text string | source string |
|---|---|
TeroSeek: An AI -Powered Knowledge Base and Retrieval Generation Platform for Terpen oid Research Xu Kang, Siqi Jiang, Kangwei Xu, Jiahao Li, Ruibo Wu * School of Pharmaceutical Sciences , Sun Yat -sen University, Guangzhou 510006, P.R. China * E-mail : wurb3@mail.sysu.edu.cn Abstract Terpenoids repre sent a pivotal cl... | https://arxiv.org/abs/2505.20663v1 |
remains static until the next training update, potentially leading to outdated information11,12. Hallucination Risks: While the probabilistic nature of LLM outputs enables extrapolation, it also introduces inaccuracies13-15. In critical fields, such errors could have catastrophic consequences16. Lack of Attribution: Ev... | https://arxiv.org/abs/2505.20663v1 |
all article types . Subsequently, LLMs were employed to read the abstracts of candidate papers to determine whether the article topics were related to terpenoids (Figure 1a). Subsequently, we collected 47,731 academic documents, which included 44,701 research papers and 3,030 literature reviews. All metadata such as jo... | https://arxiv.org/abs/2505.20663v1 |
enhance the retrieval capabilities of the knowledge base, the research module implements a more advanced workflow. For user -specified topics, the module adopts a two-stage retrieval strategy. Initially, retrieval is conducted solely within the review - specific knowledge base to establish a comprehensive understanding... | https://arxiv.org/abs/2505.20663v1 |
of paclitaxel whatsoever —it solely addresses terpenoids in the context of breast cancer, likely resulting from a retrieval error. Figure 3: Response Demonstration Case - Knowledge Q&A For the question: "What is the target of paclitaxel?", the upper section displays the output from TeroSeek -normal, while the lower sec... | https://arxiv.org/abs/2505.20663v1 |
and Qwen2 -235B -A22B rose from 0.46 to 0.74, as illustrated in Figure 4c. Based on these results, we selected DeepSeek -R1 as the backbone model for our web service, TeroSeek -Normal. We subsequently benchmarked TeroSeek -normal against other lead ing language models using the same test set. The comparative results sh... | https://arxiv.org/abs/2505.20663v1 |
retrieval. We've incorporated data from over 40,000 literature abstracts and conducted thoro ugh full -text analysis of research papers, ensuring no detail was overlooked. The system surpasses existing LLMs in answering terpenoid -specific queries while maintaining academic integrity through citation -attached response... | https://arxiv.org/abs/2505.20663v1 |
Compounds with Several Health Functions and Industrial Applications —A Comprehensive Overview. Molecules 29, 3861, doi:10.3390/molecules29163861 (2024). 8 Patel, T., Ishiuji, Y . & Yosipovitch, G. J. J. o. t. A. A. o. D. Menthol: a refreshing look at this ancient compound. 57, 873 -878 (2007). 9 Christopher, I. K. & Jö... | https://arxiv.org/abs/2505.20663v1 |
Research 53, D30 -D44, doi:10.1093/nar/gkae978 %J Nucleic Acids Research (2024). 30 Zhang, R. et al. PlantGPT: An Arabidopsis -Based Intelligent Agent that Answers Questions about Plant Functional Genomics. n/a, e03926, doi:https://doi.org/10.1002/advs.202503926 . 31 Shuster, K., Poff, S., Chen, M., Kiela, D. & Weston,... | https://arxiv.org/abs/2505.20663v1 |
arXiv:2505.20664v1 [cs.CL] 27 May 2025Self-Route: Automatic Mode Switching via Capability Estimation for Efficient Reasoning Yang He, Xiao Ding, Bibo Cai, Yufei Zhang, Kai Xiong Zhouhao Sun, Bing Qin, Ting Liu Research Center for Social Computing and Interactive Robotics Harbin Institute of Technology, China {yhe, xdin... | https://arxiv.org/abs/2505.20664v1 |
of general models, we propose the Self-Route frame- work. The core objective of this framework is to utilize the Router to select the most efficient rea- soning strategy while ensuring correctness. Our ap- proach introduces a lightweight pre-inference stage prior to formal reasoning. By collecting hidden layer represen... | https://arxiv.org/abs/2505.20664v1 |
output. To effectively train a router that could accurately detect the capability boundaries of different models, we carefully constructed a densely graded difficulty dataset, Gradient-10K , based on model difficulty estimation from multiple data sources. Subsequent experiments demonstrate that the dense difficulty gra... | https://arxiv.org/abs/2505.20664v1 |
Language models can estimate both the difficulty of a question and their own ability to answer it during the generation process (Ashok and May, 3 2025). This pre-inference stage enables the model to perform such an assessment using only a small number of tokens, allowing for an efficient capabil- ity estimation before ... | https://arxiv.org/abs/2505.20664v1 |
cross- entropy loss: L=−NX i=1[yilog(ˆyi) + (1 −yi) log(1 −ˆyi)](5) where yiis the true solution correctness label, andˆyiis the predicted label. 3 Experiments 3.1 Experimental Setup We evaluate the effectiveness of the Self-Route method on models with various parameter config- urations and inference modes. The general... | https://arxiv.org/abs/2505.20664v1 |
datasets with different levels of complexity, includ- ing Orca Math (Mitra et al., 2024), AMC AIME, Olympiads, GSM8K (Cobbe et al., 2021), and s1k (Muennighoff et al., 2025).Evaluation Dataset. For the test datasets, we selected a variety of commonly used mathematical test datasets with different difficulty levels, dom... | https://arxiv.org/abs/2505.20664v1 |
the necessity of con- structing a densely annotated difficulty gradientModel Acc. Prec. F1 Router on Gradient Qwen2.5-7B Gradient-Router 0.81 0.83 0.86 Qwen2.5-32B Gradient-Router 0.84 0.89 0.89 Qwen3-8B Gradient-Router 0.83 0.88 0.89 Router on GSM8K Qwen2.5-7B GSM8K-Router 0.70 0.70 0.80 Qwen2.5-32B GSM8K-Router 0.77 ... | https://arxiv.org/abs/2505.20664v1 |
CoT reasoning, respec- tively. "Pre" indicates the number of tokens consumed by the pre-inference module, and "Ratio" denotes the percentage of Pre over Long. sumption across datasets of varying difficulty levels and domains, as well as the ratio of pre-inference token usage relative to that of the reasoning model. Thi... | https://arxiv.org/abs/2505.20664v1 |
Related Work Recent studies on Model-based Efficient Reason- ing have primarily focused on fine-tuning large language models (LLMs) to improve their intrin- sic ability to reason concisely and efficiently. Ex- isting works leverage traditional RL optimization techniques combined with explicit length-based re- ward to c... | https://arxiv.org/abs/2505.20664v1 |
2025. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697 . Daman Arora and Andrea Zanette. 2025. Training lan- guage models to reason efficiently. arXiv preprint arXiv:2502.04463 . Dhananjay Ashok and Jonathan May. 2025. Lan- guage models can predict their own... | https://arxiv.org/abs/2505.20664v1 |
2025. URL https://arxiv. org/abs/2502.17419 . Yuxin Liang, Zhuoyang Song, Hao Wang, and Jiax- ing Zhang. 2024. Learning to trust your feelings: Leveraging self-awareness in llms for hallucination mitigation. arXiv preprint arXiv:2401.15449 . 9 Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Bake... | https://arxiv.org/abs/2505.20664v1 |
own problem-solving capacity. When applied to data of the same dif- ficulty level, larger models (e.g., Qwen2.5-32B) tend to classify a greater number of questions as within their solvable range, compared to smaller models such as Qwen2.5-7B. This suggests that more powerful models possess a more confident self-assessm... | https://arxiv.org/abs/2505.20664v1 |
arXiv:2505.20674v1 [cs.CL] 27 May 2025Pretraining Language Models to Ponder in Continuous Space Boyi Zeng1, Shixiang Song1,2,3, Siyuan Huang1, Yixuan Wang1,3, He Li1, Ziwei He3,Xinbing Wang4, Zhiyu Li2,Zhouhan Lin1,2,3∗ 1LUMIA Lab, Shanghai Jiao Tong University, 2Institute for Advanced Algorithms Research, Shanghai, 3S... | https://arxiv.org/abs/2505.20674v1 |
0.10.7 ⋯ 0.1 ⨁⨁ ⨁ ⨁⨁ ⨁ Input EmbeddingPondering EmbeddingPredicted Probability Pondering Embeddingclass PonderingLanguageModel(nn.Module): def __init__(self, lm, v, h, k): self.lm = lm # language model self.vocab_size = v self.hidden_dim = h self.pondering_steps = k self.embedding = nn.Parameter(torch. randn(v, h), req... | https://arxiv.org/abs/2505.20674v1 |
Given that pretraining fundamentally constitutes a language modeling task, we briefly review this task before detailing our proposed model. Language Modeling. Given a sequence of tokens X= [x1, x2, . . . , x n], the primary objective of language modeling is typically to maximize the likelihood of predicting each token ... | https://arxiv.org/abs/2505.20674v1 |
capabilities with those of the official Pythia suite (Biderman et al., 2023). Third, we evaluate the downstream task performance of PonderingPythia models, including nine popular general tasks and an instruction-following task, and compare the results with official Pythia, OPT (Zhang et al., 2022), Bloom (Le Scao et al... | https://arxiv.org/abs/2505.20674v1 |
training tokens or 52% of the parameters. Table 2: Language Modeling Perplexity (ppl). Lower is better. The values in parentheses indicate the improvement ( ↓) compared to the corresponding Pythia model. For simplicity, we refer to PonderingPythia as Ponder. Model Pile Wikitext Lambada Openai Lambada Standard Pythia-14... | https://arxiv.org/abs/2505.20674v1 |
all datasets and model sizes. Notably, the perplexity achieved by the PonderingPythi-70M model is comparable to that of the much larger official Pythia- 160M model. 3.3 Downstream Tasks Evaluation We use the PonderingPythia models pretrained in the previous subsection to conduct downstream task evaluations. 3.3.1 Gener... | https://arxiv.org/abs/2505.20674v1 |
60.2 64.7 55.0 29.8 62.2 75.1 46.1 93.0 37.5 58.2 GPTneo-2.7B (300B) 56.0 64.0 51.6 30.1 59.6 73.9 42.4 93.3 35.5 56.3 Bloom-3B (366B) 46.2 63.8 47.1 31.7 57.8 70.8 41.4 93.4 34.6 54.1 Pythia-2.8B (300B) 59.0 67.0 50.7 31.0 61.1 74.4 45.3 93.7 35.9 57.6 0-shot Pythia-70M (300B) 18.7 36.7 13.5 18.8 51.1 59.9 26.6 59.9 2... | https://arxiv.org/abs/2505.20674v1 |
modeling loss on the Pile validation set. This demonstrates the significant potential and scala- bility of our method, suggesting that model per- formance can be further improved by increasing the depth of pondering steps. 4 Limitations and Future Work 4.1 Limitations There are two limitations to our work. Firstly, due... | https://arxiv.org/abs/2505.20674v1 |
al., 2025; Li et al., 2025). Following the emergence of powerful OpenAI o1 and DeepSeek R1, leveraging extensive chain-of-thought prompting to scale test-time computation has become increasingly popular. Nonetheless, existing methods often come with limitations, including reliance on specially curated datasets (Allen-Z... | https://arxiv.org/abs/2505.20674v1 |
S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning , pp. 2397–2430. PMLR, 2023. Biran, E., Gottesman, D., Yang, S., Geva, M., and Globerson, A. Hopping too late: Exploring the limitations of large langua... | https://arxiv.org/abs/2505.20674v1 |
sweetness of best-of-n sampling. arXiv preprint arXiv:2406.00832 , 2024. Hackenburg, K., Tappin, B. M., Röttger, P., Hale, S. A., Bright, J., and Margetts, H. Scaling language model size yields diminishing returns for single-message political persuasion. Proceedings of the National Academy of Sciences , 122(10):e241344... | https://arxiv.org/abs/2505.20674v1 |
In Proceedings of the 8th Asia-Pacific Workshop on Networking , pp. 1–8, 2024. 11 Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreff... | https://arxiv.org/abs/2505.20674v1 |
Stroebl, B., Kapoor, S., and Narayanan, A. Inference scaling flaws: The limits of llm resampling with imperfect verifiers. arXiv preprint arXiv:2411.17501 , 2024. Su, J., Ahmed, M., Lu, Y ., Pan, S., Bo, W., and Liu, Y . Roformer: Enhanced transformer with rotary position embedding. Neurocomputing , 568:127063, 2024. S... | https://arxiv.org/abs/2505.20674v1 |
2022. Zheng, L., Chiang, W.-L., Sheng, Y ., Zhuang, S., Wu, Z., Zhuang, Y ., Lin, Z., Li, Z., Li, D., Xing, E., et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems , 36:46595–46623, 2023. Zhou, J., Pang, L., Shen, H., and Cheng, X. Think before you speak: Cul... | https://arxiv.org/abs/2505.20674v1 |
arXiv:2505.20679v1 [cs.CL] 27 May 2025SELF-PERCEPT: Introspection Improves Large Language Models’ Detection of Multi-Person Mental Manipulation in Conversations Danush Khanna1Pratinav Seth2,3Sidhaarth Murali4 Aditya Guru1Siddharth Shukla1Tanuj Tyagi1 Sandeep Chaurasia1Kripabandhu Ghosh5 1Manipal University Jaipur, Indi... | https://arxiv.org/abs/2505.20679v1 |
tactics that emerge in group dynam- ics. These limitations raise a crucial question: Can large language models effectively identify vari- ous manipulation techniques in complex, multi- turn, multi-participant dialogues that resemble real-world conversations? To answer this question, we conducted extensive experiments u... | https://arxiv.org/abs/2505.20679v1 |
use strategic and manipulative tactics, 1https://www.fandom.com/ 2https://www.imdb.com/title/tt0239195/ making it a valuable source of relevant data. Addi- tionally, the diversity of contestants from various backgrounds allows us to observe a broad range of manipulative techniques and responses. We gather transcripts f... | https://arxiv.org/abs/2505.20679v1 |
prompting, which focuses on stepwise reasoning, SELF-PERCEPT’s Stage 1 ex- plicitly extracts behavioral cues (verbal/non-verbal) to infer latent attitudes. For example, detecting a sigh (non-verbal) alongside agreement (verbal) re- veals passive-aggressive intent. This aligns with Self-Perception Theory (Fazio, 2014), ... | https://arxiv.org/abs/2505.20679v1 |
3.1 8B ,SELF-PERCEPT again leads in F1 score (0.34) and Accuracy (0.30), offering a better balance of Precision (0.17) and Recall (0.26) than other methods. For theMentalManip Dataset (C.f. Table 2), SELF-PERCEPT delivers the best performance in GPT-4o with an Accuracy of 0.45 and F1 score of 0.47. It outperforms Zero-... | https://arxiv.org/abs/2505.20679v1 |
on transcripts from the Survivor TV Series on Fandom, provides real examples of manipulative behavior in competitive, multi-turn dialogues. However, it may not fully capture the range of real-world conversa- tional nuances where manipulation can be even less predictable. While it includes genuine interactions with vari... | https://arxiv.org/abs/2505.20679v1 |
Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunx- ing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Do- han, Steve Dowling, Sheila Dunning, Adrien Ecof- fet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Sim’on Posada Fishman, Jus- ton ... | https://arxiv.org/abs/2505.20679v1 |
and Barret Zoph. 2023. Gpt- 4 technical report. None . Anne Barnhill. 2014. What is manipulation? In Ma- nipulation: Theory and Practice . Oxford University Press. Bobby J Calder and Barry M Staw. 1975. Self- perception of intrinsic and extrinsic motivation. Jour- nal of personality and social psychology , 31(4):599. G... | https://arxiv.org/abs/2505.20679v1 |
Collot, Suchin Gururangan, Syd- ney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mihaylov, Tong Xiao, Ujjwal Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent Gonguet, Vir- ginie Do, Vish V ogeti, Vladan Petrovic, Weiwei C... | https://arxiv.org/abs/2505.20679v1 |
Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pa- van Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollár, Polina Zvyagina, Prashant Ratan- chandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Raghotham Murthy, Raghu Nayani, Rahul Mitra, Raymond Li, Rebek... | https://arxiv.org/abs/2505.20679v1 |
theory. Journal of Personality and Social Psychology , 28(1):138.Matteo Antonio Senese, Giuseppe Rizzo, Mauro Drag- oni, and Maurizio Morisio. 2020. MTSI-BERT: A session-aware knowledge-based conversational agent. InProceedings of the Twelfth Language Resources and Evaluation Conference , pages 717–725, Mar- seille, Fr... | https://arxiv.org/abs/2505.20679v1 |
a predefined set of 12 possible labels to each item. The 12 labels include 11 manipulation methods and 1 for non- manipulative. The resulting Fleiss’ Kappa value was0.429 , indicating moderate agreement (0.41 - 0.60) between the annotators.A.2 Dataset Curation and Pre-Processing To ensure effective processing, the tran... | https://arxiv.org/abs/2505.20679v1 |
manipulation, as real-world interactions often blend genuine conversation with manipula- tive tactics. By capturing this complexity, the dataset mirrors the varied and multi-turn nature of manipulation in everyday life.A.4 Inferences drawn from Annotation The final labels were determined by combining an- notations from... | https://arxiv.org/abs/2505.20679v1 |
SeductionThe manipulator uses charm, emotional appeal, or logical reasoning to lower the victim’s defenses. Table 3: Definitions of the 11 manipulation techniques. Person Conversation (Dialogue) Person 1 I do believe the reasons are starting to change. Person 2 What do you mean by that? Person 1 I really think before, ... | https://arxiv.org/abs/2505.20679v1 |
: Who says I gave up? Person C : If you get knocked back, you stand back up and you take another knock in the mouth. Person A : And you know– and I don’t want to start another argument between you two guys– but at times– and I said this to you, Anthony– I felt that in his own twisted way, Rocky was actually trying to n... | https://arxiv.org/abs/2505.20679v1 |
manipulative behavior. Person1: Next two, next two. Person2: We are tied 11. Person1: First tribe to complete the tem- ple doesn’t have to go to the temple. Person3: Way to work together, guys. Person2: You guys have already solved the fire. Person1: We have the key, we have the map. Person2: (laughs) That’s right, we ... | https://arxiv.org/abs/2505.20679v1 |
more of the following categories: <Insert Definitions of the 11 manipulation techniques here (See Table 3> 3. Provide a comma-separated list of the abbreviations for all manipulation types detected. Do not add anything else. Dialogue to analyze: dialogue Manipulation Techniques: ZERO SHOT PROMPT I will provide you with... | https://arxiv.org/abs/2505.20679v1 |
to work together. Person A: You can’t just dictate to people what to do.> <FEI, SER, S_B> Example 2: <Person A : Anthony, as we get ready to go to the vote, how uncomfortable are you that it might be you? Person B : Fairly uncomfortable. I mean, my personality does clash with some people, and they may think that I’m tr... | https://arxiv.org/abs/2505.20679v1 |
is derived from publicly available re- ality show transcripts. All participant references have been replaced with generic labels (e.g., “Per- son A,” “Person B,” “Person C”) to remove any direct personal identifiers. We did not make addi- tional modifications to censor or remove offensive content, as the dialogue is in... | https://arxiv.org/abs/2505.20679v1 |
manipula- tion tactics become more advanced and common in personal interactions and online spaces, being able to spot and address these behaviors is crucial for protecting individuals and communities. This research, which leverages the MultiManip dataset and theSELF-PERCEPT framework, is a key part of this effort. Firs... | https://arxiv.org/abs/2505.20679v1 |
Users must have the option to opt-out or provide consent. Accountability and Regulation : Accountability measures and external audits are needed to prevent misuse. Governments should establish ethical guidelines to safeguard pri- vacy and civil liberties. Bias Mitigation : To avoid discrimination, detection systems sho... | https://arxiv.org/abs/2505.20679v1 |
arXiv:2505.20692v1 [cs.HC] 27 May 2025Can we Debias Social Stereotypes in AI-Generated Images? Examining Text-to-Image Outputs and User Perceptions Saharsh Barve1, Andy Mao1, Jiayue Melissa Shi1, Prerna Juneja2, Koustuv Saha1 1University of Illinois Urbana-Champaign,2Seattle University ssbarve2@illinois.edu, hanqim2@il... | https://arxiv.org/abs/2505.20692v1 |
misinformation, erode trust in AI systems, and distort public perceptions (Zhou et al. 2023; Zhang et al. 2024). Despite these concerns, our understand- ing of how to effectively identify and mitigate these biases in T2I systems remains limited. Most current approaches rely on ad-hoc audits, case studies, or red-teamin... | https://arxiv.org/abs/2505.20692v1 |
underscores the challenge of balancing ethical representation with user ex- pectations. Our study makes the following contributions: 1. A theory-driven rubric to quantify social bias in gen- erated images. At its core is the Social Stereotype In- dex ( SSI), a novel metric that systematically captures and compares ster... | https://arxiv.org/abs/2505.20692v1 |
structured rubric- based questionnaire designed to audit AI-generated outputs. Complementing this with a qualitative study, we elicit end- users’ mental image to examine how their expectations align with AI-generated images and how stereotypical representa- tions are internalized or resisted in practice. Social Bias in... | https://arxiv.org/abs/2505.20692v1 |
in practice—exhibiting unexpected be- haviors, biases, and harms ranging from misinformation, stereotyping, discrimination, exclusion, and erosion of au- tonomy (Mittelstadt et al. 2016; Sandvig et al. 2014; Floridi et al. 2018; Raji et al. 2022; Ghosh and Caliskan 2023). En- suring that these AI operates as intended r... | https://arxiv.org/abs/2505.20692v1 |
tural ,occupational , and adjectival categories, we generated 1,200 images from three state-of-the-art models: DALL-E-3, Midjourney-6.1, and Stability AI Core. These models span diverse architectures and training approaches. Rather than directly comparing model performance, our primary goal is to evaluate a broad spect... | https://arxiv.org/abs/2505.20692v1 |
ster eotypes Expectation al ignment Concerns and desir esFigure 1: Overview of our study design for identifying and mitigating social stereotypes in T2I output. queries in the format, photo of a [A] person, where [A]in- cludes adjectives like, rude, beautiful, smart , etc. Generating T2I: To generate image data for ana... | https://arxiv.org/abs/2505.20692v1 |
generated using our prompt refinement approach (e.g., Ap- pendix Figure A2). They chose their preferred set and briefly explained their reasoning. This comparison served two main purposes—1) to assess the effectiveness of prompt refine- ment by revealing whether users consistently favored refined outputs. 2) to gather ... | https://arxiv.org/abs/2505.20692v1 |
Equation 1). Essentially, SSIranges between 0 and 1, where 0 indicates no stereotypical bias in an image, and higher val- ues indicate a greater presence of stereotypical bias. SSI =1 NNX i=1xi,where xi=1,if stereotype present for item i 0,otherwise (1) LLM-powered Automated Evaluations of T2I Outputs. Next, we employ... | https://arxiv.org/abs/2505.20692v1 |
in the original image set, 2) maintain the visual coherence and avoid fragmented, collage-like outputs ob- served in our observations, and 3) preserve the original in- tent and meaning present in the initial query. Through mul- tiple rounds of experiments and discussions among the re- search team, we found that the mos... | https://arxiv.org/abs/2505.20692v1 |
observations, and captured participants’ perspec- tives on stereotypes, preferences, and expectations related to AI-generated images. Then, we employed a micro-board affinity diagramming to organize the codes, enabling us to cluster insights and identify emerging patterns across par- ticipants. Finally, we applied refl... | https://arxiv.org/abs/2505.20692v1 |
gender (17.8% →8.1%), and race/ethnicity (11.4%→4.0%) (ref: Figure 2). We observed that occupational queries often revealed gen- der and racial biases in initial T2I outputs. For example, men were predominantly depicted in corporate or leadership roles (e.g., a CEO, a manager ), often accompanied by stereotyp- ical ele... | https://arxiv.org/abs/2505.20692v1 |
reinforce societal stereotypes, such as beauty standards and the over- representation of features associated with Western cul- tures. For example, queries such as furious andrude led to male-presenting individuals, whereas beautiful predom- inantly generated white-skinned women—reflecting narrow cultural norms and a la... | https://arxiv.org/abs/2505.20692v1 |
beauty. However, upon viewing the refined image set with more diverse representations, they came to recognize and appreciate the value of inclusive imagery: “The [diverse beautiful representation] is better because there is diversity in different ways. The 2nd lady is in a suit- /formal clothes. The 1st one is in tradi... | https://arxiv.org/abs/2505.20692v1 |
stereotypical representations. We noted that some stereotypes can in- ternally be rooted in a deeply subjective process filtered through an individual’s real-life interactions and relation- ships. Individuals who had direct personal connections to the groups or professions represented often exhibited height- ened sensi... | https://arxiv.org/abs/2505.20692v1 |
sec- ond section of the interviews—the rapid-fire comparison task—participants were asked to compare initial and refined outputs corresponding to a set of queries one by one. Inter- estingly, we observed a relatively balanced distribution of preferences between the initial and refined T2I outputs. The 17 participants d... | https://arxiv.org/abs/2505.20692v1 |
reduction of stereotyped outputs. This suggests that the same technique may lead to even greater improvements for less-moderated or fine-tuned models—such as early-stage commercial deployments or third-party applications—where moderation is minimal or opaque. In these “black-box” settings, models often priori- tize fid... | https://arxiv.org/abs/2505.20692v1 |
tools. For instance, independent audits, explainability standards, and enforceable fairness benchmarks are needed to ensureaccountability—particularly when systems shape perception subtly and at scale. As some of our interview findings sug- gest, in the absence of such safeguards, generative models may shape public ima... | https://arxiv.org/abs/2505.20692v1 |
steps to respect cultural sensitivities, such as using inclusive language and allow- ing participants to skip or rephrase prompts they found un- comfortable. Our interdisciplinary research team comprises individuals with diverse gender, racial, and cultural back- grounds, including people of color and immigrants, and h... | https://arxiv.org/abs/2505.20692v1 |
Heidari, H. 2023. A validity perspective on evaluating the justified use of data- driven decision-making algorithms. In 2023 IEEE conference on secure and trustworthy machine learning (SaTML) .Das Swain, V .; and Saha, K. 2024. Teacher, trainer, counsel, spy: How generative AI can bridge or widen the gaps in worker-cen... | https://arxiv.org/abs/2505.20692v1 |
Analysis of Visual Stereotypes in Text-to-Image Generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , 12333–12347. Kawakami, A.; Chowdhary, S.; Iqbal, S. T.; Liao, Q. V .; Olteanu, A.; Suh, J.; and Saha, K. 2023. Sensing Wellbeing in the Work- p... | https://arxiv.org/abs/2505.20692v1 |
pragmatics , 59: 210–220. Mitchell, M.; Wu, S.; Zaldivar, A.; Barnes, P.; Vasserman, L.; Hutchinson, B.; Spitzer, E.; Raji, I. D.; and Gebru, T. 2019. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency , 220–229. Mittelstadt, B. D.; Allo, P.; Taddeo, M.; Wach... | https://arxiv.org/abs/2505.20692v1 |
Karahalios, K.; and Langbort, C. 2014. Auditing algorithms: Research methods for detecting discrimina- tion on internet platforms. Data and discrimination: converting critical concerns into productive inquiry , 22(2014): 4349–4357. Shatz, I. 2017. Fast, free, and targeted: Reddit as a source for re- cruiting participan... | https://arxiv.org/abs/2505.20692v1 |
Q.; Parker, A. G.; and De Choudhury, M. 2023. Synthetic lies: Understanding ai-generated misinformation and evaluating algorithmic and human solutions. In Proceedings of the 2023 CHI conference on human factors in computing systems .Appendix Figure A1: A slide showing an example of the mental model elicitation section.... | https://arxiv.org/abs/2505.20692v1 |
arXiv:2505.20693v1 [cs.CL] 27 May 2025Phir Hera Fairy: An English Fairytaler is a Strong Faker of Fluent Speech in Low-Resource Indian Languages Praveen Srinivasa Varadhan1, Srija Anand1, Soma Siddharta2, Mitesh M. Khapra1 1AI4Bharat, Indian Institute of Technology Madras, India 2Saryps Labs, India cs21d201@cse.iitm.ac... | https://arxiv.org/abs/2505.20693v1 |
total data used to train a state-of-the-art English TTS model, even when summing across 11 languages: (i) training from scratch on 11 Indian languages (IN11), (ii) fine-tuning an English-trained model on IN11, and (iii) fine-tuning on IN11 while retaining English data. Our findings challenge conven- tional assumptions ... | https://arxiv.org/abs/2505.20693v1 |
English: The model is fine-tuned on a mixture containing equal amounts of English [15] and IN11 data, allowing us to examine whether continued English exposure enhances generalization while im- proving performance on Indian languages. By evaluating these strategies, we aim to identify the most data- efficient and effec... | https://arxiv.org/abs/2505.20693v1 |
styles rendered by professional voice artists. To improve speaker di- versity, we integrate Google Crowdsourced TTS [21], enabling IN-F5 to generalize across different voices recorded in con- trolled environments. Finally, we leverage IndicV oices-R [16], a large-scale ASR-restored dataset, as a crucial factor in scali... | https://arxiv.org/abs/2505.20693v1 |
adaptation? We present the overall MUSHRA scores for the three fine- tuning strategies in Table 1. Intuitively, one might expect that fine-tuning with both English and IN11 data (EN →EN+IN) would yield the best results. However, our findings reveal a surprising trend: direct fine-tuning on IN11 alone (EN →IN) achieves ... | https://arxiv.org/abs/2505.20693v1 |
find that IN-F5 benefits from the English pre-training and generates clean, noise-free speech, making it sound more natural than human recordings. (ii) Polyglot. More intriguingly, IN-F5 emerges as an excep- tional polyglot. In our evaluations, we find that it enables speak- ers from one linguistic family (e.g., Indo-A... | https://arxiv.org/abs/2505.20693v1 |
IN-F5 is an “Excellent” expressive TTS, that marginally falls behind human recordings. 4.3. Scaling Effects in Low-Resource Adaptation In Table 4, we assess emergence across data scales, in terms of (i) naturalness (ii) speaker similarity and (iii) intelligibility. The scores reveal that fine-tuning on just 10 hours pe... | https://arxiv.org/abs/2505.20693v1 |
4.5. Advancing State-Of-The-Art for Indian Languages We benchmark IN-F5 against prior state-of-the-art models on 8 out of the 11 Indian languages included in the official Rasa [19] test set. The MUSHRA scores, presented in Table 6, highlight a significant 8-point improvement over V oiceCraft and a 14-point gain over Fa... | https://arxiv.org/abs/2505.20693v1 |
Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokar et al. , “V oicebox: Text-guided multilingual universal speech generation at scale,” Advances in neural information processing systems , vol. 36, 2024. [9] Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wang e... | https://arxiv.org/abs/2505.20693v1 |
Indian Lan- guages in Low-resource Settings,” in Proc. INTERSPEECH 2024 , 2024. [20] A. Baby, A. L. Thomas, N. Nishanthi, T. Consortium et al. , “Re- sources for indian languages,” in Proceedings of Text, Speech and Dialogue , 2016. [21] B. Abraham, D. Goel, D. Siddarth, K. Bali, M. Chopra, M. Choudhury, P. Joshi, P. J... | https://arxiv.org/abs/2505.20693v1 |
arXiv:2505.20700v1 [cs.CL] 27 May 2025Beyond Templates: Dynamic Adaptation of Reasoning Demonstrations via Feasibility-Aware Exploration Yong Wu1∗Weihang Pan1∗Ke Li2Chen Binhui3Ping Li4Binbin Lin1† 1College of Software Technology, Zhejiang University 2Fullong Technology, Ningbo, Zhejiang, China 3Ningbo Zhoushan Port Co... | https://arxiv.org/abs/2505.20700v1 |
This approach enables DART to flexibly adapt high-quality reasoning datasets to heterogeneous model populations, significantly improving reasoning elicitation under distribution shift. In summary, our contributions are as follows. •We identify the critical limitations of applying static curated reasoning datasets to di... | https://arxiv.org/abs/2505.20700v1 |
the indicator function. Such failures may arise from superficial imitation, incomplete reasoning chains, or insufficient justification for the final answer. Compounding this challenge is the substantial cost associated with constructing template datasets Dthat satisfy the Cognitive Template Demonstration criterion. Suc... | https://arxiv.org/abs/2505.20700v1 |
Overview of the DART framework. where Nsimdenotes the total number of rollouts performed for each candidate step st, with each rollout simulating a complete reasoning trajectory conditioned on the prefix s<tand the adoption of stepst. Empirically observed patterns (see Section 5.1) suggest that adaptability tends to ri... | https://arxiv.org/abs/2505.20700v1 |
effectively activate the student model’s own reasoning ability, we apply a standard cross-entropy loss on the outcome-aligned adapted trajectories generated during autonomous exploration. This training objective encourages the model to reinforce reasoning patterns that are not only aligned with the task goal but also f... | https://arxiv.org/abs/2505.20700v1 |
denotes direct zero-shot evaluation. (2) Static reflects standard offline supervised fine-tuning on the full set of expert trajectories, without any filtering or adaptability mechanism. (3) The Adaptation-Full strategy represents the complete DART pipeline, integrating imitation gap detection with outcome-consistent st... | https://arxiv.org/abs/2505.20700v1 |
5 Analysis We further analyze the internal mechanisms of DART, aiming to understand why selective imitation and autonomous exploration improve reasoning capabilities. 5.1 Step-wise Adaptability Reveals the Emergence of the Imitation Gap To empirically validate the imitation gap hypothesis introduced in Section 3, we es... | https://arxiv.org/abs/2505.20700v1 |
stability. 5.3 Capacity-Aligned Lexical Dynamics Under Adaptation To investigate how DART reshapes student model behavior at different scales, we analyze keyword frequency changes between static and adapted dataset. Table 3 lists the top 20 tokens with the largest shifts in the first sentence of each reasoning step for... | https://arxiv.org/abs/2505.20700v1 |
contrast, our work specifically addresses the unique challenges associated with training smaller-scale models for complex reasoning tasks. Data-Efficient Reasoning Elicitation A related line of work investigates how minimal supervision can elicit latent reasoning abilities in pretrained models Ye et al. [2025], Muennig... | https://arxiv.org/abs/2505.20700v1 |
editors, Proceedings of the Neural In- formation Processing Systems Track on Datasets and Benchmarks , volume 1, 2021b. URL https://datasets-benchmarks-proceedings.neurips.cc/paper_files/paper/ 2021/file/be83ab3ecd0db773eb2dc1b0a17836a1-Paper-round2.pdf . Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the kn... | https://arxiv.org/abs/2505.20700v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.