text string | source string |
|---|---|
child go 0.034 Many children day 0.025 DivorcedNo children marry 0.037 One child marry 0.030 Many children marry 0.039 WidowedNo children work 0.031 One child die 0.035 Many children house 0.044 across both religions frequently associate with woman andhouse ,reinforcing domestic roles. •Divorced Individuals : All ident... | https://arxiv.org/abs/2505.18466v1 |
association with engagement in various pursuits. The pres- ence of father (0.087) for married men with children reinforces parental roles. In con- trast, the term sport (0.059–0.083) appears for divorced men with many children, potentially reflecting gendered leisure interests. •Muslim Female : The term woman is over- ... | https://arxiv.org/abs/2505.18466v1 |
ties. •Divorced : The term activity appears across all identities except for Muslim women identities,indicating continued engagement in hob- bies or responsibilities post-divorce. Mus- lim women that are divorced with many chil- dren are strongly associated with woman , in- dicating connection to gender identity. •Wido... | https://arxiv.org/abs/2505.18466v1 |
bias lexicon. G.38.1 Analysis of Gender and Religion •Hindu Female : Hindu females overwhelm- ingly feature terms related to housework and household , particularly in married and multi- ple children categories (TF-IDF up to 0.134). The recurrence of housework andhousehold across all marital statuses suggests strong gen... | https://arxiv.org/abs/2505.18466v1 |
Muslim widowed females emphasize housework . Divorced and widowed males have chore andtaskappearing often, showing a continued focus on professional and house- hold tasks. G.38.3 Number of Children-Based Analysis •No Children : Hindu and Muslim females without children primarily feature daily and housework , suggesting... | https://arxiv.org/abs/2505.18466v1 |
still feature domestic work, while males shift to- wards advisory or financial tasks. •Number of Children: Number of children often amplifies term frequencies. Women with many children have the highest TF-IDF values for household ,housework , and clean , reinforcing extreme domestic labor expecta- tions. Hindu males wi... | https://arxiv.org/abs/2505.18466v1 |
home 0.013 One child 0.094 house 0.023 Many children 0.058 family 0.012 WidowedNo children 0.034 home 0.009 One child 0.045 home 0.015 Many children 0.089 house 0.033 Hindu MaleSingleNo children 0.012 force 0.004 One child 0.020 back 0.005 Many children 0.012 sad 0.003 MarriedNo children 0.020 good 0.007 One child 0.03... | https://arxiv.org/abs/2505.18466v1 |
one child show words likelove in married categories, and change orhero in widowed and divorced categories. This indicates a broader role, where marriage creates positive associations, having children also creates some positive associations. •Many Children: Women with many children (both Hindu and Muslim) have the highe... | https://arxiv.org/abs/2505.18466v1 |
with additional bias terms such as iso- lated for single women, highlighting a stereo- type of seclusion for those that were never married. Married Muslim females with many children have high bias scores (0.086), rein- forcing traditional roles of women in large families. The term clothe (0.007)) appears for widowed Mu... | https://arxiv.org/abs/2505.18466v1 |
children 0.014 family 0.011 One child 0.014 family 0.006 Many children 0.055 family 0.052 Hindu MaleSingleNo children 0.000 N/A N/A One child 0.009 responsibility 0.009 Many children 0.041 responsibility 0.018 MarriedNo children 0.031 kind 0.016 One child 0.028 kind 0.010 Many children 0.067 good 0.018 DivorcedNo child... | https://arxiv.org/abs/2505.18466v1 |
and Religion •Hindu Female: The highest Bias TF-IDF terms for Hindu females include chore , with high bias scores across marital statuses and children categories. This term is mentioned frequently, with the highest TF-IDF value of 0.131 in the married identities. This indi- cates a strong bias towards associating Hindu... | https://arxiv.org/abs/2505.18466v1 |
chore 0.079 One child 0.103 chore 0.075 Many children 0.192 chore 0.098 WidowedNo children 0.148 chore 0.097 One child 0.114 chore 0.065 Many children 0.133 chore 0.049 Muslim MaleSingleNo children 0.020 offer 0.020 One child 0.031 responsibility 0.017 Many children 0.030 responsibility 0.014 MarriedNo children 0.054 f... | https://arxiv.org/abs/2505.18466v1 |
categories, with chore being the most frequent bias term, par- ticularly for those who are married with many children. Muslim females show similar bias trends, although their scores are slightly lower. The bias scores for Hindu and Muslim males are significantly lower, with responsibility be- ing the primary bias term.... | https://arxiv.org/abs/2505.18466v1 |
bal- ance domestic and professional roles. Wid- owed Muslim females show higher TF-IDF values for work (0.093) and husband ,suggest- ing a continuation of duties and relation- ships despite the loss of a spouse. •Muslim Male : There are higher Overall TF- IDF values for Muslim males in comparisonto Hindu males. Muslim ... | https://arxiv.org/abs/2505.18466v1 |
0.062 Many children work 0.086 DivorcedNo children doctor 0.088 One child husband 0.067 Many children husband 0.100 WidowedNo children work 0.093 One child die 0.084 Many children husband 0.056 Muslim MaleSingleNo children friend 0.092 One child work 0.095 Many children patient 0.089 MarriedNo children wife 0.084 One c... | https://arxiv.org/abs/2505.18466v1 |
•Muslim Female : Muslim females also feature preference prominently (up to 0.328). Woman (up to 0.254) frequently appears in Muslim female narratives, highlighting gendered ex- pectations of women in Muslim contexts. There’s also an emphasis on birth , highlight- ing themes of having children. •Muslim Male : Muslim mal... | https://arxiv.org/abs/2505.18466v1 |
pref-erence , showing individual interests. There are varying terms like Muslims for Muslim divorced males, and birth for Hindu males, andwoman for Muslim females. Yet, Hindu females are continually associated with pref- erence . G.43.3 Children-Based Analysis •No Children : Individuals with no children across all iden... | https://arxiv.org/abs/2505.18466v1 |
Hindu and Mus- lim females display higher TF-IDF values fordaily compared to their male counterparts. This highlights that women are connected to routines, especially after marriage. The higher TF-IDF values for women, particularly those with one child, reinforce the association of women with routines and family life. ... | https://arxiv.org/abs/2505.18466v1 |
Status : Widowed and divorced males show lower connections to term daily , show- ing a possible shift in routine after spousal loss. Widowed and married females have high TF-IDF values for daily . Children : The presence of children, espe- cially one child, strengthens the link between daily . This emphasizes routines ... | https://arxiv.org/abs/2505.18466v1 |
ambitious andsadappears Table 49: Bias scores for all identities along with the terms with highest Bias TF-IDF values under the application Story in Telugu outputs for the original prompting method. Religion & Gen- derMarital Sta- tusChildren Bias ScoreTerm Bias TF- IDF Hindu FemaleSingleNo children 0.038 home 0.009 On... | https://arxiv.org/abs/2505.18466v1 |
strong assumptions regarding women without children. For married males, terms like happy and grow reflect societal views that equate family and reproduction with fulfillment and personal growth. •Divorced: Divorced Hindu females are linked to terms like home , while divorced Mus- lim females exhibit a bias toward death... | https://arxiv.org/abs/2505.18466v1 |
are linked to terms like ambitious and conflict , which reflect a stereotype of men as ambitious and potentially involved in conflict, but with lower overall bias scores. Hindu males are often associated with back , rein- forcing the stereotype of backwardness. •Marital Status Bias: Widowed and divorced individuals, pa... | https://arxiv.org/abs/2505.18466v1 |
protection are present, the over- all bias is relatively low compared to Mus- lim females. G.46.2 Marital Status-Based Analysis •Single: Single Hindu and Muslim females are primarily associated with cook ,suggesting a bias that views them through the lens of Table 50: Bias scores for all identities along with the terms... | https://arxiv.org/abs/2505.18466v1 |
ated with family and nurturing. Interestingly, widowed Muslim women with one child are associated with traditional .Widowed women have lower bias scores than married indi- viduals. G.46.3 Number of Children-Based Analysis •No Children: Single women with no children have high bias scores. Hindu females with no children ... | https://arxiv.org/abs/2505.18466v1 |
marital statuses, with a particular emphasis on women with children. The bias score for Hindu females ranges from 0.138 to 0.203, indicating notable bias, especially for those who are married or divorced, with a higher bias for those with many children. Additionally, women that are married without children also have hi... | https://arxiv.org/abs/2505.18466v1 |
take on domestic roles remains high, despite the loss of a spouse. The bias scores for wid- owed males are again low, further emphasiz- ing the less scrutinized societal image of men in comparison to women. G.47.3 Number of Children-Based Analysis •No Children: Both Hindu and Muslim fe- males without children are assoc... | https://arxiv.org/abs/2505.18466v1 |
For males, terms like provide andmain- tenance indicate the stereotype of men being expected to provide for large families, but the bias scores are still lower than for fe- males. G.47.4 Summary of Findings •Gender-Based Bias: Female identities show the highest bias scores, with chore andcook as recurring terms. This r... | https://arxiv.org/abs/2505.18466v1 |
females. This could suggest stronger gendered associations with tradi- tional Muslim female roles . The terms admit (0.102) and woman (up to 0.113) are promi- nent for Muslim females with many children, reflecting a narrative where the female iden- tity is closely tied to motherhood and familial duties. These terms als... | https://arxiv.org/abs/2505.18466v1 |
a high TF- IDF for woman ,underscoring the contin- uing emphasis on gender identity despite marital status. •Widowed : There is an emphasis on numer- ical values , such as two, possibly indicating duality or description of family sizes. For male identities, there is an emphasis on wife, showing references to loss of th... | https://arxiv.org/abs/2505.18466v1 |
like father (up to 0.15) and sport (up to 0.117). While these terms still reflect a family-oriented identity, there are also associations to gendered activities. The term widow appears for divorced males, and childless for widowed males with no children. This suggests that divorced males are some- times associated with... | https://arxiv.org/abs/2505.18466v1 |
,father , man,woman , and family appear for those with many children. There are higher TF-IDF val- ues for woman than man,showing stronger ties to womanhood and motherhood, con- trasting fatherhood. Men, however, are more Table 53: Highest Overall TF-IDF terms and values for all identities under the application Hobbies... | https://arxiv.org/abs/2505.18466v1 |
most frequent terms for Hindu females include daily (up to 0.272), woman (up to 0.135), every (up to 0.158), and mother (0.119). The term daily indicates structured routine. Woman andmother sug- gest a connection to gender identity and motherhood. Every appears for widowed identities, possibly indicating structured ac-... | https://arxiv.org/abs/2505.18466v1 |
0.101 MarriedNo children husband 0.096 One child husband 0.130 Many children husband 0.104 DivorcedNo children day 0.107 One child family 0.108 Many children family 0.069 WidowedNo children daily 0.105 One child daily 0.144 Many children clean 0.078 tially signifying an emphasis on maintain- ing order or religious clea... | https://arxiv.org/abs/2505.18466v1 |
and Muslim men with many chil- dren, reinforcing gendered role assignments.•Religion-Based Analysis Muslim women had the highest TF-IDF values for daily , espe- cially when they were married with one child (0.344), highlighting a strong association be- tween routine tasks and Muslim women. The term husband appeared uni... | https://arxiv.org/abs/2505.18466v1 |
and struggling. These terms show a slightly negative bias, indicating a stereotype of Mus- lims being involved in difficult or struggle- driven scenarios, especially when single or divorced. In contrast, married individuals are associated with positive terms. G.51.2 Marital Status-Based Analysis •Single: The highest bi... | https://arxiv.org/abs/2505.18466v1 |
0.003 Muslim FemaleSingleNo children 0.025 isolate 0.005 One child 0.013 home 0.004 Many children 0.027 death 0.008 MarriedNo children 0.056 house 0.019 One child 0.060 respect 0.009 Many children 0.042 family 0.008 DivorcedNo children 0.028 death 0.010 One child 0.057 house 0.013 Many children 0.023 depression 0.005 W... | https://arxiv.org/abs/2505.18466v1 |
child, and divorced men with no children. There are terms like happiness (0.013) often associated with married men, showing satis- faction in marriage. Contrastingly, violent appears often especially for divorced males with no children. Furthermore, single men are associated with selfish . These terms suggest societal ... | https://arxiv.org/abs/2505.18466v1 |
stereotype of single women as more focused on emotional and caregiving roles. For single men, the highest bias term is selfish , suggesting negative stereotypes about single men being self-centered or individualistic. This term reflects societal bias that associates single men with autonomy, often perceived as undesira... | https://arxiv.org/abs/2505.18466v1 |
higher bias scores compared to their male counterparts, reflecting the stronger societal pressure on women to conform to traditional gender roles. •Marital Status Bias: Married individuals, es- pecially women, are associated with terms likefamily , indicating the expectation to ful- fill familial roles. Divorced and wi... | https://arxiv.org/abs/2505.18466v1 |
of both genders face notable bias, with Hindu and Muslim fe- males often associated with the term clean (ranging from 0.037 to 0.160). The associa- tion of cleanliness with single women high-lights the stereotype of women being ex- pected to maintain domestic spaces regard- less of their personal status. Hindu and Mus-... | https://arxiv.org/abs/2505.18466v1 |
0.159 cook 0.039 MarriedNo children 0.201 clean 0.031 One child 0.157 clean 0.033 Many children 0.152 cook 0.030 DivorcedNo children 0.195 clean 0.057 One child 0.085 chore 0.021 Many children 0.144 clean 0.034 WidowedNo children 0.088 clean 0.024 One child 0.072 clean 0.022 Many children 0.121 iron 0.032 Muslim MaleSi... | https://arxiv.org/abs/2505.18466v1 |
: Hindu males predominantly fea- ture terms like wife (up to 0.056), work (up to 0.052), and doctor (up to 0.133) signify- ing the emphasis on marital relationships and careers. The term name (up to 0.037) emerges frequently across different marital statuses, suggesting a connection to identity or legacy. •Muslim Femal... | https://arxiv.org/abs/2505.18466v1 |
work andhouse , which re- flect a balance between domestic and career related spaces. Doctor appears for females, while wife appears for males, showing depen- dence on various individuals.G.54.3 Children-Based Analysis •No children : For individuals with no children, terms like work andgoappear frequently, es- pecially... | https://arxiv.org/abs/2505.18466v1 |
with maternal identity. •Hindu Male : For Hindu males, artistic and father are frequently mentioned, especially for those with children. The term father is partic- ularly frequent for Hindu males with one child (up to 0.138), suggesting the importance of paternal identity. Despite the strong familial associations, Hind... | https://arxiv.org/abs/2505.18466v1 |
, while males are associated with religious . •Many Children : The presence of multiple children notably increases associations with artistic , reinforcing the connection between multiple children and creativity. Widowed in- dividuals with many children are consistently highly associated with artistic . Table 59: Highe... | https://arxiv.org/abs/2505.18466v1 |
= up to 0.154) and routine (up to 0.150). The term clean appears most often for single, married, and divorced Hindu females, reinforcing gendered expectations around domestic labor and hygiene. The presence ofroutine across all marital statuses suggests anemphasis on structured, repetitive daily tasks. •Hindu Male : Th... | https://arxiv.org/abs/2505.18466v1 |
routine anddaily appear frequently across married identities. Hindu females also feature clean and Muslim fe- males feature woman ,indicating an associa- tion with gender identity and chores. Con- trastingly, Hindu males are associated with work when they have one child, showing an emphasis on their career or providing... | https://arxiv.org/abs/2505.18466v1 |
like love andhouse appear in some cat- egories, particularly for married and widowed women, reinforcing traditional gender roles. The bias scores for Hindu females range from 0.035 to 0.076, with married women and those with multiple children experienc- ing the highest bias. These scores indicate there is a high expect... | https://arxiv.org/abs/2505.18466v1 |
0.027 happy 0.006 One child 0.038 love 0.011 Many children 0.041 happy 0.009 DivorcedNo children 0.007 struggle 0.002 One child 0.014 fix 0.002 Many children 0.014 fight 0.003 WidowedNo children 0.008 assault 0.002 One child 0.006 bad 0.002 Many children 0.004 provide 0.001 terms are found otherwise. However, bias scor... | https://arxiv.org/abs/2505.18466v1 |
with positive associations like indepen- dent for single men and love for married men. •Religious Stereotypes: Muslim males are dis- proportionately associated with aggression-related words like angry andthreaten , rein- forcing societal stereotypes of Muslim men as violent. Similarly, Hindu males are de- scribed with ... | https://arxiv.org/abs/2505.18466v1 |
The most frequent bias terms include stress (Bias TF- IDF of 0.027) and love (up to 0.046), in married and divorced categories respectively, showing positivity associated with those that are married. However, the presence of terms like evil(0.012) suggests potential bias associating Muslim males with negative connotati... | https://arxiv.org/abs/2505.18466v1 |
One child 0.105 family 0.035 Many children 0.167 family 0.053 MarriedNo children 0.143 family 0.038 One child 0.195 respect 0.030 Many children 0.200 family 0.086 DivorcedNo children 0.119 home 0.023 One child 0.078 family 0.029 Many children 0.198 family 0.060 WidowedNo children 0.037 family 0.016 One child 0.070 fami... | https://arxiv.org/abs/2505.18466v1 |
score (0.548) is observed for married Hindu women with many children, reinforc- ing the stereotype of domestic responsibil- ity increasing with marriage and mother- hood. The bias scores for Hindu females are generally higher than bias scores for Muslim females. •Hindu Male: The dominant term for Hindu males is respons... | https://arxiv.org/abs/2505.18466v1 |
child 0.057 responsibility 0.016 Many children 0.067 responsibility 0.026 DivorcedNo children 0.045 offer 0.020 One child 0.023 responsibility 0.015 Many children 0.035 responsibility 0.035 WidowedNo children 0.057 competitive 0.032 One child 0.028 responsibility 0.028 Many children 0.060 responsibility 0.046 •Married:... | https://arxiv.org/abs/2505.18466v1 |
est bias scores, reinforcing the stereotype of women as primary caregivers.G.60 Top Overall Terms for Kannada - Story To analyze the most frequent terms from the appli- cations and 48 identities Kannada generations of stories under the original prompting method, we highlight all 48 identities and terms with the high- e... | https://arxiv.org/abs/2505.18466v1 |
time, marriage, and youth. •One Child : Across most identities, boyorgirl appears, reinforcing the centrality of children in their narratives. Muslim widowed males with one child are associated with loss, show- ing an emphasis on emptiness associated with spousal loss. •Many Children : Identities with many chil- dren f... | https://arxiv.org/abs/2505.18466v1 |
0.143), and family (up to 0.138). Terms like family andagricul- tureare strongly associated with women with many children, suggesting a narrative cen- tered on domesticity and rural living. The use of ideal for single Hindu women with- out children implies that having no children before married is considered acceptable... | https://arxiv.org/abs/2505.18466v1 |
larly prominent for those with no children, reflecting an association of spiritual pursuits. Spiritual practices are prominent for Muslim males that are single with no children. Mus- lim women are associated often with woman , emphasizing gender identity. •One Child : One child significantly influences the terms agricu... | https://arxiv.org/abs/2505.18466v1 |
children more likely to have terms like family andagriculture , further ce- menting family values and rural, traditional practices. G.62 Top Overall Terms for Kannada - To-do List To analyze the most frequent terms from the ap- plications and 48 identities Kannada generations of daily activities under the original prom... | https://arxiv.org/abs/2505.18466v1 |
0.138 Muslim FemaleSingleNo children chore 0.135 One child chore 0.147 Many children chore 0.122 MarriedNo children chore 0.168 One child chore 0.128 Many children cook 0.123 DivorcedNo children task 0.091 One child household 0.102 Many children chore 0.109 WidowedNo children chore 0.119 One child chore 0.134 Many chil... | https://arxiv.org/abs/2505.18466v1 |
Pedagogy-R1: Pedagogically-Aligned Reasoning Model with Balanced Educational Benchmark Unggi Lee1†, Jaeyong Lee2†, Jiyeong Bae1, Yeil Jeong2, Junbo Koh2 Gyeonggeon Lee3,Gunho Lee1,Taekyung Ahn1,Hyeoncheol Kim4 1Enuma, Inc.,2Seoul National University,3Nanyang Technological University,4Korea University Corresponding Auth... | https://arxiv.org/abs/2505.18467v1 |
al ., 2024; Puech 1arXiv:2505.18467v1 [cs.AI] 24 May 2025 QwQ -32B Stage 1. Building Well -balanced Educational Benchmark Stage 2. Pedagogical Reasoning and CoP Prompting Educational Data LLM HumanRefinement Collection Curation DeepSeek -R1Format GuideReasoning DistillationPedagogy Distillation Pedagogy -R1 Stage 3. Mi... | https://arxiv.org/abs/2505.18467v1 |
and the limitations of LRMs in educational contexts. 1.1 Contributions •Pedagogical Reasoning Framework : We pro- pose a distillation-based training pipeline that 2 aligns LLMs with teacher-like reasoning through pedagogically filtered instruction tuning. •Chain-of-Pedagogy Prompting : We develop CoP, a prompting strat... | https://arxiv.org/abs/2505.18467v1 |
to pedagogically guided reasoning. We integrate CoP in two core stages of our frame- work. First, we apply CoP prompting during infer- ence time to guide the model’s step-by-step reason- ing toward educationally aligned outputs. Second, 3 we leverage CoP prompts during data generation by the teacher model—instructing i... | https://arxiv.org/abs/2505.18467v1 |
ensure format consistency with other benchmark compo- nents. All items were originally written in Korean. To build the dataset, we collected official exam PDFs from the KICE website and applied OCR to extract the text. OCR errors, especially in older documents, were manually reviewed and corrected to ensure fidelity. T... | https://arxiv.org/abs/2505.18467v1 |
is tailored for programming educa- tion and was released by the CSEDM Workshop at LAK 2019 (CSEDM Workshop, 2019). Since the original questions and knowledge conceptswere unavailable, we generated them from student- submitted Java code using GPT-4o-mini. Each in- teraction includes a reconstructed question, a gen- erat... | https://arxiv.org/abs/2505.18467v1 |
25.89 (+2.38) Pedagogy-R1-CoP-7B + CoP (Ours) 27.08 (-2.61) 29.70 (+4.52) 57.36 (-2.45) 63.30 (+22.7) 7.02 (-0.06) 25.91 (+2.49) turn-level annotations (e.g., teacher uptake, focus- ing questions, student reasoning), expert observa- tion scores (CLASS, MQI), teacher and student demographics, value-added scores, and sur... | https://arxiv.org/abs/2505.18467v1 |
tuning leads to better performance in educational reasoning tasks than simple instruc- tion tuning based on problem-answer formats. Among the base models, performance trends partially correlate with model size. In the Qwen2.5 series, SK increases from 26.83% (1.5B) to 35.65% (32B), and KT-AUC increases from54.77% (1.5B... | https://arxiv.org/abs/2505.18467v1 |
KT, COP prompting sometimes increases both reasoning tokens and UT Score but reduces contrast transitions, indicating a complex interplay between prompting and transition strategies. 5.2 Qualitative Analysis To probe how the CoP prompt shaped the model’s reasoning, we adopted Schön’s tripar- tite conception of professi... | https://arxiv.org/abs/2505.18467v1 |
reconstruction and policy interpretation typical of these domains. This disciplinary fingerprint aligns with cognitive task demands: procedural subjects privilege iterative plan-execute cycles, while interpretive subjects fa- vor comparative judgment and narrative synthesis. 5.2.4 Theme 3: Disciplinary Signatures in Re... | https://arxiv.org/abs/2505.18467v1 |
Related work 6.1 Large Reasoning Models The release of OpenAI’s o1 (Jaech et al ., 2024) marked the beginning of a new era in LRMs, tran- sitioning from simple generation tasks to perform- ing complex, multi-step reasoning, known as Test- time Scaling (TTS). Following this, several ef- forts attempted to replicate o1’s... | https://arxiv.org/abs/2505.18467v1 |
and educational levels, but is not openly accessible to the broader research community (AI-For-Education.org, 2025). In contrast, MathTutorBench offers an open-source framework for domain-specific evaluation in math- ematics, utilizing datasets and metrics grounded in learning sciences research (Macina et al ., 2025). ... | https://arxiv.org/abs/2505.18467v1 |
ing Dataset Based on Online Student Evaluation. arXiv preprint arXiv:2208.12651 (2022). AI-For-Education.org. 2025. The Pedagogy Bench- mark. https://benchmarks.ai-for-education. org/#moreinfo-about-the-benchmark Dana AlZoubi. 2022. From data to actions: Unfold- ing instructors’ sense-making and reflective practice wit... | https://arxiv.org/abs/2505.18467v1 |
Fu, Hao Guan, Kounianhua Du, Jianghao Lin, Wei Xia, Weinan Zhang, Ruiming Tang, Yasheng Wang, and Yong Yu. 2024. Sinkt: A structure-aware inductive knowledge tracing model with large lan- guage model. In Proceedings of the 33rd ACM Inter- national Conference on Information and Knowledge Management . 632–642. Rujun Gao,... | https://arxiv.org/abs/2505.18467v1 |
Michael Sears, Filip Bar, Mia Mesar, Mana Jab- bour, Arslan Chaudhry, James Cohan, Sridhar Thia- garajan, Nir Levine, Ben Brown, Dilan Gorur, Svet- lana Grant, Rachel Hashimshoni, Laura Weidinger, Jieru Hu, Dawn Chen, Kuba Dolecki, Canfer Akbu- lut, Maxwell Bileschi, Laura Culp, Wen-Xin Dong, Nahema Marchal, Kelsie Van... | https://arxiv.org/abs/2505.18467v1 |
Nico Daheim, Ido Hakimi, Manu Ka- pur, Iryna Gurevych, and Mrinmaya Sachan. 2025. MathTutorBench: A Benchmark for Measuring Open-ended Pedagogical Capabilities of LLM Tu- tors. arXiv preprint arXiv:2502.18940 (2025). Oliver McGarr. 2021. The use of virtual simulations in teacher education to develop pre-service teacher... | https://arxiv.org/abs/2505.18467v1 |
ucation: A survey and outlook. arXiv preprint arXiv:2403.18105 (2024). Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang, et al .2025. Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMs. arXiv preprint arXiv:2501.18585 (2025). Jerom... | https://arxiv.org/abs/2505.18467v1 |
1 Investigating AI Rater Effects of L arge L anguage M odels: GPT, Claude, Gemini, and DeepSeek Hong Jiao University of Maryland, College Park Dan Song Won- Chan Lee University of Iowa Abstract Large language models (LLM s) have been widely explored for automated scoring in low -stakes assessment to facilitate learning... | https://arxiv.org/abs/2505.18486v1 |
BERT, de BERTa based on transformer (Vaswani et al., 2017) and large language models (LLMs) such as GPT, Claude, and Gemini . Developing feature -based or SLM -based automated scoring systems often requires foundational skills in supervised machine learning, NLP, and deep learning. This dramatically reduces the accessi... | https://arxiv.org/abs/2505.18486v1 |
introduce score errors for students whose true scores lie at the two ends of the score scale. Such rater effects are as severe as score accuracy as they may compromise the fairness, reliability, and validity of the assigned scores. Though in operational rating sessions, different strategies would be implemented to mini... | https://arxiv.org/abs/2505.18486v1 |
specific scores ranging from 0 to 6 (Song & Tang, 2025). LLM s Trained for Automated Scoring 4 In alignment with the development of LLMs, ChatGPT 3.5 was first used to sco re all 120 essays and other AI engines were sequentially explored when new versions or new engines were released . Data collection took approximatel... | https://arxiv.org/abs/2505.18486v1 |
rubrics. In total, 2 human raters and 1 0 AI raters scored 120 essays across four writing tasks. Analysis Using the analytic and holistic scores for the essays, this study investigate s scoring accuracy, reliability, and rater effects. Quadratic Weighted K appa (QWK ; Cohen, 1960; 1968) was computed to evaluate the sco... | https://arxiv.org/abs/2505.18486v1 |
infit and outfit mean square. Infit mean square i ndicates the consistency a rater uses the rating scale across students and criteria while outfit mean squares quantifies the outlier score patterns. The expected values of the infit and outfit mean squares are 1. Values smaller than 1 show less variation than expected i... | https://arxiv.org/abs/2505.18486v1 |
0.543 0.304 0.656 0.584 A7 – Gemini 2.0 0.201 0.131 0.319 0.199 0.289 0.143 0.716 0.619 A8 – Deep Seek V3 0.378 0.233 0.610 0.439 0.619 0.448 0.664 0.667 A9 - Gemini 1.5 pro 0.267 0.169 0.698 0.576 0.631 0.640 0.590 0.537 A10 - DeepSeek R1 0.579 0.522 0.417 0.293 0.453 0.368 0.727 0.703 A11= Ensemble 0.606 0.426 0.605 ... | https://arxiv.org/abs/2505.18486v1 |
yielded the highest QWK with each of the two raters on both SN1 and ER1. For SN2 , Gemini 1.5 Pro yielded the highest QWK with both human raters. O nly two ensemble models improved scoring accuracy over the standalone LLM S for SN1 and ER1. Unlike analytic scores for Task Completion, AI -human QWK was consistently high... | https://arxiv.org/abs/2505.18486v1 |
0.368 0.417 0.553 0.581 A9 - Gemini 1.5 pro 0.447 0.304 0.626 0.677 0.599 0.622 0.495 0.528 A10 - DeepSeek R1 0.432 0.382 0.216 0.246 0.455 0.459 0.665 0.743 A11=Ensemble 0.460 0.382 0.429 0.417 0.644 0.558 0.618 0.685 A12=A11 -A7 0.460 0.382 0.560 0.455 0.672 0.709 0.652 0.716 A13=A12 -A9 0.432 0.364 0.509 0.397 0.690... | https://arxiv.org/abs/2505.18486v1 |
The red colored numbers indicate the AI rater who assigned scores with the highest mean score among all AI raters or a human rater with higher scores compared with the other human rater while the blue- colored numbers indicate the lowest mean scores assigned among AI raters . To some degree, t his pattern reflect s the... | https://arxiv.org/abs/2505.18486v1 |
holistic scoring as shown in Figure 1, SN2 wa s the most difficult essay , while E R2 was the easiest essay. Gemini 2.0 was the most stringent rater, which is consistent with the lowest mean scores assigned while R2 and ChatGPT 3.5 were more lenient raters as they assigned the highest or higher mean scores as presented... | https://arxiv.org/abs/2505.18486v1 |
a smaller difference of about 0.08. Consistent with what presented in Figure 2, Gemini 2.0 was a severe rater with a logit value of 0.76. DeepSeek V3 was the most neutral rater with a rater parameter estimate close to 0. All AI raters did not have severe rater effects except OpenAI o1 (0.5) at the harsh end. OpenAI o1 ... | https://arxiv.org/abs/2505.18486v1 |
R 2 was the most lenient raters. R1 was the second most lenient rater like Claude 3.5 and Gemini 1.5 Pro on Delivery trait scoring. Table 10 presents the rater parameter estimates for the analytic scores for Delivery . Human rater R2 was the most lenient. Gemini 2.0 was the most severe rater with a rater parameter of 1... | https://arxiv.org/abs/2505.18486v1 |
Discussion This study aims to examine the rater effects of LLM -based AI raters in automated scoring. To illustrate the methods for exploring AI rater effects, this study investigated the use of LLMs for automated holistic and analytic scoring using a fully crossed rating design. Ten LLMs were compared including ChatGP... | https://arxiv.org/abs/2505.18486v1 |
with DeepSeek V3 and Gemini 1.5 Pro as the two other lenient raters. Whether an AI rater was more neutral raters depends on the task. There was no evident central tendency effects for the AI raters. However, the two human raters displayed some minor central tendency effect in holistic scoring. As Wolfe (2004) pointed o... | https://arxiv.org/abs/2505.18486v1 |
C. M. (2007). Applying the Rasch model: Fundamental measurement in the human sciences (2nd ed.). Mahwah, NJ: Erlbaum Bond, T. G , Yan , Z., & Heene, M. (2021). Applying the Rasch model: Fundamental measurement in the human sciences (4th ed). Routledge Bui, N.M., & Barrot, J.S. (2024). ChatGPT as an automated essay scor... | https://arxiv.org/abs/2505.18486v1 |
03209-9 Linacre , J. M. (1994). Many-Facet Rasch Measurement , (2nd ed). Chicago , IL: MESA Press . Linacre, J. M. (2003). Size vs. Significance: Infit and Outfit Mean-Square and Standardized Chi- Square Fit Statistic. Rasch Measurement Transactions , 17:918 Linacre, J. M. (2024). FACET computer program for many- facet... | https://arxiv.org/abs/2505.18486v1 |
Ritchie, D., ... & Warschauer, M. (2024). Can AI provide useful holistic essay scoring?. Computers and Education: Artificial Intelligence , 7, 100255. https://doi.org/10.1016/j.caeai.2024.100255. Wijekumar, K. K., McKeown, D., Zhang, S., Lei, P. W., Hruska, N., & Pirnay-Dummer, P. (2024). We write automated scoring: Us... | https://arxiv.org/abs/2505.18486v1 |
more. He is listening to music. The second picture: The boy puts everything under his bed, covers it with a bedsheet, and pretends he has cleaned the room. The third picture: His dad removes the bedsheet, sees the mess, and is very surprised and angry. The fourth picture: His dad goes downstairs a nd tells his son what... | https://arxiv.org/abs/2505.18486v1 |
repeated interference from another language 2) Limited grammatical structures, with frequent errors that obscur e meaning Score 2: 小蔡在床上，爸爸来房间，爸爸回到房间看，儿子没有整理。 Score 2: 今天爸爸来到我的房间，让我打扫，可是我不想打扫。 Score of 3: Adequate Suggests competence in interpersonal writing -TASK COMPLETION: 1) E-mail addresses topic directly but may ... | https://arxiv.org/abs/2505.18486v1 |
0.67 1.38 1.03 1.06 1.17 1.47 1.46 Task_SN1 1.03 0.85 0.79 0.79 1.25 0.79 1.10 0.90 0.68 1.15 1.02 0.84 Task_SN2 0.89 0.78 0.82 0.92 1.07 0.66 1.03 1.16 0.74 1.04 0.68 1.11 Task_ER1 1.20 1.06 1.53 0.82 1.18 1.13 1.30 0.83 0.86 1.09 1.31 1.40 Task_ER2 1.40 1.28 1.17 1.23 1.25 0.73 1.28 1.00 0.98 1.10 1.44 1.46 Delivery_... | https://arxiv.org/abs/2505.18486v1 |
arXiv:2505.18488v1 [cs.LG] 24 May 2025Synthesizing and Adapting Error Correction Data for Mobile Large Language Model Applications Yanxiang Zhang∗, Zheng Xu∗, Shanshan Wu∗, Yuanbo Zhang, Daniel Ramage Google {zhangyx, xuzheng, shanshanw, zyb, dramage }@google.com Abstract Error correction is an important capability whe... | https://arxiv.org/abs/2505.18488v1 |
applications is challenging because of the domain shift and privacy considerations on user data. Production LLM mobile applications have developed pipelines to synthesize error correction data. Liu et al. [18]collects public web data, and then uses trained task-specific models [ 16,26] to detect grammatical errors. A t... | https://arxiv.org/abs/2505.18488v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.