text string | source string |
|---|---|
†Email: wanli.jwl@alibaba-inc.com ‡Email: langyu.yzy@alibaba-inc.com 1arXiv:2505.20336v1 [cs.CL] 24 May 2025 combination of preferences once training is completed, whereas reward soups allows for more flexi- bility by combining preferences at the inference stage but at the cost of increased computational and training o... | https://arxiv.org/abs/2505.20336v1 |
achieving a balanced alignment of user preferences. However, MORLHF requires retraining for each combination of preferences. To address this, Re- warded Soups (Ram ´e et al. 2023) proposes a novel method for parameter combination. It involves training separate reward models and policy models for each individual prefere... | https://arxiv.org/abs/2505.20336v1 |
question-answer (Q,A) sequence into preferences such as helpfulness, harmlessness, or honesty. Within each head (e.g., helpful), we classify the intensity of the preference. We then optimize our reward model by calculating the accuracy of intensity classification for each head. The overall architecture is showed in 2. ... | https://arxiv.org/abs/2505.20336v1 |
consistent metric for comparability. To address these challenges, we propose a reward mapping function, which can scale both dimen- sions and intensities of preferences. During the training phase, we record the moving average and standard deviation of each preference intensity within every preference head. During infer... | https://arxiv.org/abs/2505.20336v1 |
demonstrating that our reward model achieves significant performance gains over GPT-4 annotators. The detailed results compared with GPT-4 is showed in Appendix E,Figure 9. Figure 3: Construction process of reward model training datasets. Data Construction: We utilize two open-sourced datasets to train our reward model... | https://arxiv.org/abs/2505.20336v1 |
that is used to train the reward model. We compare our MOSLIM with baseline methods, MORLHF, Rewarded Soups (RSoups), and RiC, to demonstrate the effectiveness of our training paradigm. All three policy models, except RiC, are trained using PPO as the policy optimization algorithm, while RiC directly uses the SFT resul... | https://arxiv.org/abs/2505.20336v1 |
while being more efficient in terms of training time.Among the baselines, MORLHF performs the worst, showing a clear drop in performance as the task difficulty increases. Rewarded Soups achieves slightly better results than MORLHF, with relatively stable scores across different datasets. RiC further improves over RSoup... | https://arxiv.org/abs/2505.20336v1 |
MORLHF RSopu 0.0 0.2 0.4 0.6 0.8 1.0 Preference Intensity <harmless score n>0.50.60.70.80.91.0harmless score MOSLIM MORLHF RSopu Figure 5: Controllability experiment results of preference intensity. From left to right, the subfigures represent preference goals <helpfulness n> ,<honesty n> , and<harmless n> , with the y... | https://arxiv.org/abs/2505.20336v1 |
language model alignment through reinforcement learning. Both Jang et al. (2023) and Li et al. (2024c) focus on personalizing language models to generate content tailored to individual users, extending alignment objectives from general preferences to personal objectives. These works aim to push the bound- aries of mult... | https://arxiv.org/abs/2505.20336v1 |
al. The llama 3 herd of models. ArXiv , abs/2407.21783, 2024. URL https: //api.semanticscholar.org/CorpusID:271571434 . Shangmin Guo, Biao Zhang, Tianlin Liu, Tianqi Liu, Misha Khalman, Felipe Llinares, Alexandre Rame, Thomas Mesnard, Yao Zhao, Bilal Piot, et al. Direct language model alignment from online ai feedback.... | https://arxiv.org/abs/2505.20336v1 |
weights fine-tuned on diverse rewards. ArXiv , abs/2306.04488, 2023. URL https://api.semanticscholar.org/CorpusID:259096117 . Sander Schulhoff, Jeremy Pinto, Anaum Khan, Louis-Franc ¸ois Bouchard, Chenglei Si, Svetlina Anati, Valen Tagliabue, Anson Liu Kost, Christopher Carnahan, and Jordan Boyd-Graber. Ig- nore this t... | https://arxiv.org/abs/2505.20336v1 |
Data Type Category Description Categories DataType 1 <helpfulness 1 > <honesty 1 > <harmless 1 >to<harmless 2 >4 DataType 2 <helpfulness 1 >to<helpfulness 2 > <honesty 1 >to<honesty 2 > <harmless 1 >to<harmless 2 >8 DataType 3 <helpfulness 1 >to<helpfulness 3 > <honesty 1 >to<honesty 3 > <harmless 1 >to<harmless 2 >18 ... | https://arxiv.org/abs/2505.20336v1 |
arXiv:2505.20338v1 [cs.CL] 24 May 2025JOURNAL OF L ATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 1 Assessing the Capability of LLMs in Solving POSCOMP Questions Cayo Viegas‡, Rohit Gheyi‡, M´arcio Ribeiro† ‡UFCG, Brazil cayo.viegas@ccc.ufcg.edu.br and rohit@dsc.ufcg.edu.br †UFAL, Brazil marcio@ic.ufal.br Abstract —Rece... | https://arxiv.org/abs/2505.20338v1 |
Nunes et al. [7] conducted a comprehensive analysis of ChatGPT-3.5 and ChatGPT-4 on ENEM questions from 2009–2017 and 2022, revealing that ChatGPT-4 with Chain-of-Thought achieved an impressive 87% accuracy on the 2022 exam. While these studies demonstrate promising results for LLMs in high school and undergraduate-lev... | https://arxiv.org/abs/2505.20338v1 |
(GAOKAO), excluding those containing images. Zero-shot prompting and human evaluation were used. Results showed that the models performed well on knowledge-based questions but struggled with specific logical reasoning, math problems, and reading comprehension of longer Chinese texts. In our work, we evaluated four LLMs... | https://arxiv.org/abs/2505.20338v1 |
2022, the study used multiple prompt strategies, including Chain-of-Thought (CoT) [6]. ChatGPT- 4 with CoT achieved 87% accuracy on the 2022 exam, outperforming ChatGPT-3.5 by 11 points. In our study, we evaluated four LLMs on two POSCOMP editions using zero- shot prompts. In our work, ChatGPT-4 achieved the highest pe... | https://arxiv.org/abs/2505.20338v1 |
with 617 and 778 participants, respectively. The previous edition in 2022 was held in person. POSCOMP provides participants with detailed individual results, including correct/incorrect answers, overall averages, and standard deviations. Furthermore, official exams and an- swer keys are published online [1]. Since 2006... | https://arxiv.org/abs/2505.20338v1 |
only provided with the translated text. The evaluation was conducted between March 16 and March 22, 2024, using a zero-shot prompt approach, where no prior examples were given to the LLM [20], [21]. We used the English version of each question as the prompt input. Each exam contains three images: three class diagrams a... | https://arxiv.org/abs/2505.20338v1 |
indicating a tendency to be less decisive or perhaps more exploratory in its response strategy. In contrast, Mistral, ChatGPT-4, and Gemini tended to leave some questions unan- swered in 2022, suggesting possible gaps in their knowledge or caution in their response approach. The question shown in Figure 5 (Question 29 ... | https://arxiv.org/abs/2505.20338v1 |
on the models’ responses. ChatGPT-4 correctly answered Question 14 from POSCOMP 2023 (see Figure 8), which none of the other LLMs answered accurately. The goal is to find an expression from the thruth table. ChatGPT-4 constructed the truth table for all options and explained why option A was correct. Gemini asserted th... | https://arxiv.org/abs/2505.20338v1 |
sorting method represented by a linear list consisting of elements with keys s1, . . . , s n, satisfying the following property: si≤s⌈1/2⌉, for 1i≤n? TABLE IV MODELS ’RESPONSES TO POSCOMP 2022 BY LEVEL OF EXPLANATION . Explanation ChatGPT Gemini Claude Mistral Topic 38/69 52/69 50/69 38/69 Selec. option 22/69 48/69 68/... | https://arxiv.org/abs/2505.20338v1 |
excelled in many areas, Gemini, Claude, and Mistral also achieved high success rates. For example, in 8 questions related to Analytical Geometry, Differential and Integral Calculus, and Artificial Intelligence, all models performed strongly with success rates above 60%. However, differences emerged in 6 questions relat... | https://arxiv.org/abs/2505.20338v1 |
the statement, f3(n) = O(2∧n) was changed to f3(n) = O(2∧(3∧n)). Options B and C were swapped with options D and E. •Question 30: Statement III was modified from “Integer types are used to store values that belong to the set of natural numbers (without fractional parts)” to “Theboolean type stores only the ‘False’ valu... | https://arxiv.org/abs/2505.20338v1 |
“efficient” was changed to “IN- EFFICIENT.” •Question 65: The subnet mask in the question statement was changed from 255.255.255.128 to 255.255.255.64. Options A) 126, B) 128, D) 255.255.255.128, and E) 256 were changed to A) 128, B) 62, D) 255.255.255.64, and E) 255.255.255.128. To ensure accuracy in the modified resp... | https://arxiv.org/abs/2505.20338v1 |
69, translating to a success rate of around 15.9%. The model performed poorly across all subjects, particularly in Mathematics and Computing Technology. Gemini’s challenges were exacerbated by hallucination issues during the test, leading to incorrect or nonsensical answers. As observed, both LLMs faced challenges in t... | https://arxiv.org/abs/2505.20338v1 |
Computer Science (CS) Technologies category and demon- strated notably high accuracy in CS Fundamentals, particularly in the 2023 and 2024 exams. These results suggest that Gemini excels in algorithmic reasoning and technical problem-solving. Model o1 also demonstrated strong and consistent per- formance, especially in... | https://arxiv.org/abs/2505.20338v1 |
in Mathemat- ics, 24/30 in CS Fundamentals, and 18/20 in CS Technologies. Gemini 1.0 (13/19, 23/30, 17/20) and Mistral (14/19, 21/30, 15/20) followed closely. Claude 3 reached 9/19 in Mathematics and 22/30 in CS Fundamentals, placing among the top 10%. Among the newer models, Gemini 2.5 and o3-mini-high reached 100% ac... | https://arxiv.org/abs/2505.20338v1 |
50 out of 69 (72.5%) in 2023. While its results were comparable to those of Gemini 1.0 and Claude 3, Mistral showed noticeable variability across topics, performing better in Computer Science Technologies than in Mathematics or Fundamentals. RQ 5To what extent do recent LLMs match or surpass human performance in POSCOM... | https://arxiv.org/abs/2505.20338v1 |
POSCOMP. Other models, such as Gemini 1.0 Advanced and Le Chat Mistral, demonstrated problem-solving capabilities across var- ious subjects but didn’t match the achievements of ChatGPT-4. Claude 3 Sonnet showed promise in explanation and compre- hension but lagged slightly behind, struggling particularly with Mathemati... | https://arxiv.org/abs/2505.20338v1 |
L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017, pp. 5998–6008. [4] N. C. Mendonc ¸a, “Evaluating ChatGPT-4 vision on brazil’s national un- dergraduate computer science exam,” ACM Transactions on Computing Education , vol. 24, no. 3, pp. 1–56, 2024.... | https://arxiv.org/abs/2505.20338v1 |
G. Neubig, “Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing,” ACM Computing Surveys (CSUR) , vol. 55, no. 9, pp. 1–35, 2023. [22] J. Sallou, T. Durieux, and A. Panichella, “Breaking the silence: the threats of using llms in software engineering,” in ACM/IEEE 46th... | https://arxiv.org/abs/2505.20338v1 |
Dynamic Manifold Evolution Theory: Modeling and Stability Analysis of Latent Representations in Large Language Models Yukun Zhang The Chinese University Of Hongkong HongKong, China 215010026@link.cuhk.edu.cnQi Dong Fudan University Shanghai, China 19210980065@fudan.edu.cn Abstract We introduce Dynamic Manifold Evolutio... | https://arxiv.org/abs/2505.20340v1 |
time, conceptual- izes the LLM generation process as a dynamical system evolving on high-dimensional manifolds and establishes a rigorous mathematical link be- tween latent representation evolution and generated text quality. We develop both continuous-time and discretized dynamical system models, provide ex- 1arXiv:25... | https://arxiv.org/abs/2505.20340v1 |
manifold structure to generation qual- ity. Latent Trajectory Analysis in Language Mod- els Examining how hidden states evolve during text generation has illuminated RNN behavior (Li et al., 2016; Mardt et al., 2018) and Transformer residual streams (Elhage et al., 2021). Recent work investigates trajectory bifurcation... | https://arxiv.org/abs/2505.20340v1 |
stable region, explaining why LLMs generate focused, logically connected passages in- stead of disjoint word sequences. 3.3 Dynamical System Modeling and Transformer Mapping We model latent evolution as a controlled dynami- cal system on the semantic manifold: dh(t) dt=−∇V h(t) +g h(t),u(t) , where Vis an energy po... | https://arxiv.org/abs/2505.20340v1 |
in latent space can be in- terpreted as stable representations of grammat- ical states. High clustering quality indicates that such grammatical regimes are clearly sepa- rated, allowing the model to reliably identify and maintain correct syntax. For instance, syntactic rules—such as subject-verb agreement or tense con-... | https://arxiv.org/abs/2505.20340v1 |
the four latent-dynamics metrics from each generated sequence. Algorithm 1 Latent Dynamical System Validation Require: Transformer model M, input text x Ensure: Dynamics metrics {δ,J, s, ρ} 1:H←GetHiddenStates (M, x) 2:δ←ComputeDistances (H) 3:J ← DetectJumps (δ) 4:V←ReduceDim (H) 5:s←ClusterStates (V) 6:ρ←ComputePersi... | https://arxiv.org/abs/2505.20340v1 |
yields highly deter- ministic, smooth trajectories converging on major attractors. Moderate temperature (0.6≤τ≤1.2) enables a balance of exploration and convergence, maximizing both continuity and topological persis- tence. High temperature (τ≥1.3) leads to more stochastic, jumpy trajectories and weaker cluster- ing. L... | https://arxiv.org/abs/2505.20340v1 |
real-time or large-scale de- ployment. Second, while we demonstrate strong correlations, causal relationships between latent dynamics and text quality remain to be established; developing interventions to directly manipulate la- tent trajectories will be crucial. Fourth, our frame- work rests on the idealized manifold ... | https://arxiv.org/abs/2505.20340v1 |
Nelson Elhage, Neel Nanda, Catherine Olsson, and oth- ers. A mathematical framework for transformer cir- cuits. Anthropic Technical Report , 2021. R. Thomas McCoy, Harsh Trivedi, Richard Socher, and others. Linguistic structure inherent in GPT embed- dings. arXiv preprint arXiv:2211.00603 , 2022. Alex Hernandez-Garcia,... | https://arxiv.org/abs/2505.20340v1 |
Science , 2000. Matheus Viana Santos, Yuan Gao, Aditi Krishnapriyan, and Michael W. Mahoney. A theory of learning dy- namics in transformers with application to optimiza- tion. arXiv preprint arXiv:2306.01129 , 2023. Junjie Shi, Suhang Wang, and Dongwon Lee. Detailed evaluation of output stability in large language mod... | https://arxiv.org/abs/2505.20340v1 |
Implementa- tion specifics for computing dynamic metrics (state continuity, clustering, persistence) and text-quality metrics (perplexity, lexical diver- sity, grammar, coherence). •Appendix D: Supplementary Results (corre- sponds to Sec. 4.2) Additional visualizations, including single-sequence trajectory phases and t... | https://arxiv.org/abs/2505.20340v1 |
zation, multi-head self-attention (MHSA) can ap- proximate any Lipschitz-continuous function g:Rd×Rn×d→Rd arbitrarily well. Proof Sketch. Standard MHSA computes MHSA( ht,X) =hX i=1WO i nX j=1αij(ht,X)·WV ixj! , (7) where the attention weights are defined as αij(ht,X) =exp (WQ iht)⊤(WK ixj)/√dk Pn j′=1exp (WQ iht)⊤(W... | https://arxiv.org/abs/2505.20340v1 |
Coherence = 0 .7×Local + 0 .3×Global . •Factuality: Checked against a knowledge graph; factuality score is #correct facts #verifiable facts. C.3 Supplementary Results Trajectory Evolution Analysis. Figure 7-10 de- pict the evolution of latent representations along the generation trajectory of a single sequence. By trac... | https://arxiv.org/abs/2505.20340v1 |
along a generation trajectory in 2D latent space (temperature=0.1 and top_K=0.6). 15 Figure 9: Dynamic evolution along a generation trajectory in 2D latent space (temperature=0.1 and top_K=1.0). Figure 10: Dynamic evolution along a generation trajectory in 2D latent space (temperature=2.0 and top_K=0.6). 16 D.4 Paramet... | https://arxiv.org/abs/2505.20340v1 |
Towards Emotionally Consistent Text-Based Speech Editing: Introducing EmoCorrector and The ECD-TSE Dataset Rui Liu1, Pu Gao1, Jiatian Xi1, Berrak Sisman2, Carlos Busso3, Haizhou Li4,5 1Inner Mongolia University, Hohhot, China 2Center for Language and Speech Processing, Johns Hopkins University, USA 3LTl, Carnegie Mello... | https://arxiv.org/abs/2505.20341v1 |
Data Traditional T ext-based Speech Editing Model Proposed Method😡 😡😁 😁 😁Figure 1: Our approach lies in correcting the emotional mis- match or inconsistency issue of traditional TSE methods. out altering the speaker’s identity. However, directly using an EVC model presents several issues. 1) Before applying EVC, i... | https://arxiv.org/abs/2505.20341v1 |
which advances emo- tional consistency modeling and evaluation in the field of TSE. 2. Dataset: ECD-TSE For the emotionally consistent modeling of TSE, we need a database where sentences conveying similar lexical informa- tion with few modifications elicit different emotions. Unfor- tunately, popular databases such as ... | https://arxiv.org/abs/2505.20341v1 |
1 in Fig.2), 2) Speaker-Emotion Disentanglement Pre-training (Block 2 in Fig.2), and 3) Emotion Post-Correction for TSE (Block 3 in Fig.2). The first two modules require pre- training, and then the third module is trained end-to-end. 3.1. Text-Speech Emotion Retrieve Database Construction Text-Speech Emotion Contrastiv... | https://arxiv.org/abs/2505.20341v1 |
is the sum of these losses: L=Ls+Le+L(e) s+L(s) e. 3.3. Emotion Post-Correction for TSE As illustrated in Fig. 2 (Block 3), the purpose of emotional post- correction is to perform emotional correction on edited speech with inconsistent emotional expressions based on the emotion of the edited text, ensuring that the fin... | https://arxiv.org/abs/2505.20341v1 |
contained in the text to calculate the accu- racy. Due to Qwen2’s output characteristics, we calculate ac- curacy group-wise and then average the results. 2) Emotional Cosine Similarity (ECS) adopts emotion2vec [22] to extract the emotional features of the audio before and after correction Table 3: Emotional comparison... | https://arxiv.org/abs/2505.20341v1 |
Analysis of parameter K: In cross-modal emotion retrieval, the number of retrieved Top Ksamples may have an impact on the final emotion expression effect. We set Kto{3,5,10}to an- alyze the speech generation results, since 1 may result in insuf- ficient information and a selection between 5 and 10 might not demonstrate... | https://arxiv.org/abs/2505.20341v1 |
in Findings of the Association for Computational Linguistics: ACL 2023 , 2023, pp. 11 655– 11 671. [2] H. Bai, R. Zheng, J. Chen, M. Ma, X. Li, and L. Huang, “A3t: Alignment-aware acoustic and text pretraining for speech synthe- sis and editing,” in International Conference on Machine Learn- ing. PMLR, 2022, pp. 1399–1... | https://arxiv.org/abs/2505.20341v1 |
2021, pp. 920–924. [14] J. Duret, M. Rouvier, and Y . Est `eve, “Msp-podcast ser chal- lenge 2024: L’antenne du ventoux multimodal self-supervised learning for speech emotion recognition,” arXiv preprint arXiv:2407.05746 , 2024. [15] C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, a... | https://arxiv.org/abs/2505.20341v1 |
arXiv:2505.20343v1 [cs.CL] 24 May 2025DOLLM S HAVE A GENDER (ENTROPY ) BIAS? Sonal Prabhune, Balaji Padmanabhan, and Kaushik Dutta {saprabhune@usf.edu, bpadmana@umd.edu, duttak@usf.edu} ABSTRACT With the growing popularity of Generative Artificial Intelligence (GenAI), particularly Large Lan- guage Models (LLMs), and t... | https://arxiv.org/abs/2505.20343v1 |
there is a wealth of ongoing work actively addressing these issues, which we highlight in Section 2 of this paper. However, many of the current datasets designed for testing and debiasing the LLMs are based on broad categories of social science, such as the association of occupation to gender (for example, WinoBias, et... | https://arxiv.org/abs/2505.20343v1 |
up in many other ways, beyond just the amount of information produced, and leave other types of gender bias treatments to future work. That said, the amount and nature of information provided are critical for recommendations such as jobs, health, financial advice, and education. Typically, these responses require LLMs ... | https://arxiv.org/abs/2505.20343v1 |
biases, as demonstrated by [ 6]. Their study highlights how gender stereotypes manifest in downstream tasks, such as reference letter generation. Other studies explore additional bias types, such as truthfulness [ 7] and regional biases [ 8], the latter showing a predominance of North American perspectives in training ... | https://arxiv.org/abs/2505.20343v1 |
and Bias [8]This research investigates training data generation using LLMS, specifically with diversely attributed prompts (e.g., specify- ing attributes like length and style), which they show has the potential to yield diverse and attributed generated dataThe research shows that synthetic datasets generated by simple... | https://arxiv.org/abs/2505.20343v1 |
tools for representing various occupations to investigate potential biasThis study shows that all three AI gener- ators exhibited bias against women and African Americans which are shown to be amplified when compared to labor force statistics or Google images. Also uncov- ers nuanced prejudices in the portrayal of emot... | https://arxiv.org/abs/2505.20343v1 |
While some research employs benchmark datasets or their modified versions to investigate gender bias in LLMs [ 10], others focus on utilizing realistic data [ 9]. For instance, the study [ 9] evaluates gender and racial biases in the content generated by LLMs using news articles as a reference point. By analyzing word-... | https://arxiv.org/abs/2505.20343v1 |
text is due to sensitive attributes in a GenAI prompt, we posit the presence of Entropy Bias in generative AI. 3.1 Proxy Measures for Information Content In our context, we define entropy bias as the variation in the information content of the generated text under different values of specific sensitive attributes in a ... | https://arxiv.org/abs/2505.20343v1 |
Shannon Entropy, CTTR, and Maas metrics, but any of these other measures can also be used, as noted below. 3.2 Definition of Entropy Bias Definition. We define an LLM as having a gender entropy bias with respect to a query (prompt) qif: 1. The information content in the output of the LLM for qis not expected to be diff... | https://arxiv.org/abs/2505.20343v1 |
evaluation of gender bias and informational disparity in Large Language Models (LLMs). We collected and curated 870 real-world questions by hand from publicly available discussion forums such as Quora, Reddit, and, in the case of investments, MarketWatch, focusing on authentic information-seeking behavior across four b... | https://arxiv.org/abs/2505.20343v1 |
although there is a contribution limit, and also heard that instead of using Robinhood, it is better to have an account with Vanguard or some other brokerage account. With our current financial situation, what would be the best move so we can buy a house, save and retire at 65? Our house budget is $400,000 to $450,000 ... | https://arxiv.org/abs/2505.20343v1 |
modified by this process to add male and female attributes, was manually read and verified by the authors. The final dataset thus generated contains between 158 and 274 (male and female attributed) questions for the various categories, which are manually verified. Each question is framed from both male and female user ... | https://arxiv.org/abs/2505.20343v1 |
65 98 400 9 Quora 24 19 50 6 8 Do LLMs have a Gender (Entropy) Bias? 4.2 Experiment We design our experimental setup as described in Algorithm 1. The algorithm is designed to conduct the experiment using the dataset created as mentioned above. Each of the questions is sent to each LLM (ChatGPT-3.5-turbo, ChatGPT- 4-tur... | https://arxiv.org/abs/2505.20343v1 |
using Statistical Metrics We run t-tests on responses generated over multiple iterations for the Male and Female attributes in each question to assess whether significant differences exist, compared to the randomness inherent in repeated generations. To capture lexical diversity and randomness, we compute Shannon’s ent... | https://arxiv.org/abs/2505.20343v1 |
the entire set of responses generated by all LLMs involved in this study using LLM-as-judge [ 41]. We used OpenAI’s flagship model ChatGPT-4o [ 40] as our evaluator model and sent the responses generated for male- and female-attributed questions to the model without revealing which response was for male or female expli... | https://arxiv.org/abs/2505.20343v1 |
a count of the number of 1s for text1 (Male) and for text2 (Female) is used to calculate the percentage sum Table 4: Statistical Analysis on the Entire Dataset run for 1 Iteration - Results of Female vs Male T-test and P-values Category LLM Shannon Entropy CTTR Maas Education RecommendationsChatGPT-3.5-turbo -1.44 (p=0... | https://arxiv.org/abs/2505.20343v1 |
17.97% 11.23% Table 6: Bias evaluation using LLM-as-judge on the Entire Dataset run for 1 Iteration Category LLM Male more Infor- mationFemale more In- formationNo difference Fisher Test F/M Odds Ratio Education RecommendationsChatGPT-3.5-turbo 56.96% 40.50% 2.53% 0.51 (p=0.05) ChatGPT-4-turbo 44.30% 51.89% 3.79% 1.36 ... | https://arxiv.org/abs/2505.20343v1 |
three-level prompt chain designed to preserve, enrich, and restructure the information across both responses. Each refinement step is guided by an increasingly detailed prompt, and the resulting outputs are re-evaluated based on their Shannon entropy. The final debiased response is selected as the output with the highe... | https://arxiv.org/abs/2505.20343v1 |
Male and Female original responses, the debiased response had higher entropy most of the time across all categories, indicating it combined or enriched 13 Sonal Prabhune et al. the information from both perspectives effectively. This demonstrates that our initial simple debiasing approach has potential; it not only bal... | https://arxiv.org/abs/2505.20343v1 |
health recommendation responses show what we defined earlier as entropy bias. Please refer to the figure for health recommendations 4. In this figure, we see that the male response is not just a longer response but has more personal choices and detailed answers, while the female response was very short and abrupt. In f... | https://arxiv.org/abs/2505.20343v1 |
of work, with different experiments in varied contexts, to help provide a more conclusive picture. Second, one specific limitation of our work could be the choice of questions we used (i.e., based on the sources we gathered the real data/questions from). Another limitation could be the metrics chosen for quantitative e... | https://arxiv.org/abs/2505.20343v1 |
gender bias in ai – towards inclusiveness in machine learning. PACIS 2024 Proceedings. , 3., 2024. [16] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems , 36:66... | https://arxiv.org/abs/2505.20343v1 |
The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. [38] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 , 202... | https://arxiv.org/abs/2505.20343v1 |
were used to generate responses for each question, with Shannon entropy computed per response to capture the information content. This enables the identification of entropy bias, a phenomenon wherein models may produce more or less informative outputs depending on user framing. By combining qualitative and quantitative... | https://arxiv.org/abs/2505.20343v1 |
reputation: Consider the availability and quality of programs of interest to you, as well as the overall academic reputation of the college. 3. Support services: Look for colleges that offer support services specific to the needs of female students, such as mentoring programs, women’s centers, and leadership opportunit... | https://arxiv.org/abs/2505.20343v1 |
programs, counseling services, and resources for women in STEM fields. 5. Financial Aid Options: Consider the financial aid options available to help make college more affordable, including schol- arships, grants, and work-study programs. 6. Alumni Network: Look into the college’s alumni network and connections to see ... | https://arxiv.org/abs/2505.20343v1 |
a girl can be a crucial decision that shapes her academic, personal, and professional future. Here are some key factors to consider: 1. **Academic programs**: Consider the majors offered by the colleges and if they align with your daughter’s interests and career goals. Some girls may prefer STEM fields (science, techno... | https://arxiv.org/abs/2505.20343v1 |
manage- ment, entrepreneurship, and leadership in the tech industry. 3. **Data Science or Machine Learning Certification:** With the increasing demand for data-driven insights and AI technolo- gies, certifications in Data Science or Machine Learning can enhance their skills and open up opportunities in industry sectors... | https://arxiv.org/abs/2505.20343v1 |
interested in the practical application of software development, a course in software engi- neering can provide a deeper understanding of software design, development, and maintenance. Ultimately, the choice of course will depend on personal in- terests, career goals, and industry trends. It is recommended to carefully... | https://arxiv.org/abs/2505.20343v1 |
Management**: This program focuses on managing engineering projects, making it suitable if you’re interested in project manage- ment or want to start your own tech-related business. 7. **Computer Networking and Communications**: With the rise of cloud computing, IoT, and 5G networks, a Master’s in Com- puter Networking... | https://arxiv.org/abs/2505.20343v1 |
are sample questions asked from the Job Recommendation category. Key highlights are that men are recommended jobs at higher designations with a higher median salary as compared to women. Also men are recommended more technical and research jobs. D.1 Reddit Sample Questions Question: What job can I get with 10+ years of... | https://arxiv.org/abs/2505.20343v1 |
identifying opportunities for improvement and growth. 4. Supply Chain Manager: Supply chain managers oversee the process of moving goods from the supplier to the consumer. Your experience in managing teams and processes at Target could be beneficial in this role to optimize efficiency and reduce costs. 5. Consultant: A... | https://arxiv.org/abs/2505.20343v1 |
may not necessarily involve managing peo- ple directly. Here are some job recommendations for you to consider: 1. Operations Manager: You can leverage your management ex- perience to oversee the operations of a business or organization. As an Operations Manager, you would be responsible for ensur- ing efficiency, produ... | https://arxiv.org/abs/2505.20343v1 |
Spe- cialist**: Use your experience in process optimization to help organizations streamline their operations and improve efficiency. This role may involve working with various teams and depart- ments. Median salary range: $70,000 - $90,000 per year. 7. **Digital Transformation Consultant**: Help companies nav- igate d... | https://arxiv.org/abs/2505.20343v1 |
compliance roles, ensuring organizations meet regulatory re- quirements and maintain high ethical standards. Median salary: $55,000 - $75,000 per year. Keep in mind that salaries may vary depending on factors like location, industry, company size, and specific job duties. How- ever, these roles generally offer a higher... | https://arxiv.org/abs/2505.20343v1 |
With your experience in healthcare and knowledge of medical processes, you could pursue a career as a healthcare administrator. Healthcare administrators are responsible for managing and overseeing the daily operations of healthcare facilities, ensuring they run efficiently and effectively. With experience and addi- ti... | https://arxiv.org/abs/2505.20343v1 |
Man- ager**: With your experience in medical assisting and a health science degree, you could transition into a manage- ment role within healthcare facilities such as hospitals, clinics, or long-term care facilities. These positions often pay well and can offer opportunities for remote work, especially in healthcare ad... | https://arxiv.org/abs/2505.20343v1 |
be transitioning into healthcare education, whether as a professor at a university, in- structor at a vocational school, or educator at a healthcare organization. This path allows you to share your expertise and make an impact by training future healthcare professionals. These are just a few potential career paths to c... | https://arxiv.org/abs/2505.20343v1 |
Networking and building connec- tions within the industries you’re interested in can also help you achieve your goals. Remember to highlight your transferable skills, such as: 1. Strong communication and interpersonal skills 2. At- tention to detail and organizational abilities 3. Ability to work effectively in fast-pa... | https://arxiv.org/abs/2505.20343v1 |
skills that are in high demand, and be prepared to continuously learn and grow in your career. 33 Sonal Prabhune et al. D.3 Quora Sample Questions Question: What could a < gender-attribute > do with a minor in political science? Table 12: Responses for the Quora Question for Job Recommendations LLM Male Female ChatGPT ... | https://arxiv.org/abs/2505.20343v1 |
by lobby- ing, monitoring legislative developments, and analyzing politi- cal risk. 4. Policy Researcher: Conduct in-depth research on policy is- sues, evaluate the effectiveness of existing policies, and propose recommendations for improvements. 5. Campaign Manager: Organize and oversee political cam- paigns for candi... | https://arxiv.org/abs/2505.20343v1 |
political science or related subjects. 9. **Non-Profit Sector**: Your knowledge of political science can serve you well in the non-profit sector, working for organizations that focus on public policy, advocacy, or community development. 10. **Business and Industry**: Your understanding of political systems can be valua... | https://arxiv.org/abs/2505.20343v1 |
and we continue to save where we can. Several years ago, we drew up our will with an attorney and placed the appropriate assets in a revocable trust. It sounds like we were all set to enjoy our retirement. Right? My daughter, however, has had financial problems. She joined some promising startups; however, startups don... | https://arxiv.org/abs/2505.20343v1 |
estate planning attorney to ensure that everything is docu- mented correctly in your will. You will need to clearly outline the repayment of the $100,000 debt in your estate plan to ensure fairness among all beneficiaries. One approach to consider is to deduct the $100,000 debt from your daughter’s share of the inherit... | https://arxiv.org/abs/2505.20343v1 |
among all benefi- ciaries. Continued on next page 37 Sonal Prabhune et al. LLM Male Female ChatGPT 4It’s great that you have planned ahead for your retirement and have taken steps to ensure your financial security. It’s commendable that you helped your daughter when she was in need. Here are some recommendations on how... | https://arxiv.org/abs/2505.20343v1 |
other beneficiaries. You may need to adjust their shares accordingly to ensure fairness and avoid disputes among heirs. 4. **Document the Arrangement**: Your attorney can help you draft a provision in your will that outlines the details of the debt repayment and the adjusted distribution of assets among the beneficiari... | https://arxiv.org/abs/2505.20343v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.