text string | source string |
|---|---|
Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. 2016. Analysing how people orient to and spread rumours in social media by looking at conversational threads. PloS one, 11(3):e0150989. A Additional Information on Data Collection We provide URLs of each fact-checking source used during data collection in ... | https://arxiv.org/abs/2505.18916v1 |
annotation (I’m pretty sure about the annotation, but might be in high chance other annotators may label it in a different category) •2 - not confident about the annotation (I’m not sure about the annotation, it seems it also belongs to other categories, but you can still include this instance as a “silver standard in-... | https://arxiv.org/abs/2505.18916v1 |
generated text exhibits a greater linguistic vari- ety, while lower-resource languages such as Hindi and Czech exhibit less variety and thus have higher values of VOR and lower values of JSD. The results of fine-tuning and adaptive self- training using different values of fine-tuning data 13 Fine-Tuning Data t GMR wF2 ... | https://arxiv.org/abs/2505.18916v1 |
0.430 0.550 0.269 0.500 0.413 0.478 Comm 0.499 0.416 0.525 0.502 0.457 0.441 0.542 0.571 0.672 Macro 0.515 0.481 0.531 0.545 0.544 0.447 0.514 0.525 0.534 Deepseek-R1 Conf 0.307 0.293 0.418 0.424 0.472 0.494 0.289 0.424 0.141 Rej 0.381 0.315 0.499 0.394 0.489 0.286 0.312 0.428 0.415 Ques 0.370 0.473 0.542 0.437 0.388 0... | https://arxiv.org/abs/2505.18916v1 |
Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments 1stAmel Muminovic Faculty of Engineering International Balkan University Skopje, North Macedonia amel.muminovic@ibu.edu.mk Abstract —As online platforms grow, comment sections increas- ingly host harassment that undermin... | https://arxiv.org/abs/2505.18927v2 |
human capacity [10]. Yet despite advances in natural language processing, enforcement remains inconsis- tent. Abusive comments that rely on sarcasm, coded language, or emotional manipulation often slip through AI moderation, even when they clearly violate platform policies. These sub- tleties continue to pose challenge... | https://arxiv.org/abs/2505.18927v2 |
address this gap, but progress remains limited [21]. Recent research also suggests that models often over-flag emotionally intense or sensitive language even when used in supportive contexts, leading to user frustration and moderation fatigue. This tension between under-flagging subtle harm andover-flagging emotional e... | https://arxiv.org/abs/2505.18927v2 |
whose pilot abuse rate exceeded 20%. From the four retained videos we then drew a uniform random sample, yielding 5 080 comments in total. Only public endpoints were accessed; private, removed, or shadow-banned comments are not included. B. Data Cleaning and Anonymization Each comment underwent a two–step preprocessing... | https://arxiv.org/abs/2505.18927v2 |
harmful comment From this, the following performance metrics were calcu- lated: Precision measures the proportion of comments that the model correctly identified as harmful out of all the comments it flagged as harmful. In other words, it reflects how accurate the model is when it predicts that a comment is harmful. It... | https://arxiv.org/abs/2505.18927v2 |
number of harmful comments, with 1 167 true positives, but also had the highest number of false positives, mistakenly labeling 354 non-harmful com- ments as harmful. GPT had a solid performance overall, producing 1 122 true positives and only 143 false positives. Claude was more selective, correctly identifying 961 har... | https://arxiv.org/abs/2505.18927v2 |
These false positives were grouped by model agreement patterns to highlight shared tendencies and specific weaknesses. While most flagged content reflected strong language or emotional tone, it often lacked any abusive intent. Below are four common patterns observed. 1) All Three Models vs. Human: Shared Over-flagging ... | https://arxiv.org/abs/2505.18927v2 |
illness framed as jokes. 2) GPT: Missed Hostility Behind Humor: GPT missed a substantial number of harmful comments flagged by human reviewers due to sarcasm, coded ridicule, or indirect hostility. Phrases like “every time I see you in the hospital I smile” conveyed sustained mockery and dismissiveness toward some- one... | https://arxiv.org/abs/2505.18927v2 |
than harassment. Comments referencing public figures or using layered sar- casm were also commonly misclassified. For instance, “It sounds like UserNameProtected Cobain’s remembrance video” was flagged by GPT, Gemini, and Claude, though human reviewers judged it to be a stylistic comparison rather than an attack. Anoth... | https://arxiv.org/abs/2505.18927v2 |
to frequent false positives. Claude achieved the highest precision 0.920 and the lowest false positive rate 0.022 yet its recall slipped to 0.720. GPT delivered the best overall balance with an F1 score 0.863, precision 0.887, and recall 0.841. Qualitative analysis showed that sarcasm, coded insults, and mixed-language... | https://arxiv.org/abs/2505.18927v2 |
The effect of anonymity on cyberbullying frequency,” Psychol. Pop. Media Cult. , vol. 4, no. 2, pp. 70–79, 2015. doi:10.1037/a0034335 [5] L. Huang et al., “The severity of cyberbullying affects bystander inter- vention among college students: The roles of feelings of responsibility and empathy,” Psychol. Res. Behav. Ma... | https://arxiv.org/abs/2505.18927v2 |
Eds., Online, Nov. 2020, pp. 3356–3369. doi:10.18653/v1/2020.findings-emnlp.301 [20] A. Arora, “Sarcasm detection in social media: A review,” in Proc. Int. Conf. Innov. Comput. Commun. (ICICC), Dec. 2020, pp. 1–4. doi:10.2139/ssrn.3749018. [21] M. S. Jahan and M. Oussalah, “A systematic review of hate speech automatic ... | https://arxiv.org/abs/2505.18927v2 |
https://arxiv.org/abs/2106.10328 [37] T. Bolukbasi, K.-W. Chang, J. Zou, V . Saligrama, and A. Kalai, “Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,” in Proc. NeurIPS , Barcelona, Spain, 2016, pp. 4356–4364. doi: 10.48550/arXiv.1607.06520. [38] H. Welbl, A. Stiennon, and Y . Bai, “C... | https://arxiv.org/abs/2505.18927v2 |
arXiv:2505.18929v1 [cs.AI] 25 May 2025Meta-aware Learning in text-to-SQL Large Language Model* Wenda Zhang1 Abstract — The advancements of Large language models (LLMs) have provided great opportunities to text-to-SQL tasks to overcome the main challenges to understand com- plex domain information and complex database s... | https://arxiv.org/abs/2505.18929v1 |
devices on open-source LLMs [1]. These issues can impede LLMs from effectively incorporating the essential business domain-specific information and exacerbate issues related to hallucinations [11]. Fine-tuning an open-source LLM for specific business environments on text-to-SQL tasks is a commonly adopted solution [1],... | https://arxiv.org/abs/2505.18929v1 |
[25], and hallucination [11], where models generate inaccurate or unrelated SQL queries to the provided schemas and contexts. To address these issues, researchers have explored strate- gies in LLM fine-tuning for text-to-SQL applications, includ- ing parameter-efficient fine-tuning (PEFT) methods [3], [27], domain-spec... | https://arxiv.org/abs/2505.18929v1 |
interact with the data structures effectively [30]. Prompt framework tokenization refers to the structuring of prompts with tokenized tags, which serve to clarify the task requirements and enhance the model’s comprehension of the specific objectives, facilitating task-specific understanding and performance. The propose... | https://arxiv.org/abs/2505.18929v1 |
scenario I, we aimed to assess the per- formance of the prompt structure tokenization method, where we provided a list of structure tags, such as <system> /<instruction> /<question> /<answer> , in the tokenizer to structure the prompt. To control the effect of table variation, the dataset was generated from a single do... | https://arxiv.org/abs/2505.18929v1 |
and domain knowledge enhancement learning, enabling the model to enhance its understanding of the data from multiple perspectives, thereby strengthening its metadata knowl- edge. •Schema-CoT-OPT : This approach combines schema- based learning and Chain-of-Thought learning with key information tokenization, which tokeni... | https://arxiv.org/abs/2505.18929v1 |
average, 9.9% longer training time than the base prompt, this increase is due to the prompt structure tokenization step, which expands the embedding size and thus increases computa- tional complexity. Despite this added cost, the tokenized prompt method offers clear advantages: it consistently de- layed overfitting com... | https://arxiv.org/abs/2505.18929v1 |
<0.1 <0.1 in all testing cases lower than 0.123. Progressive learning [39] is a strategy for sequentially learning tasks while retaining knowledge from prior tasks in multi-task learning. However, progressive learning can be prone to catastrophic forgetting [39], [40], where knowledge from previous tasks is lost as wei... | https://arxiv.org/abs/2505.18929v1 |
in this material are those of the author(s) and do not necessarily reflect the views of the funding parties. DATA AVAILABILITY STATEMENT Due to Walmart’s Privacy Requirements, models and datasets are not open to public. REFERENCES [1] Z. Hong, Z. Yuan, Q. Zhang, H. Chen, J. Dong, F. Huang, and X. Huang, “Next-generatio... | https://arxiv.org/abs/2505.18929v1 |
, vol. 37, no. 11, 2023, pp. 13 067–13 075. [16] K. O’Shea, “An introduction to convolutional neural networks,” arXiv preprint arXiv:1511.08458 , 2015.[17] Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, T. Liu et al. , “A survey on in-context learning,” arXiv preprint arXiv:2301.00234 , 2022. [18... | https://arxiv.org/abs/2505.18929v1 |
arXiv preprint arXiv:1806.03852 , 2018. [34] W. Cui and Q. Wang, “Ada-instruct: Adapting instruction generators for complex reasoning,” arXiv preprint arXiv:2310.04484 , 2023. [35] T. Yu, R. Zhang, K. Yang, M. Yasunaga, D. Wang, Z. Li, J. Ma, I. Li, Q. Yao, S. Roman et al. , “Spider: A large-scale human-labeled dataset... | https://arxiv.org/abs/2505.18929v1 |
arXiv:2505.18931v1 [cs.AI] 25 May 2025Can Large Language Models Infer Causal Relationships from Real-World Text? Ryan Saklad1, Aman Chadha*2, Oleg Pavlov1, and Raha Moraffah1 1Worcester Polytechnic Institute 2Amazon Gen AI Abstract Understanding and inferring causal relation- ships from texts is a core aspect of human ... | https://arxiv.org/abs/2505.18931v1 |
text (Veldhuis et al., 2024; Hosseinichimeh et al., 2024; Oh, 2025; Jin et al., 2024; Joshi et al., 2024a; Lasheras and Pinheiro, 2025), they often use synthetically generated or simplified texts with explicitly stated causal links. This approach, however, falls short of real-world scenarios where causal relationships ... | https://arxiv.org/abs/2505.18931v1 |
requiring the synthesis of a complete causal graph. These approaches, detailed further in Appendix A, generally do not address the challenge of causal graph construction under the full diversity and complexities of real-world conditions, where causality is often implicit and embedded within long narratives. ReCAST diff... | https://arxiv.org/abs/2505.18931v1 |
location of each primary causal graph in the paper is labeled to ensure accurate annotation (e.g., “Top diagram on page 7”). 3.2 Annotation As ReCAST is a text-based benchmark, it is impor- tant that ground-truth causal graphs are converted from images to a text-based representation to be used as the ground-truth answe... | https://arxiv.org/abs/2505.18931v1 |
save on computational costs). Therefore, Mistral is tasked to output the markdown auto- regressively while skipping over non-textual ele- ments (such as images, charts, or other figures), in-line citations, references, publication informa- tion, and appendices. The output of this step is a well-formatted markdown versi... | https://arxiv.org/abs/2505.18931v1 |
include these failed answers to avoid artificially depressing scores. We also provide the LLM the number of expected nodes, as fixing Visolates the causal graph’s abstraction from the causal reason- ing objective. Without doing so, there are many valid levels of granularity for the causal graph, greatly complicating au... | https://arxiv.org/abs/2505.18931v1 |
hallucinations, reversing causality, and other factors to be appropriately in- corporated into overall metrics while minimizing ambiguities in grading. The LLM judge, R1, is pro- vided with the source text and ground-truth graph, and outputs a categorical assessment across several criteria for each node and edge. These... | https://arxiv.org/abs/2505.18931v1 |
of graphs. However, normalized SHD is low, due to most graphs being sparse. We also find that there is a relationship be- tween model size and performance. The worst- performing model, Llama-8B , is just 8 billion pa- rameters, while the best performing model is R1, with 685 billion. There is also a relationship be- tw... | https://arxiv.org/abs/2505.18931v1 |
F1of 0.57, yet this reduces by almost half to 0.31 for samples where more than half of the nodes are confounding. This trend shows that while models struggle for all degrees of confounding, they espe- cially struggle when they must use causal reasoning to infer information from the text. This finding is corroborated by... | https://arxiv.org/abs/2505.18931v1 |
tion highlights that the poor performance of LLMs on ReCAST is due to fundamental limitations in their causal reasoning capabilities rather than being limited to errors in surface-level entity recognition. 5.5 Case Study - R1 Model Answer To illustrate model performance on realistic causal reasoning, we show a benchmar... | https://arxiv.org/abs/2505.18931v1 |
under real- world conditions has been limited by the lack of appropriate benchmarks. This paper introduced ReCAST, which is, to our knowledge, the first benchmark to assess LLM causal reasoning capa- bilities from text under realistic conditions. Re- CAST draws diverse samples from academic lit- erature, featuring text... | https://arxiv.org/abs/2505.18931v1 |
sourced from economics students par- ticipating for opt-in extra credit. No personallyidentifiable information (PII) besides name was collected (to assign credit), which has since been destroyed, guaranteeing complete anonymization. Participants were explicitly informed that their anonymized annotation data would be us... | https://arxiv.org/abs/2505.18931v1 |
exogenous galore. Notion. Available at https://sangmino.notion. site/1a897b8106ca44eeaf31dcd5ae5a61b1?v= ff7dc75862c6427eb4243e91836e077e . OpenAI. 2025. Openai o3-mini system card. Ope- nAI, January 31, 2025. https://cdn.openai.com/ o3-mini-system-card.pdf . Judea Pearl. 2009. Causality . Cambridge University Press, C... | https://arxiv.org/abs/2505.18931v1 |
al., 2025) examines how LLMs un- derstand explicit connectives in sentence pairs, re- sulting in 2-node links. By design, it uses very short, often crafted inputs and focuses on explicit cues, thereby avoiding the complexities of implicit causality and information integration from exten- sive texts that ReCAST targets.... | https://arxiv.org/abs/2505.18931v1 |
Document 2 ✓ /exclamati⌢n-triangle ✓ From text to map (Hosseinichimeh et al., 2024) Graph Construction ✗ Short Narratives 15 ✗ ✗ ✓ Failure Modes (Yamin et al., 2024) ID/Graph Construction ✗ Short Narrative 20 ✗ /exclamati⌢n-triangle /exclamati⌢n-triangle From Text to Model (Veldhuis et al., 2024) Sentence Classificatio... | https://arxiv.org/abs/2505.18931v1 |
ensure suf- ficient context to classify while minimizing total tokens. We use the following prompt: Domain Classification Prompt You are an expert at correctly labeling domains. You will be given a published paper’s title and abstract. You will label each other domain of the paper based on the content. You may pick mor... | https://arxiv.org/abs/2505.18931v1 |
(Team, 2025b) narrowly outperforms it. However, all mod- els perform poorly overall regardless of domain, showing that this is not the factor that leads to poor causal reasoning. Figure 7 visualizes how often each domain ap- pears in the benchmark samples. Engineering & Technology and Environmental & Earth Sciences com... | https://arxiv.org/abs/2505.18931v1 |
to a unified eco- nomic outcome, total household income , re- flecting a broader—but valid—abstraction of the two-hop causal path present in the ground-truth graph (Total grain output →County GDP ). This demonstrates that the task is tractable for hu- mans, and the importance of evaluation that dif- ferentiates stylist... | https://arxiv.org/abs/2505.18931v1 |
capabilities of the graph embed- ding model for this task, as substantial information being provided to models has little effect on the final embedding score. F Alternative Measure of Degree of Confounding As shown in Figure 2, degree of confounding has a noticeable effect on model performance. For this, we determine w... | https://arxiv.org/abs/2505.18931v1 |
being outputted. Another failure mode was nonsensical generations, such as generating Chinese despite the text and in- structions being in English. We show this example below. Foreign-Language Base Model Output <think>建立一个因果关系图需要遵循以下步 骤： 1.确定因果关系：首先需要确定因果关系，即 哪些变量是因，哪些变量是果，以及它们之间 的关系是什么。在本例中，我们已经确定了一 些变量之间的因果关系，例如生产、生活... | https://arxiv.org/abs/2505.18931v1 |
or re-evaluations, benefit greatly from caching the expensive text embedding, mak- ing the iterative evaluation process highly econom- ical. This efficient design ensures that ReCAST can be utilized and extended by researchers with- out imposing prohibitive computational or financial burdens. J Inter-Annotator Agreemen... | https://arxiv.org/abs/2505.18931v1 |
us to treat papers published in 2024 and onwards as more likely “unseen” by these models.Our dataset contains 35 samples derived from papers published in 2024 or later. For Llama-8B , 229 samples were from papers published before 2024, and for o3-mini , 235 samples were from papers published before 2024. We compare the... | https://arxiv.org/abs/2505.18931v1 |
24 0 0 0 0 630 25 32 0 0 0 2 617 27 23 0 0 0 0 588 10 16 0 0 0 0 574 19 32 0 0 0 0 566 24 23 1 0 0 1 558 21 24 0 0 0 0 552 16 15 0 0 0 0 536 37 23 0 0 0 0 497 15 20 0 0 1 0 491 12 21 0 0 0 1 486 28 32 0 0 0 0 481 20 28 0 0 0 0 458 18 24 0 0 0 0 449 43 88 0 0 0 0 440 12 15 1 0 0 0 435 26 27 0 0 2 5 410 9 16 0 0 0 0 393 ... | https://arxiv.org/abs/2505.18931v1 |
samples have confounding levels. We detail the prompt used for this below. These node-level labels are used as the basis for calculation of degree of confounding. Label Unobserved Confounders Prompt You will be given a causal graph in economics and a source text. Your task is to label each node in the graph to determin... | https://arxiv.org/abs/2505.18931v1 |
combined into "Number of dogs". - "number of dogs" and "Number of dogs" should be combined into "Number of dogs". Negative examples (do not combine): - "<variable>" and "variable" should not be combined since it is clear that they are intended to be distinct. - NEVER change any variables with < or > in the name. - "Num... | https://arxiv.org/abs/2505.18931v1 |
the ReCAST benchmark, it is important to remove any explicit references to the causal graph, which make the task trivial. During this step, we also correct any references to non-existent elements which were removed in previous pre-processing steps (for ex- ample, referencing an image). We utilize a normal- ization tool... | https://arxiv.org/abs/2505.18931v1 |
times, use a slightly longer start_string and include some of the original text in your replacement to maintain context. - Do not "redact" the text; remove references entirely rather than replacing them with generic text. - Both the start and end strings will be included in the text that gets replaced. Changes are appl... | https://arxiv.org/abs/2505.18931v1 |
text and the name of each node in the graph. Ensure that each node is included at least once in the generated causal graph. Do not use the node’s name in the graph; instead, use the id corresponding to the node. For the example nodes below (not the same as the ones you will be provided), whenever you want to include th... | https://arxiv.org/abs/2505.18931v1 |
it seems tedious, redundant, or unnecessary. Do this for each node or edge you are evaluating; there is no time limit. Be sure to fully to fully think through each node or edge you are tasked with evaluating fully before moving onto the next one. i. It is helpful to quote supporting evidence from the provided texts and... | https://arxiv.org/abs/2505.18931v1 |
interpretation needed; meaning partially preserved - SEMANTIC_NA: Not applicable Abstraction Labels (select one): - ABSTRACTION_BROADER: Represents a more general concept that includes text concepts - ABSTRACTION_ALIGNED: Represents approximately the same scope and specificity as the text - ABSTRACTION_NARROWER: Repres... | https://arxiv.org/abs/2505.18931v1 |
in graph) Inference Labels (select one): - INFERENCE_DIRECT: Relationship matches text’s explicit causal claims - INFERENCE_DERIVED: Relationship logically follows from text - INFERENCE_STRETCHED: Relationship possible but weakly supported - INFERENCE_NA: Not applicable or relationship does not exist Abstraction Labels... | https://arxiv.org/abs/2505.18931v1 |
ground-truth graph) is computed by averaging the numerical scores of its constituent labels. Precision metrics (node precision, edge preci- sion) for each item generated by the LLM are de- termined by comparing it against both the ground- truth graph and the source text. If the item is la- beled asPRESENCE_NO_MATCH aga... | https://arxiv.org/abs/2505.18931v1 |
arXiv:2505.18933v1 [cs.AI] 25 May 2025REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing Haitian Zhong1, Yuhuan Liu2, Ziyang Xu3, Guofan Liu1,4 Qiang Liu1,Shu Wu1,Zhe Zhao4,Liang Wang1,Tieniu Tan1 1NLPR, MAIS, Institute of Automation, Chinese Academy of Sciences 2C... | https://arxiv.org/abs/2505.18933v1 |
” is cor- rected to “Luka Doncic plays in the NBA team of Lakers .” In an overfit scenario, when queried with “Who does Luka Doncic play with?”, the model may still disproportionately favor the edit target but not the correct answer—assigning a high prob- ability to “Mavericks”—while the probabilities for 1 more contex... | https://arxiv.org/abs/2505.18933v1 |
the objective is to update this to (s=Luka Doncic, r=plays in the NBA team of, o∗=Lakers ). Such an editing operation is denoted by e= (s, r, o, o∗). Given a model f and an edit e, we define the editing operator as K(f, e) =f∗, where f∗represents the model after applying the edit. Unlike conventional approaches that mo... | https://arxiv.org/abs/2505.18933v1 |
+,i,h(l) −,i)}N i=1 for each layer l. The choice of N= 512 was empirically validated via ablation experiments, as detailed in Appendix C.1. Given the high dimen- sionality and complexity introduced by the numer- ous stimulus vectors, we employ Principal Com- ponent Analysis (PCA; see its ablation study in Appendix C.2)... | https://arxiv.org/abs/2505.18933v1 |
To enforce this, we introduce a regularization term that minimizes the divergence between the output distributions of the edited model f∗and the orig- inal model fover a dataset of unrelated prompts. Formally, we define the local consistency loss as: Lloc= E (p′,x)∼Dloch DKL Pf∗(x|p′) Pf(x|p′)i (4) where p′denotes a ... | https://arxiv.org/abs/2505.18933v1 |
of diverse knowledge domains. Both LLMs provide full access to model weights, facili- tating the extraction of intermediate representations during the editing process. 4.2 Knowledge Editing Baselines Our method is compared against several established knowledge editing techniques: Fine-Tuning (FT) FT updates model param... | https://arxiv.org/abs/2505.18933v1 |
unintended generalization. We next introduce the key probability-based met- rics used to quantify overfitting. In an overfitting evaluation, a prompt does not necessarily retrievethe original object, since not all prompts explicitly invoke the subject-relation pair. Correct Answer Probability (CAP) measures the probabi... | https://arxiv.org/abs/2505.18933v1 |
Ro- bust Knowledge Editing. In addition to reliabil- ity, locality, and generality, our approach achieves notably high portability scores. Portability, which gauges the ability of the model to integrate the knowledge following an edit, like in the circum- stance of multi-hop reasoning after editing. Com- pared to basel... | https://arxiv.org/abs/2505.18933v1 |
underscores the growing need for controllability and adaptability in modern LLMs, ensuring that their responses remain accurate and up-to-date without extensive retraining. Representation Engineering Representation En- gineering (Zou et al., 2023) is derived as a novel approach that shifts the focus from neurons and ci... | https://arxiv.org/abs/2505.18933v1 |
Mor Geva. 2024. Evaluating the ripple effects of knowledge editing in language models. Transac- tions of the Association for Computational Linguis- tics, 12:283–298. Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Edit- ing factual knowledge in language models. In Pro- ceedings of the 2021 Conference on Empirical Met... | https://arxiv.org/abs/2505.18933v1 |
Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2023. Editing large language models: Prob- lems, methods, and opportunities. arXiv preprint arXiv:2305.13172 . Mengqi Zhang, Xiaotian Ye, Qiang Liu, Pengjie Ren, Shu Wu, and Zhumin Chen. 2024a. Uncovering over- fitting in large language model editing. Preprint , arXiv:241... | https://arxiv.org/abs/2505.18933v1 |
(s, r, o, o∗): Mrel=E e∼D edit1n arg max o{Pf∗(o|p(s, r)) =o∗}o 10 Generality Mgenevaluates the model’s capacity to apply the edit correctly to in-scope data, ensuring that the model maintains generalization capabili- ties: Mgen=E e∼D edit p∗∼N (e)1n arg max o{Pf∗(o|p∗(s, r)) =o∗}o where the N(e)stands for the rephrase... | https://arxiv.org/abs/2505.18933v1 |
city of p(s, r′) Houston is located in p(s, r, o′;sneighbor , rneighbor )What is the twin city of Houston? It is Prague. Regensburg is a twin city of Table 3: Notations and their meanings. Details of evaluation metrics The key probability-based metrics used to quantify the effectiveness of Overfit editing tasks for a g... | https://arxiv.org/abs/2505.18933v1 |
to update the fact that Apple A5 was created by Google . This is absolutely true in the following context. Given this es- tablished fact, please tell me: Apple A5 was created by Unprompted Input (generic factual comple- tion): Apple A5 was created by C Ablation Studies C.1 Ablation Study on the Number of Stimulus Vecto... | https://arxiv.org/abs/2505.18933v1 |
for Qwen-2.5 and 1.04B for Llama3.1. 13 D.1.1 REACT Parameters Llama3.1 Qwen2.5 Iters 20000 20000 Edit Layerall layer of all layer of Transformer Module Transformer Module Optimizer Adam Adam Learning Rate 1e−5 1e−5 cedit 1 1 cloc 0.1 0.1 cedit,cls 1 1 cloc,cls 0.1 0.1 D.1.2 FT Parameters Llama3.1 Qwen2.5 Max Steps 25 ... | https://arxiv.org/abs/2505.18933v1 |
arXiv:2505.18942v2 [cs.CY] 27 May 2025Language Models Surface the Unwritten Code of Science and Society Honglin Bao1,2, Siyang Wu1,2∗, Jiwoong Choi1∗, Yingrong Mao1∗, James A. Evans1,2,3 1Knowledge Lab 2Data Science Institute 3Department of Sociology University of Chicago honglinbao@uchicago.edu; jevans@uchicago.edu Ab... | https://arxiv.org/abs/2505.18942v2 |
biases and heuristics is now well-established in literature, framing these models as inherently cultural and social technologies [ 13,17,65,49,39]. This paper moves beyond merely acknowledging LLM biases; it proposes leveraging them as diagnostic tools to uncover and critically examine the implicit, unwritten societal ... | https://arxiv.org/abs/2505.18942v2 |
the cognitive and social aspects of scientific reasoning and judgment. Data : We collected data from the OpenReview API and Paper Copilot3, a website that aggregates peer review scores for top computer science conference submissions. Our final dataset includes metadata, peer review scores, and review comments for 26,73... | https://arxiv.org/abs/2505.18942v2 |
and comparing the papers’ extended abstracts, and then generating five hypotheses to explain why the better scoring paper may appear stronger than the other. Self-consistent Refinement : We apply the five generated hypotheses to all pairs in the dataset. For each pair and hypothesis, we ask the LLM to determine the deg... | https://arxiv.org/abs/2505.18942v2 |
judgments can often be explained by a compact set of generalizable hypotheses, reflecting recurring evaluative patterns across diverse cases. We will discuss the generated hypotheses in the result section. Prior and Posterior : From a Bayesian perspective, querying an LLM with a general question “What do you think make... | https://arxiv.org/abs/2505.18942v2 |
each pair multiple rounds would lead to prohibitive computational and monetary costs. For demonstration purposes in our case study, we randomly selected 5,000 pairs for the following experiments. Our goal is to showcase the effectiveness of our conceptual framework and to inspire future research. 4 Results 4.1 Prior an... | https://arxiv.org/abs/2505.18942v2 |
storytelling quality and interdisciplinary relevance. When treated as a peer reviewer, the LLM initially aligns with conventional scientific ideals. In practice, however, it leans more heavily on accessible, narrative-driven criteria that appear more explanatory in distinguishing lower from higher rated papers (Figure ... | https://arxiv.org/abs/2505.18942v2 |
as 68% of its prior and 50% of human mentions. LLMs, as expected, initially even exaggerate stereotyped evaluation norms — overweighting "prior" and underweighting "posterior" standards — more sharply than humans do. Humans indeed explicitly reward internal qualities, but their stated reward patterns explain very littl... | https://arxiv.org/abs/2505.18942v2 |
studies and generate hypotheses about why some are amplified while others are overlooked or even misreported. It is important to emphasize that our goal is not to advocate for the automation of judgment tasks by LLMs. Rather, we propose using LLMs as investigative tools to replicate and expose the biases and heuristics... | https://arxiv.org/abs/2505.18942v2 |
criteria, they often play a critical role in actual decision-making. If the evaluation processes really did, as some claim, cleave close to explicit review criteria that focus on intrinsic qualities of a research study, then heuristics, biases, and instances of consistent discrimination would not systematically occur. ... | https://arxiv.org/abs/2505.18942v2 |
A stricter experiment building on the paper would involve explicitly instructing the LLM that it is not appropriate to follow a certain norm, while observing whether it still privately engages in the behavior. In sum, we encourage the research community to build upon our proposed framework to explore these "unwritten c... | https://arxiv.org/abs/2505.18942v2 |
of variance in human visually evoked affect. Proceedings of the National Academy of Sciences , 122(4):e2306025121, 2025. [16] Paula Czarnowska, Yogarshi Vyas, and Kashif Shah. Quantifying social biases in nlp: A generalization and empirical comparison of extrinsic fairness metrics. Transactions of the Association for C... | https://arxiv.org/abs/2505.18942v2 |
actionable survey. arXiv preprint arXiv:2210.07700 , 2022. [34] Bruno Latour. Science in action: How to follow scientists and engineers through society . Harvard University Press, 1987. [35] Carole J Lee, Cassidy R Sugimoto, Guo Zhang, and Blaise Cronin. Bias in peer review. Journal of the American Society for Informat... | https://arxiv.org/abs/2505.18942v2 |
Esther Vidal, Salvatore Ruggieri, Franco Turini, Symeon Papadopoulos, Emmanouil Krasanakis, et al. Bias in data-driven artificial intelligence systems—an introductory survey. Wiley Interdis- ciplinary Reviews: Data Mining and Knowledge Discovery , 10(3):e1356, 2020. [51] Uwe Peters and Benjamin Chin-Yee. Generalization... | https://arxiv.org/abs/2505.18942v2 |
Bekiranov, and Aidong Zhang. Improving scientific hypothesis generation with knowledge grounded large language models. arXiv preprint arXiv:2411.02382 , 2024. [67] Zonglin Yang, Xinya Du, Junxian Li, Jie Zheng, Soujanya Poria, and Erik Cambria. Large language models for automated open-domain scientific hypotheses disco... | https://arxiv.org/abs/2505.18942v2 |
strategies. Implicit Bias in LLMs : Concerns about bias in AI systems have long focused on the ways these technologies can absorb and amplify harmful content from large-scale human-generated datasets – particularly prejudices related to race, gender, and other social dimensions [ 50]. Large language models (LLMs), trai... | https://arxiv.org/abs/2505.18942v2 |
implicit heuristics and tacit values that often shape peer review decisions but remain unspoken. By prompting LLMs to hypothesize why one paper is rated more highly than another, we reveal not only individual reviewer preferences but also broader cultural norms and evaluative codes that govern scientific judgment. In d... | https://arxiv.org/abs/2505.18942v2 |
such as the Holder and Besov spaces, we show that, by considering anisotropic smoothness, they can alleviate exponential dependency on the dimensionality but they only depend on the smoothness of the target functions. Our theoretical analysis supports the great practical success of convolutional networks. Furthermore, ... | https://arxiv.org/abs/2505.18942v2 |
Crucially, this smoothness is allowed to vary across input coordinates: Some directions may be very "rough" (i.e., important to capture precisely), while others can be ignored or treated coarsely. This idea mirrors how images contain both low-frequency (broad structure) and high-frequency (detail) components, and not a... | https://arxiv.org/abs/2505.18942v2 |
contribution by offering comprehensive analyses of deep learning’s learnability in infinite-dimensional spaces. Its results: Support the practical success of CNNs in real-world tasks involving high-dimensional data. Justify the use of dilated convolutions as an effective tool for handling sparse or long-range dependenc... | https://arxiv.org/abs/2505.18942v2 |
0.20 13 0.67 one paper is not theoretically rigorous 1 0.92 0.18 20 0.57 one paper’s design is over-engineered, not elegant, and unnecessarily com- plicated1 0.50 0.02 18 0.74 one paper does not use a fair bench- mark for evaluation1 0.32 0.08 15 0.16 one paper lacks implementation and reproducibility details1 0.44 0.1... | https://arxiv.org/abs/2505.18942v2 |
they assign within the conference. The results are shown below. Table 2: Regression Results Hypothesis Coef. Std. Err. t P >|t|[0.025, 0.975] const 5.966 0.100 59.508 0.000 [5.769, 6.163] one paper lacks justification regarding its novelty0.231 0.096 2.410 0.016 [0.043, 0.420] one paper is not theoretically rigorous 0.... | https://arxiv.org/abs/2505.18942v2 |
bias is most pronounced when the quality gap is small [ 40,56]. Second, we randomize the positions of Paper 1 and Paper 2 across voting, while ensuring logically consistent prompts and vote aggregation. We then assess whether position bias is indeed stronger when the quality signal gap is small. Specifically, we regres... | https://arxiv.org/abs/2505.18942v2 |
arXiv:2505.18943v1 [cs.CL] 25 May 2025MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems Xuanming Zhang1, Yuxuan Chen2, Min-Hsuan Yeh1, Yixuan Li1∗ ∗ 1Uniersity of Wisconsin-Madison 2Tsinghua University xzhang2846@wisc.edu, sharonli@cs.wisc.edu Abstract Human social interactions depend on t... | https://arxiv.org/abs/2505.18943v1 |
bedtime schedule. Nana calmly accepts the responsibility of discretion, appreciating the consensus she's reached without displaying distress or irritation. Generation & ValidationFinal Response Figure 1: MetaMind multi-agent framework. The architecture comprises three collaborative agents—Theory- of-Mind Agent, Domain ... | https://arxiv.org/abs/2505.18943v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.