are used to score all documents for each concept via zero-shot prompting on a 0–1 scale (score step). For further implementation details, we refer the reade... | https://arxiv.org/abs/2505.19240v1 |
that both humans and the model often LLLMs: A Data-Driven Survey of Evolving Research on Limitations of Large Language Models 11 Table 5. Comparison of evidence extraction between human annotators and Llama-3.1-70b using Prompt 3. Text highlighted in green indicates parts which both the model and human annotators selec... | https://arxiv.org/abs/2505.19240v1 |
analysis. In the final classification step, Llama-3.1-70b-Instruct assigns each LLM-focused paper 12 Last Name et al. 0 1 2 3 4 50 1 2 3 4 52 3 0 0 0 0 80 23 2 0 0 105 24 3 0 23 7 0 14 5 9 (a)Annotators’ agreement 0 1 2 3 4 5 Predicted0 1 2 3 4 5True52 0 8 2 1 0 11 60 34 1 1 0 7 7 124 27 8 0 3 0 21 20 23 0 0 0 0 9 26 4... | https://arxiv.org/abs/2505.19240v1 |
2022Q12022Q22022Q32022Q42023Q12023Q22023Q32023Q42024Q12024Q22024Q32024Q42025Q10.200.250.300.35Proportion Limitation Papers As Proportion Of LLM Papers ACL arXiv(ii) Proportion of LLM limitation pa- pers among all LLM papers. Fig. 5. Trends in LLM and LLM limitation research over time. Figure 5i shows the share of LLM a... | https://arxiv.org/abs/2505.19240v1 |
and -cultural performance. Long Context: Handling long inputs and extended memory. Social Bias: Fairness, stereotypes (especially gender and cultural), and societal impacts. Healthcare Application: Challenges in domain-specific adaptation for clinical and mental health use. Code Generation: Errors in logic, syntax, or ... | https://arxiv.org/abs/2505.19240v1 |
political considerations, including ethical ,moral ,political , LLLMs: A Data-Driven Survey of Evolving Research on Limitations of Large Language Models 15 Reasoning Hallucination Security Generalization Social Bias Long Context Uncertainty (i) ACL dataset Security Risks Social Bias Hallucination Context & Memory Limit... | https://arxiv.org/abs/2505.19240v1 |
(EMNLP 2022 [59]) (6) Long Context“However, they face challenges in managing long documents and extended conversations, due to significantly increased computational requirements, both in memory and inference time [...].” (EMNLP 2023 [53]) (7) Uncertainty[...] “We find that methods aimed at improving usability, such as ... | https://arxiv.org/abs/2505.19240v1 |
for arXiv, adding concerns such as Alignment ,Trustworthiness , and Generalization . •Reasoning ,Generalization , and Hallucination are top-ranking topics in both ACL and arXiv. Beyond these shared concerns, ACL places additional emphasis on Knowledge Editing , while arXiv is led by Trustworthiness andAlignment . Figur... | https://arxiv.org/abs/2505.19240v1 |
generalization. •Alignment Limitations (arXiv): Highlights challenges in aligning LLMs with human values or safety protocols (see example 20 which discusses how models can generate outputs that are untruthful, toxic, or unhelpful despite alignment efforts). •Prompt Sensitivity (arXiv): Highlights performance instabilit... | https://arxiv.org/abs/2505.19240v1 |
number of limitation papers on topic 𝑘and the total number of limitation papers, respectively, in quarter 𝑞. (i) How are limitation topics represented in the broader growth of LLM research? Key Insights •The presence of limitation topics in LLM research is increasing across both ACL and arXiv datasets. Topics like Ha... | https://arxiv.org/abs/2505.19240v1 |
doubling in share), such as Multimodality (+133%), Long Context (+108%), Catastrophic Forgetting (+140%) in ACL 2024, and Hallucination (+223%), Security Risks (+163%), and Alignment Limitations (+102%) on arXiv in 2023, reflecting heightened attention to certain types of LLLMs following widespread LLM deployment. Othe... | https://arxiv.org/abs/2505.19240v1 |
a few showing significant shifts. •Long Context increases significantly in ACL, while Multimodality ,Security Risks , and Align- ment Limitations rise in arXiv. •Generalization andBias and Fairness decline significantly in arXiv, but no topics show statisti- cally significant decline in ACL. 22 Last Name et al. 2022-Q2... | https://arxiv.org/abs/2505.19240v1 |
over time: Reasoning ,Generalization , andHallucination fluctuate between 10–35%, while Knowledge Editing varies more widely (18–39%) and Multimodality stays between 6–11%, with a slight increase after early 2024. Privacy Risks (3–6%), Security Risks (peaking at 15% in 2022-Q3 but mostly under 10%), and Catastrophic Fo... | https://arxiv.org/abs/2505.19240v1 |
do show consistent growth within limitation-focused work, the majority remain flat or variable despite gaining visibility across the broader LLM field, as discussed earlier. However, the Mann-Kendall test only captures consistent upward or downward trends, not short-term changes, which may explain why topics like Secur... | https://arxiv.org/abs/2505.19240v1 |
Alignment Limitations , both their overall topic composition and temporal dynamics vary by field. Figure 18 (right) in the supplementary material, shows how topic shares differ across arXiv cate- gories, while Figure 12 illustrates how topics in the four largest categories evolve over time. Key trends in the largest ca... | https://arxiv.org/abs/2505.19240v1 |
the primary arXiv category. Topics that make up less than 3% are grouped under Other . reflecting ethical and user-centered concerns. cs.CR (Cryptography and Security) is dominated by Security Risks (57.7%), consistent with its focus on adversarial threats and privacy vulnerabilities. cs.IR (Information Retrieval) dist... | https://arxiv.org/abs/2505.19240v1 |
also correlate strongly, as reflected by high Spearman𝜌values. For instance, Multimodality (𝜌=0.86) and Multilinguality (𝜌=0.92) achieve strong and statistically significant trend similarity across methods. Nonetheless, slight divergences remain: although trend directions often align, significance levels or trend sh... | https://arxiv.org/abs/2505.19240v1 |
in the supplementary material, which show the most closely aligned topics across methods. Some large limitation areas appear relatively stable across clustering strategies: e.g. topics such as Reasoning ,Hallucination ,Security Risks , and Bias and Fairness align well across datasets 3AMI compares how often data points... | https://arxiv.org/abs/2505.19240v1 |
Other prominent topics include Generalization ,Hallucination ,Bias, and Security . Beyond these, LLM limitation research is notably diverse. Our clustering analyses (HDBSCAN and LlooM) reveal a broad spectrum of concerns, ranging from code generation and benchmark contamination toprompt sensitivity andlong context . Th... | https://arxiv.org/abs/2505.19240v1 |
stability of the main findings. We validate our results by comparing HDBSCAN+BERTopic (single-topic, density-based) with LlooM (LLM-based, multi-topic). Despite methodological differences, both clustering approaches identify the high-frequency topics, most notably Reasoning ,Hallucination , andSecurity Risks , showing ... | https://arxiv.org/abs/2505.19240v1 |
Eric Pan, Garry Kuwanto, and Derry Wijaya. 2023. Dune: Dataset for unified editing. arXiv preprint arXiv:2311.16087 (2023). [6]Mudassar Hassan Arsalan, Omar Mubin, Abdullah Al Mahmud, Imran Ahmed Khan, and Ali Jan Hassan. 2025. Mapping Data-Driven Research Impact Science: The Role of Machine Learning and Artificial Int... | https://arxiv.org/abs/2505.19240v1 |
Clustering. arXiv preprint arXiv:2502.09667 (2025). [21] Qinxu Ding, Ding Ding, Yue Wang, Chong Guan, and Bosheng Ding. 2024. Unraveling the landscape of large language models: a systematic review and future perspectives. Journal of Electronic Business & Digital Economics 3, 1 (2024), 3–19. [22] Steffen Eger, Yong Cao,... | https://arxiv.org/abs/2505.19240v1 |
2024. Large language models for software engineering: A systematic literature review. ACM Transactions on Software Engineering and Methodology 33, 8 (2024), 1–79. doi:10.1145/3695988 [37] Jie Huang and Kevin Chen-Chuan Chang. 2022. Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403 (2... | https://arxiv.org/abs/2505.19240v1 |
2023. Counterfactual reasoning: Testing language models’ understanding of hypothetical scenarios. arXiv preprint arXiv:2305.16572 (2023). [53] Yucheng Li, Bo Dong, Chenghua Lin, and Frank Guerin. 2023. Compressing context to enhance inference efficiency of large language models. arXiv preprint arXiv:2310.06201 (2023). ... | https://arxiv.org/abs/2505.19240v1 |
Capabilities and Limitations. arXiv preprint arXiv:2501.04040 (2025). [70] Leland McInnes, John Healy, Steve Astels, et al .2017. hdbscan: Hierarchical density based clustering. J. Open Source Softw. 2, 11 (2017), 205. [71] Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform manifold approximation and p... | https://arxiv.org/abs/2505.19240v1 |
A primer in BERTology: What we know about how BERT works. TACL 8 (2021), 842–866. doi:10.1162/tacl_a_00349 [88] Domenic Rosati, Giles Edkins, Harsh Raj, David Atanasov, Subhabrata Majumdar, Janarthanan Rajendran, Frank Rudz- icz, and Hassan Sajjad. 2024. Defending against reverse preference attacks is difficult. arXiv ... | https://arxiv.org/abs/2505.19240v1 |
Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al .2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023). [104] Dave Van Veen, Cara Van Uden, Louis Blankemeier, Jean-Benoit Delbrouck, Asad Aali, Ch... | https://arxiv.org/abs/2505.19240v1 |
Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, et al .2025. Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Models. arXiv preprint arXiv:2501.09686 (2025). [117] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucinatio... | https://arxiv.org/abs/2505.19240v1 |
Research on Limitations of Large Language Models 37 Evaluate the following paper's title and abstract to determine if it discusses Large Language Models (LLMs) and whether it addresses their limitations. If it does, indicate specific evidence from the abstract or title that points to the limitations and identify the do... | https://arxiv.org/abs/2505.19240v1 |
and focuses on other topics. **Example 3:** Title: Injecting New Knowledge into Large Language Models via Supervised Fine-Tuning Abstract : "In recent years, Large Language Models (LLMs) have shown remarkable performance ... However, adapting these models to incorporate new, out-of-domain knowledge remains a challenge ... | https://arxiv.org/abs/2505.19240v1 |
only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation. [...] ” [4], EMNLP 2023No mention of LLMs or their limitations. 1 “ We introduce the framework of ‘social learning’ in the context of LLMs, whereby models share knowledge with each other in a privac... | https://arxiv.org/abs/2505.19240v1 |
topic based on the topic probability distribution, but only if they fall within a certain distance threshold to ensure proximity to an existing cluster. 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Distance020406080100FrequencyUMAP Centroid Distance All Points Inliers Outliers 0.0 0.2 0.4 0.6 0.8 Distance05010015020025... | https://arxiv.org/abs/2505.19240v1 |
problems caused by models forgetting previ- ously learned information during fine-tuning or continual learning? Data Contamination Does the text example discuss issues of training data containing test data, causing overestimation of model performance through memorization? LLLMs: A Data-Driven Survey of Evolving Researc... | https://arxiv.org/abs/2505.19240v1 |
10.26% 28.21% Generalization 5.88 →5.99→5.83→5.95 1.87% -2.67% 2.06% Hallucination 1.28 →4.14→4.64→4.40 223.44% 12.08% -5.17% Alignment Limitations 1.87 →3.78→4.51→5.23 102.14% 19.31% 15.96% Bias and Fairness 3.73 →4.14→4.25→4.12 10.99% 2.66% -3.06% Security Risks 1.04 →2.74→3.78→3.68 163.46% 37.96% -2.65% Multimodalit... | https://arxiv.org/abs/2505.19240v1 |
the main category only and broken down by category. 0 10 20 30 40 50 60 Percentage of Papers (%)cs.CL cs.LG cs.AI cs.CV cs.CY cs.IR cs.CR cs.SE cs.RO cs.HC58.7% 13.3% 8.7% 6.6% 3.2% 1.6% 1.2% 0.6% 0.6% 0.6%T op ArXiv Categories (LLM Limitation Papers) TrustworthinessReasoning Generalization Alignment LimitationsHalluci... | https://arxiv.org/abs/2505.19240v1 |
Topic Jaccard Overlap Multimodality Multimodality 0.379 Alignment Limitations Hallucination 0.160 Hallucination Hallucination 0.477 Trustworthiness Security Risks 0.263 Generalization Reasoning 0.116 Language and Cultural Limitations Multilinguality 0.393 Bias and Fairness Social Bias 0.288 Privacy Risks Security Risks... | https://arxiv.org/abs/2505.19240v1 |
and Knowledge Editing . However, these topics are the least represented, and since 2025 data covers only one quarter, recent drops should be interpreted with caution. These comparisons reflect annual, macro-level trends in topical focus and relative prominence. To examine shorter-term dynamics and account for changes i... | https://arxiv.org/abs/2505.19240v1 |
quarters. Due to the inconsistency of this change, it is not LLLMs: A Data-Driven Survey of Evolving Research on Limitations of Large Language Models 49 2022-Q2 2022-Q3 2022-Q4 2023-Q2 2023-Q3 2023-Q4 2024-Q1 2024-Q2 2024-Q3 2024-Q4 Quarter01020304050Percentage of All Papers (%) iclr2022 acl2022 naacl2022 aacl2022emnlp... | https://arxiv.org/abs/2505.19240v1 |
but no sustained change. These patterns are consistent with the Mann-Kendall test, which detects no significant trends for any of these topics. Some topics also show sharp, isolated changes. Social Bias spikes in 2022-Q3 (26%), then drops to around 10% by early 2023. This may reflect heightened concern about bias in th... | https://arxiv.org/abs/2505.19240v1 |
arXiv Dataset Topic Topic Shares (2022 →2025)→2023 (%)→2024 (%)→2025 (%) Social Bias 0.17 →0.68→1.37→1.61 300.00% 101.47% 17.52% Security Risks 0.11 →0.59→1.33→1.66 436.36% 125.42% 24.81% Reasoning 0.12 →0.46→0.79→1.32 283.33% 71.74% 67.09% Context & Memory Limitations 0.17 →0.39→0.78→1.0 129.41% 100.00% 28.21% Multimo... | https://arxiv.org/abs/2505.19240v1 |
co-occurrence matrices for ACL and arXiv LLM limitation papers, clustered by LlooM approach. In both ACL and arXiv papers, topic co-occurrences tend to concentrate around the largest clusters (see Figure 21). The main patterns are as follows: •Reasoning shows the highest co-occurrence with other limitations, especially... | https://arxiv.org/abs/2505.19240v1 |
arXiv:2505.19250v1 [cs.CL] 25 May 2025PATS: Process-Level Adaptive Thinking Mode Switching Yi Wang*, Junxiao Liu*, Shimao Zhang*, Jiajun Chen, Shujian Huang† National Key Laboratory for Novel Software Technology, Nanjing University {yiw,junxiaoliu,smzhang}@smail.nju.edu.cn chenjj@nju.edu.cn, huangsj@nju.edu.cn Abstract... | https://arxiv.org/abs/2505.19250v1 |
employed by LLMs. Prior work has ex- plored mechanisms for switching between System 1 and System 2 in LLMs, broadly categorized into training-based (Su et al., 2024; Saha et al., 2024; Cheng et al., 2025) and training-free methods (Yao et al., 2024). This study focuses on the training-free setting. HDFLOW (Yao et al., ... | https://arxiv.org/abs/2505.19250v1 |
voked to re-solve the problem. The switch between thinking systems occurs after a complete solution is derived, representing a solution-level strategy adjustment with coarse granularity. In contrast, our approach enables finer-grained, step-level mode switching during the reasoning process. 2.2 Test-time Scaling Test-t... | https://arxiv.org/abs/2505.19250v1 |
1as the reasoning context (Wang et al., 2025).Given a problem q, the policy model performs step-by-step reasoning, generating a solution path S={s1, s2, ..., s n}, where sndenotes the n-th reasoning step. BFS consists of two iterative oper- ations: expansion and selection. In the expansion stage of step i, the model ge... | https://arxiv.org/abs/2505.19250v1 |
path (Wang et al., 2023). Studies use PRM score of the final answer to evaluate problem difficulty (Snell et al., 2024), with higher scores indicating easier problems for the model. Addi- tionally, inspired by reward signal can guide re- source allocation (Sun et al., 2024), Fu et al. (2024) collect the terminal reward... | https://arxiv.org/abs/2505.19250v1 |
al., 2023), Minerva Math (Lewkowycz et al., 2022), AMC23 , and AIME24 , which collectively span elementary, intermediate, and advanced levels of mathematical reasoning. Policy Models. We utilize the Instruct variants of theQwen2.5 family (Yang et al., 2024a) and con- duct experiments with models of varying parameter si... | https://arxiv.org/abs/2505.19250v1 |
1929.9 23.3 5821.0 61.3 2808.0 Table 1: Comparison of average accuracy and token usage on five math reasoning benchmarks. Based on the average metrics, Our method PATS achieves effective and efficient reasoning with an excellent accuracy-efficiency balance. •Solution-verification Switch : It first gener- ates a complet... | https://arxiv.org/abs/2505.19250v1 |
adaptability. This underscores the rationality and effectiveness of our pipeline’s switching strategy design. In summary, PATS dynamically adjusts reason- ing strategies based on step-wise difficulty, has been empirically validated. The results demon- strate that well-designed adaptive thinking mode 6 SettingLevel 1 (E... | https://arxiv.org/abs/2505.19250v1 |
and aligning with the greater complexity of AMC23 problems. This shows that harder tasks 7 Setting GSM8K MATH500 AMC23 MinervaMATH AIME24 Average Acc↑Token ↓Acc↑Token ↓Acc↑Token ↓Acc↑Token ↓Acc↑Token ↓Acc↑Token ↓ PATS-No-Penalty 93.9 829.4 79.0 1931.3 47.5 2604.0 42.6 1836.8 16.7 4612.9 55.9 2362.9 PATS-Infinite-Penalt... | https://arxiv.org/abs/2505.19250v1 |
highlight the robustness of our adaptive paradigm across a wide range of policy models and PRMs. 6 Conclusion In this paper, we propose a novel reasoning paradigm— Process-Level Adaptive Thinking Mode Switching (PATS) . This method leverages PRM scores during the reasoning process to es- timate the current difficulty a... | https://arxiv.org/abs/2505.19250v1 |
Chen, Yingy- ing Zhang, Fei Yin, Jiahua Dong, Zhiwei Li, Bao- Long Bi, Ling-Rui Mei, Junfeng Fang, Zhijiang Guo, Le Song, and Cheng-Lin Liu. 2025. From system 1 to system 2: A survey of reasoning large language models. Preprint , arXiv:2502.17419. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen ... | https://arxiv.org/abs/2505.19250v1 |
. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, and 1 others. 2024b. Qwen2. 5-math technical report: Toward mathe- matical expert model via self-improvement. arXiv preprint arXiv:2409.12122 . Wenlin Yao, Haitao Mi, and Dong Yu. 2024. Hdflow... | https://arxiv.org/abs/2505.19250v1 |
Unveiling Dual Quality in Product Reviews: An NLP-Based Approach Rafał Po ´swiata, Marcin Michał Miro ´nczuk, Sławomir Dadas, Małgorzata Gr˛ ebowiec, Michał Perełkiewicz National Information Processing Institute al. Niepodległo ´sci 188b, 00-608 Warsaw, Poland {rposwiata, mmironczuk, sdadas, mgrebowiec, mperelkiewicz}@... | https://arxiv.org/abs/2505.19254v1 |
preparation, through extensive evaluation of differ- ent approaches, to deployment. To our knowledge, no available dataset or model is aimed at recog- nizing dual quality-related reviews. While several articles (discussed further in Section 2) approach 1In this article, we use the terms ‘reviews’ and ‘opinions’ interch... | https://arxiv.org/abs/2505.19254v1 |
may justify product variations based on local market prefer- ences, research suggests that these practices often lack transparency and leave consumers feeling de- ceived (Bartkova and Veselovska, 2023). Moreover, comparative consumer tests confirm that dual qual- ity is not confined to food products but also extends to... | https://arxiv.org/abs/2505.19254v1 |
amount of training data. The model was im- plemented using the SetFit (Sentence Transformer Fine-tuning) framework (Tunstall et al., 2022) and a sentence transformer for the Polish language st- polish-paraphrase-from-distilroberta4. 4⃝Apply the model trained in step 3⃝to all re- views of the CENEO / WIZAZ dataset. The ... | https://arxiv.org/abs/2505.19254v1 |
dual quality, 281as other problems, and the rest are standard opinions. Of the dual quality reviews, 1076were from the Internet, 265from the CENEO / WIZAZ collection, and 168 from our demo system. The dataset is unbalanced, with over half of the reviews belong to the standard class. This characteristic was intentionall... | https://arxiv.org/abs/2505.19254v1 |
77.7±0.5 75 .6±0.8 65.9±0.9 68.4±0.7 para-multi-mpnet-base-v2 72.8±1.7 66.4±2.4 69.4±2.0 75.9±1.4 72.4±2.2 66.8±2.5 68.8±2.6 para-multi-MiniLM-L12-v2 69.4±2.2 58.7±3.3 63.6±2.7 71.2±1.2 65.8±1.3 58.2±1.7 60.2±1.7 multi-e5-small 68.7±1.6 68.0±1.3 68.3±0.8 72.8±0.7 70.4±0.8 58.9±0.9 60.3±1.3 multi-e5-base 72.2±1.2 79 .0±... | https://arxiv.org/abs/2505.19254v1 |
the larger, language-specific models such as polish-roberta-large-v2 (84.6%) and herbert-large- cased (81.5%), exhibited significantly stronger performance, comparable even with state-of-the- art conversational large language models (LLMs). Among LLMs, instructive prompting strategies (providing clear definitions of cl... | https://arxiv.org/abs/2505.19254v1 |
applied to the test set for robustness verification. Modification gpt-4o polish-roberta herbert period 4.0±0.0 4.2±1.0 5.0±0.9 first_letter 4.0±0.0 2.8±0.7 2.6±0.8 lower 5.0±0.0 4.6±0.5 4.2±0.7 pl_chars 5.0±0.0 4.6±1.2 4.6±0.8 pl_chars_once 4.0±0.0 4.0±1.4 3.6±0.8 Table 4: Robustness verification results for GPT-4o (ze... | https://arxiv.org/abs/2505.19254v1 |
likely to be accurate. A product with several dual quality reviews will be selected for further analysis to verify whether this issue genuinely exists in its case. The proposed solution is implemented as a stan- dalone service within a local infrastructure and is exclusively dedicated to UOKiK employees(Poland’s Office... | https://arxiv.org/abs/2505.19254v1 |
Faculty of Economics and Administration , 26. I. Botunac, M. Brki ´c Bakari ´c, and M. Mateti ´c. 2024. Comparing fine-tuning and prompt engineering for multi-class classification in hospitality review analy- sis.Applied Sciences (Switzerland) , 14. Chambers. Dual Quality of Food Products. https://chambers.com/legal-tr... | https://arxiv.org/abs/2505.19254v1 |
Junlong Li, Junxiao Song, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Xu, Leyi Xia, Liang Zhao, Litong Wang, Liyue Zhang, Meng Li, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Ning Tian, Panpan Huang, Peiyi Wang, Peng Zhang, Qiancheng Wang, Qiha... | https://arxiv.org/abs/2505.19254v1 |
Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers) , pages 878–891, Dublin, Ireland. Association for Computa- tional Linguistics. Gilad Fuchs, Ido Ben-shaul, and Matan Mandelbrod. 2022. Is it out yet? automatic future product releases extraction from web data. In Proceedings of th... | https://arxiv.org/abs/2505.19254v1 |
Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Ful- ford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo- Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Green... | https://arxiv.org/abs/2505.19254v1 |
of the Association for Com- putational Linguistics (Volume 5: Industry Track) , pages 744–751, Toronto, Canada. Association for Computational Linguistics. Rafał Po ´swiata, Sławomir Dadas, and Michał Perełkiewicz. 2024. PL-MTEB: Polish Mas- sive Text Embedding Benchmark. Preprint , arXiv:2405.10138. Nils Reimers and Ir... | https://arxiv.org/abs/2505.19254v1 |
practices. However, challenges remain in enforce- ment and uniform interpretation across Member States (EU Monitor). Recent research shows that while the prevalence of dual quality food products declined from 31% in 2018 to 24% in 2021, con- cerns persist regarding non-food items, as similar discrepancies have been ide... | https://arxiv.org/abs/2505.19254v1 |
more than half of the reported problems concern probable counterfeit products, differences dependent on the place of pur- chase within the same market, quality deterioration over time, mismatches between received products and orders, misleading information, suspicions of fraud, and variations related to packaging, batc... | https://arxiv.org/abs/2505.19254v1 |
Source Process DatasetFinal DQ (PL)Final Test (Multi)NStep NFigure 4: Diagram showing the process of preparing DQ and multilingual datasets. Original review text Translated review text Label Additional Comment Fantastyczny zapach i produkt z chemii niemieckiej, wi˛ ec o wiele bardziej intensywny ni˙z te, produkowane na... | https://arxiv.org/abs/2505.19254v1 |
FacebookAI/xlm-roberta-large herbert-base-cased allegro/herbert-base-cased herbert-large-cased allegro/herbert-large-cased polish-roberta-base-v2 sdadas/polish-roberta-base-v2 polish-roberta-large-v2 sdadas/polish-roberta-large-v2 deepseek-v3* deepseek-ai/DeepSeek-V3 gpt-4o* - Table 8: Model names as referenced in the ... | https://arxiv.org/abs/2505.19254v1 |
ace gł˛ ebszej analizy np. pogorszenie jako ´sci z upływem czasu; praktyki niezgodne z prawem i/lub naruszaj ˛ ace prawa klienta np. produkt jest prawdopodobnie podrobiony, podejrzenie oszustwa, wprowadzanie klienta w bł ˛ ad, brak instrukcji w wymaganym j˛ ezyku, brak daty wa ˙zno´sci itp. "standard" – standardowa opi... | https://arxiv.org/abs/2505.19254v1 |
only the name of the class, without additional comment. Review text: <review> few-shot Assign the following review to one of three classes: “dual quality”, “other problems” or “standard”. Examples: The capsules are better than those on the Polish market from the same company. – dual quality Good coffee taste. Country o... | https://arxiv.org/abs/2505.19254v1 |
the Polish market from the same company.", "Good coffee taste. Country of origin: Germany. It is not as acidic as the one bought in the country." "other problems" – The review does not indicate an issue of dual quality but provides information on other problems, which can include: differences in products resulting from... | https://arxiv.org/abs/2505.19254v1 |
Next Token Prediction Is a Dead End for Creativity: Why It’s Impossible to Lose Yourself in the Moment Ìbùkún Ọlátúnjí¹ Mark Sheppard²* ¹Computational Foundry, Swansea University, Crymlyn Burrows, Skewen, Swansea SA10 6JW, UK ²University of Kent, Canterbury, Kent CT2 7NZ, U Abstract This position paper argues that toke... | https://arxiv.org/abs/2505.19277v1 |
standards that reflect the dialogic, embodied, and adversarial nature of language in the wild. _____________ Corresponding authors: 2030349@swansea.ac.uk, ms2403@kent.ac.uk Table1 compares the functional capabilities of next-token prediction models with the demands of improvisational co-creation in contexts like rap ba... | https://arxiv.org/abs/2505.19277v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.