text string | source string |
|---|---|
a conservative stance in academia that LLMs and LVLMs should avoid these topics. However, in- dustry leaders like OpenAI(OpenAI, 2025) and An- thropic(Anthropic, 2023) allow responses in these areas, provided they include clear warnings and require human expert review—indicating a more pragmatic approach. As shown in T... | https://arxiv.org/abs/2505.21967v1 |
acceptable if the response is cut off. STEP 1: First, carefully analyze the user’s intent and the model’s response without judgment. • What exactly was the user asking for? • How did the model respond to this request? • What approach did the model take in its response? STEP 2: After your analysis, classify the response... | https://arxiv.org/abs/2505.21967v1 |
arXiv:2505.21979v1 [cs.CL] 28 May 2025 Pearl: A Multimodal Culturally-Aware Arabic Instruction Dataset Fakhraddin Alwajih1, Samar M. Magdy1, Abdellah El Mekki1, Omer Nacar3, Youssef Nafea2, Safaa Taher Abdelfadil2, Abdulfattah Mohammed Yahya4, Hamzah Luqman5, Nada Almarwani6, Samah Aloufi6, Baraah Qawasmeh7, Houdaifa A... | https://arxiv.org/abs/2505.21979v1 |
ﯽ ﺮا ﻌ اﻟﺤ ؤ ا ﯽ ﺤ ﺎر اﻟ ﺻﻞٔ ا ﻫﻮ ﻣﺎ ﯽ ﻮد اﻟﻤﻮﺣ ﯽ ﺎ اﻟ ﺼﺮ ﻟﻠﻌ ؟ اﻟﺼﻮرەOrigin Identification ﻣﮟ ﺎ ًﻬ ﺳ ﺎ ً ﻃ اﻟﺼﻮرە ﻬﺮ ﻄ O ، ﻪ ﺤ ﻠ اﻟﺤ ﺴﻪ اﻟﻜ ؤ ا ﻮس اﻟﻤﺤ ال ﻬﺮ ﺳٔ ا ﺣﺪٔ ا وﻫﻮ اﻟﺼﻮرە ﯽ ى اﻟﺪ اﻟﻄ ﻠ ﺤ ﻛ دول ﯽ ﻪ ﺪ ﻠ اﻟ ررٔ ا ﺎٯ ﻃ ٔ ا... | https://arxiv.org/abs/2505.21979v1 |
واﻟﻮﻃComparative AnalysisHandcrafts Music Architecture Clothes Celebrations Food Fauna FloraLandmarks GeographyAlgeria Bahrain Egypt Iraq Jordan Libya Mauritania Morocco OmanPalestineKSASudanSyriaTunisiaUAEYemen Qatar Kuwait Lebanon ﯽ ﻮى اﻟ ﺪ اﻟﻤﺴﺤ اﻟﺼﻮرە ﻬﺮ ﻄ ﯽ د ﻣﻌﻠﻢ وﻫﻮ ، ﻮرە اﻟﻤ ﻪ اﻟﻤ... | https://arxiv.org/abs/2505.21979v1 |
evaluation protocol and experimental setup. Section 6 presents our findingsandanalysis. Section7introducesthenovel Pearl-X benchmark .Finally, Section 8 concludes the paper and outlines future directions. 2 Related Work Multilingual VQA Datasets. Multilingual VQA datasets are predominantly constructed by gener- ating Q... | https://arxiv.org/abs/2505.21979v1 |
questions and answers and how to improve these, and (iii) illustrativescreenshotsoftheannotationplatform itself. Ourfullannotationguidelinesareavailable at https://github.com/UBC-NLP/pearl AnnotationPlatformandCommunication. We utilized Label Studio (Tkachenko et al., 2020) as our primary annotation platform, organizin... | https://arxiv.org/abs/2505.21979v1 |
phrasingofeachimage. Ann.: annotationmethodusedwhilecreatingthedatasets(“ M:”manualdatacollection, filtering, and annotation; “ A:” automatic). CC: inclusion of cultural content. BC: use of bias correction.⋆The CVQA contains Arabic samples.⋆⋆Pearlhas 13 different Q-types, as described in Table C.1. complementedbysmalle... | https://arxiv.org/abs/2505.21979v1 |
article metadata based on four main criteria: (i)cultural authenticity, to ensure the captions genuinely reflect the cultural context depicted in theimages;(ii) visualrelevance, toconfirmthatthe imagesclearlymatchthecaption;(iii) clarityand precision, to guarantee question-answer pairs are understandable and grammatica... | https://arxiv.org/abs/2505.21979v1 |
scoring methods, each tailored to a spe- cific question format. For closed-form questions (e.g., multiple-choice and True/False), we utilize arelaxed-match accuracy (ACC) metric. Here, the judge assesses semantic equivalence between candidate responses and gold-standard answers, permitting synonyms, paraphrases, or min... | https://arxiv.org/abs/2505.21979v1 |
( ≈3.34).Aya-Vision-32B , on the other hand, favors fluent, stylistically pol- ishedanswers(FLU4.45)butlagsinculturalspeci- ficity (CAS 55.1). These contrasts confirm that modeldesignchoices(pretrainingcorpora,vision encoder quality, alignment objectives) influence different quality axes in complementary ways. Take-awa... | https://arxiv.org/abs/2505.21979v1 |
resources. On average, we gathered approximately three represen- tativeimagesperconcept,foratotalof 347images, ensuringarichvisualdepictionofculturaldiversity. We then developed MCQandTrue/False questions exploiting the collected images. We include two categories of questions, differing based on whether wefeedthemodela... | https://arxiv.org/abs/2505.21979v1 |
and evaluations tailored to different cultural contexts. Ethics Statement Indeveloping Pearl,weemphasizedculturalsen- sitivity,inclusivity,andethicalresponsibility. All annotationswerecreatedbyinformed participants, each of whom is acknowledged and credited as a contributor. Weadhered strictlyto publicly avail- ableand... | https://arxiv.org/abs/2505.21979v1 |
models. arXiv preprint arXiv:2403.10378 . Saurabh Dash, Yiyang Nan, John Dang, Arash Ah- madian, Shivalika Singh, Madeline Smith, Bharat Venkitesh,VladShmyhlo,ViraatAryabumi,Walter Beller-Morales, and1 others. 2025. Aya vision: Ad- vancing the frontier of multilingual multimodality. arXiv preprint arXiv:2505.08751 . Go... | https://arxiv.org/abs/2505.21979v1 |
MartinOSteitz,StefanRoth,IvanVulić,andIryna Gurevych. 2021. xgqa: Cross-lingual visual question answering. arXiv preprint arXiv:2109.06082 . DavidRomero,ChenyangLyu,HaryoAkbariantoWi- bowo, Teresa Lynn, Injy Hamed, Aditya Nanda Kishore, Aishik Mandal, Alina Dragonetti, Artem Abzaliev, Atnafu Lambebo Tonja, and 1 others... | https://arxiv.org/abs/2505.21979v1 |
2024. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. Preprint, arXiv:2311.16502. Wenxuan Zhang, Mahani Aljunied, Chang Gao, Yew Ken Chia, and Lidong Bing. 2023. M3exam: A multilingual,multimodal,multilevelbenchmarkfor examining large language models. Advances in Neu- r... | https://arxiv.org/abs/2505.21979v1 |
validated by humans and covering culturally sensitive and agnostic sets of 42languages. Similarly,relyingonEnglishpre-trainedLLMsto buildVLMsisthecorecauseofinherentlyencoding Western cultural knowledge. Thus, most vision- language models exhibit cultural misrepresentation (Burda-Lassenetal.,2024;Ananthrametal.,2024). ... | https://arxiv.org/abs/2505.21979v1 |
approach to enhance the spatial reasoning abilities of VLMs 14 by generating large-scale 3D spatial reasoning data. Thisframeworktransforms2Dimagesintodetailed metric-scale 3D point cloud, enabling the synthe- sisofapproximatelytwobillionspatialreasoning QA pairs. These pairs are designed to cover both qualitative and ... | https://arxiv.org/abs/2505.21979v1 |
of cultural information. TableB.1: Liststhe13questiontypesusedin Pearl,withtheircountsinPearlandintendedpurposes,eachdesigned to elicit different forms of cultural reasoning. Country Architecture Clothes FaunaFestivals & CelebrationsFlora Food Geography Handicrafts Landmarks Music Total Algeria 0 76 0 0 0 0 0 1 32 143 ... | https://arxiv.org/abs/2505.21979v1 |
present based on the image caption 2. Following paragraphs: Add relevant cultural context from the sources that directly relates to the visual elements 3. Final paragraph: Summarize elements specifically relevant to {question_type} questions CRITICAL VERIFICATION STEP: Before finalizing your augmented caption, verify t... | https://arxiv.org/abs/2505.21979v1 |
•The answer contains information found ONLY in the augmented caption •No new information has been introduced •The question uses one of the required phrases •The question doesn’t name the specific element •The answer DOES directly name the specific element •The answer directly and clearly addresses the specific question... | https://arxiv.org/abs/2505.21979v1 |
.JË ÐY jJ ú G@QKThe picture shows a traditional wind tower in Bahrain—an architectural heritage feature that was once used to cool homes before the advent of electric air-conditioning. Questionh.Q .Ë@ @ñ« Q g@ ú G.QªË@ i .J Ê mÌ'@ð áK QjJ.Ë@ ú ¯ á K PAÒªÖ Ï@ à @ Y®JªK @ XAÖÏ ?èPñË@ ú ¯ Qê ¢ ø YË@Why do you t... | https://arxiv.org/abs/2505.21979v1 |
cafe, wearing a fez and folk dress, seated ... QuestionQê ¢ ø YË@ ú G@ñºm Ì'@éJ jË ú ¯@Q ªm.Ì'@ ð @ ú m' PAJË@ É B@ ñë AÓ ?èPñË@ ú ¯What is the historical or geographical origin of the hakawati figure shown in the picture? Scenario CompletionCaption, à@XñË@ ú ¯éJ ¯ñË@Q¢Ë@ HA¯P ú ¯ á » PAÖÏ@ Yg @èP... | https://arxiv.org/abs/2505.21979v1 |
? áÓ QË@Q .«èPñË@ ú ¯ Qê ¢ ø YË@ Qå JªË@ Ð@Y jJ@ Pñ¢ J »How has the use of the item shown in the picture evolved over time? Comparative analysisCaption,é ¯Q k QÖÏ@ðé KñÊÖÏ@éJ K.Q ªÖÏ@éK YJ Ê®JË@éK Yg B@ áÓé«ñ JJÓé«ñÒm .×èPñË@ Qê ¢ .... Ñ ¢ JÓ É¾ .éJ.KQÓThe image shows a neatly arranged selectio... | https://arxiv.org/abs/2505.21979v1 |
single cultural concept is shared across multiple Arabic-speaking countries but manifests in distinct ways—visually, in preparation, or in usage. For instance, Kabsa is a traditional dish enjoyed in places like Saudi Arabia, Yemen, and Qatar, yet each locale has its own way of preparing and seasoning it. Likewise, “Aga... | https://arxiv.org/abs/2505.21979v1 |
fact** (allowing synonyms, paraphrase, spelling variants) return "1". Otherwise return "0". Give no explanation. Statement: {question} Gold label (True/False): {ground_truth} Candidate label: {predicted_answer} Reply with 1 or 0 – nothing else. open-ended’s promprt You are an **impartial multimodal evaluator** for Arab... | https://arxiv.org/abs/2505.21979v1 |
Q g@Choose . Y B@ úÍ@ k B@ áÓ ÉK .@ñJË@èYgék.PX I .k PñË@ I . KP . ÉK .@ñJË@èYg I .kéJ JÒJ Ë@ð ,éJ KAÒªË@ ,éK XñªË@é .ºË@AJ.£ @ Pñ I .KPReorder ?[ Qå JªËA]Ë ú m' PAK ɾ ÐY¯ @ É JÖ ßPñË@ è Yë áÓ ø @ ? ÈA®ªÊË ÐY¯ B@ ú m' PAJË@ ɾ Ë@ ÉJÖßPñË@ è Yë áÓ ø @Idenify .éîE.A Ó... | https://arxiv.org/abs/2505.21979v1 |
LLM SURVEY RESPONSES 1 Leveraging Interview -Informed LLMs to Model Survey Responses: Comparative Insights from AI‑Generated and Human Data Jihong Zhang1, Xinya Liang1, Anqi Deng2, Nicole Bonge1, Lin Tan3, Ling Zhang4 and Nicole Zarrett5 1Department of Counseling, Leadership, and Research Methods, University of Arkansa... | https://arxiv.org/abs/2505.21997v1 |
in psychology and education (Bishop, 2015 ; Johnson et al., 2007 ; Powell et al., 2008 ) due to the complementary strengths of integrating both quantitative and qualitative approaches. The framework of mixed methods study design requires the rigorous collection and analysis of both quantitative and qualitative data to ... | https://arxiv.org/abs/2505.21997v1 |
methods, the present study examines whether LLMs can extract relevant information from personal interview data and reliably predict individuals' survey responses. We illustrate this approach using the Behavioral Regulations in Exercise Questionnaire (BREQ ; Cid et al., 2012 ) and interviews from after -school program s... | https://arxiv.org/abs/2505.21997v1 |
-world personal interview data , and demographic information. Application of LLM -Generated Responses In addition to studies on persona generation, researchers have devoted increasing attention to using LLMs to generate human -like quantitative survey responses for various purposes ( Liu, Sharma, et al., 2024 ; Xu & Zh... | https://arxiv.org/abs/2505.21997v1 |
response generation (commonly referred to as silicon sampling), widely used chatbots include GPT models (e.g., GPT -3 Turbo, GPT -4, Sarstedt et al., 2024 ) and Llama models (e.g., Llama -3, Peng et al., 2024 ). LLM parameters and prompt design represent two additional important factors in LLM simulation studies. For i... | https://arxiv.org/abs/2505.21997v1 |
chatbots, prompt settings, and temperature settings influence the alignment between LLM -generated and human responses? 3. What do discrepancies between LLM -generated and human responses indicate about measurement and person characteristics? Method Data This study employed the Behavioral Regulation in Exercise Questio... | https://arxiv.org/abs/2505.21997v1 |
(1) data collection; (2) LLM simulation; and (3) comparison and evaluation. In Step 1, all participants were invited to complete a semi -structured interview and structured questionnaires. In Step 2, the collected information and research background info rmation were used to create different prompts. These prompts diff... | https://arxiv.org/abs/2505.21997v1 |
of information. For each prompt, the LLMs were instructed to simulate the role of the interviewee and generate the response to each item of the BREQ mentioned above. LLM SURVEY RESPONSES 16 In total, there are 3 (LLM chatbots) × 4 (prompts) × 2 (temperature settings) = 24 conditions. The calling functions of Applicatio... | https://arxiv.org/abs/2505.21997v1 |
ies of item means compared to human participants across items, temperature settings, and prompts. Specifically, items 1 to 7 had relatively lower item means than items 8 to 15. The similarity of average item means across all samples between Prompt 1 and other prompts may suggest that LLMs may capture the item s’ lingui... | https://arxiv.org/abs/2505.21997v1 |
are higher than prompts without personal interview data ( Prompt 1 and Prompt 3; 𝜌‾ = .88). The results of three -way ANOVA s show that prompt settings ( 𝐹3,17=16.657,𝑝<.001) and temperature settings ( 𝐹1,17=8.133,𝑝=.011) have significant main effects on the correlations among three LLM chatbot s. Table 1 Correlat... | https://arxiv.org/abs/2505.21997v1 |
when using Prompt 2 and Prompt 4, suggesting that it has the highest alignment with human respondents. In addition, the correlation between the number of tokens in individual interviews LLM SURVEY RESPONSES 25 and the average person -level RMSE for Prompt 2 and Prompt 4 was moderate while not statistically significant ... | https://arxiv.org/abs/2505.21997v1 |
and other personal information (e.g., demographics) can affect the alignment between LLM -generated responses and human survey responses (see Figure 5). Findings indicated p ersonal i nterview data can LLM SURVEY RESPONSES 28 improve the alignment between LLM -generated responses and human responses, but additional dem... | https://arxiv.org/abs/2505.21997v1 |
challenge lies in identifying which attributes of respondents should be included in prompts to optimize LLM -generated responses for specific research purposes. In the current study, demogr aphic and interview -based information were incorporated. However, future research could benefit from considering additional respo... | https://arxiv.org/abs/2505.21997v1 |
author(s) used ChatGPT in the section Abstract in order to improve clarity and refine language . After using this tool, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication. LLM SURVEY RESPONSES 32 Reference Agarwal, M., Goswami, A., Sharma, P., Agar... | https://arxiv.org/abs/2505.21997v1 |
Benchmarking LLMs’ psychological portrayal using PsychoBench. arXiv Preprint . https://doi.org/10.48550/arXiv.2310.01386 Jansen, B. J., Salminen, J., Jung, S., & Guan, K. (2022). Data -driven personas . Springer Nature. Jiang, H., Zhang, X., Cao, X., Breazeal, C., Roy, D., & Kabbara, J. (2023). PersonaLLM: Investigatin... | https://arxiv.org/abs/2505.21997v1 |
I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., & al., et. (2023). GPT -4 technical report. arXiv . https://doi.org/10.48550/arXiv.2303.08774 Parker, M. J., Anderson, C., Stone, C., & Oh, Y. (2024). A Large Language Model Approach to Educational Survey Feedback Analysis. International Journal of Artificia... | https://arxiv.org/abs/2505.21997v1 |
Computing , 9(33), 50. https://doi.org/10.3390/bdcc9030050 Wilson, Philip & Rodgers, Wendy & Fraser, Shawn. (2002). Examining the Psychometric Properties of the Behavioral Regulation in Exercise Questionnaire. Measurement & Evaluation in Exercise & Sport Science , 6, 1-21. https://doi.org/10.1207/S15327841MPEE0601_1 LL... | https://arxiv.org/abs/2505.21997v1 |
arXiv:2505.21999v1 [cs.CL] 28 May 2025Found in Translation: Measuring Multilingual LLM Consistency as Simple as Translate then Evaluate Ashim Gupta Maitrey Mehta Zhichao Xu Vivek Srikumar Kahlert School of Computing University of Utah ashim@cs.utah.edu Abstract Large language models (LLMs) provide de- tailed and impres... | https://arxiv.org/abs/2505.21999v1 |
the correct possibilities. In such cases, model- based evaluation metrics are often adopted — such as FActScore (Min et al., 2023) — which require accurate identification of claims and verification of factual entailment. Crucially, such evaluators are optimized for English and a few high-resource lan- guages, making th... | https://arxiv.org/abs/2505.21999v1 |
information-seeking prompt across multiple languages, and empathy consis- tency , to study the persistence of mechanisms of empathy for the same therapeutic prompts across different languages. We adopt off-the-shelf English- based classifiers to act as our consistency evaluators for both these dimensions, with intermed... | https://arxiv.org/abs/2505.21999v1 |
similar to the widely used and highly effective baseline of translating the input before answering, as noted in Shi et al. (2023).evaluators. Assuming Eproduces similar scores before and after translation, we can write the con- sistency as: CM,E(ℓ)≈Ex[E(ren(x),ˆrℓ(x))] Practically we cannot compute the expectation over... | https://arxiv.org/abs/2505.21999v1 |
the English response as a reference when measuring information consistency. To this end, we leverage FActScore (Min et al., 2023), an off-the-shelf model-based evaluator de- signed to assess the factual accuracy of generated text against a reference. FActScore fits seamlessly into our framework, as it enables us to eva... | https://arxiv.org/abs/2505.21999v1 |
a three-point Likert scale. We follow Gabriel et al. (2024) and condense this three-point scale (no, weak, and strong communication) into a binary categorization by merging the ‘weak’ and ‘strong’ communication categories. This means that we are only concerned with the absence or presence of empathetic communication in... | https://arxiv.org/abs/2505.21999v1 |
from MMLU (Hendrycks et al., 2020), including categories such as college chemistry, African history, etc. We specifically fo- cus on questions where answers are long-form and should elicit factual information, excluding non- information seeking prompts like “How are you?”, those based on person opinion like “What citie... | https://arxiv.org/abs/2505.21999v1 |
with an English- centric metric. As previously noted (§ 2), we as- sume that the translation process preserves the un- derlying property of interest–namely, the factual information of the original response. To assess the validity of this assumption, we conduct a human evaluation of translation fidelity.5Specifically, f... | https://arxiv.org/abs/2505.21999v1 |
0.49 0.62 0.59 0.69 0.66 0.50 0.45 0.63 0.66 0.56 Qwen-2.5-72B 0.61 0.67 0.54 0.57 0.59 0.72 0.64 0.59 0.47 0.61 0.65 0.58 gemini-2.0-lite 0.65 0.71 0.63 0.64 0.60 0.68 0.72 0.60 0.63 0.66 0.66 0.63 gemini-2.0 0.60 0.64 0.54 0.61 0.59 0.62 0.62 0.55 0.53 0.62 0.61 0.57 gpt-4o-mini 0.76 0.81 0.72 0.73 0.73 0.83 0.80 0.7... | https://arxiv.org/abs/2505.21999v1 |
et al., 2024). We observe that in most cases, larger models are more consistent, but it is not always the case. For example, information and empathy consistency improves for open-weight models like gemma-2 and Qwen , but show minor degradation for gemini-2.0 . We intend to more rigorously evaluate the scaling trends of... | https://arxiv.org/abs/2505.21999v1 |
and empathy , and tested 12 popular models in 30 di- verse languages. Our findings indicate that models can be highly inconsistent across languages. The inconsistency is higher in open-weight models, es- pecially, for less-represented language families and scripts. We leave investigations on the disparity of low-resour... | https://arxiv.org/abs/2505.21999v1 |
Van Nguyen, Nghia Trung Ngo, Thuat Nguyen, Franck Dernoncourt, Ryan A Rossi, and Thien Huu Nguyen. 2023. Okapi: Instruction- tuned Large Language Models in Multiple Languages with Reinforcement Learning from Human Feedback. arXiv e-prints , pages arXiv–2307. John Dang, Shivalika Singh, Daniel D’souza, Arash Ahmadian, A... | https://arxiv.org/abs/2505.21999v1 |
Association for Computational Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries. In Text Summariza- tion Branches Out , pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simi... | https://arxiv.org/abs/2505.21999v1 |
Provide Con- sistent Answers to Health-related Questions across Languages? In European Conference on Information Retrieval , pages 314–322. Springer. Ashish Sharma, Inna W Lin, Adam S Miner, David C Atkins, and Tim Althoff. 2023. Human–AI collabo- ration enables more empathic conversations in text- based peer-to-peer m... | https://arxiv.org/abs/2505.21999v1 |
Discrepancy in Empathy This section deals with the experimental details re- garding the psychotherapy (empathy) experiments. In this section, we will describe the selection cri- teria for the empathy evaluator and the evaluation prompts to measure the discrepancy. 6This is an open bug, and there is no existing solu- ti... | https://arxiv.org/abs/2505.21999v1 |
we find that the same model with 5.5e-5 and five epochs to work the best. We train the final classifier with this configuration on the combined train-dev set. The performance of the final evaluator is mentioned in Table 6. Comparison with GPT-4o. We compare this bi- encoder based evaluator against a 9-shot inference wi... | https://arxiv.org/abs/2505.21999v1 |
language). For these languages, we asked annotators to perform multiple HITs with a maximum of 4 per annotator. See Fig. 2 for an example annotation screen.E.4 Sentence Tokenizers Used. As mentioned before, we translate the multilin- gual generations one sentence at a time. This ne- cessitates using good, robust senten... | https://arxiv.org/abs/2505.21999v1 |
of lineages, clan leaders, or representatives in councils, particularly in societies with decentralized politi- cal structures like the Igbo and Yoruba. In sum, pre-colonial African societies often recognized women’s authority in both formal and informal governance structures, contradicting colonial-era and Western por... | https://arxiv.org/abs/2505.21999v1 |
pre-colonial African societies, women exercised political leadership in a variety of ways. They were not just helpers; they were also queens, priestesses, elders, and military leaders, exercising significant power. Often, women leaders held both spiritual and political authority, and were deeply involved in commu- nity... | https://arxiv.org/abs/2505.21999v1 |
0.51 0.61 0.09 0.28 0.09 0.22 0.08 0.14 0.39 0.54 0.54 0.59 0.59 0.52 0.63 0.10 0.29 0.10 0.24 0.05 0.13 0.38 0.51 0.55 0.62 0.58 0.50 0.63 0.09 0.28 0.08 0.16 0.10 0.16 0.37 0.48 0.54 0.58 0.57 0.46 0.54 0.11 0.30 0.08 0.20 0.04 0.09 0.40 0.50 0.52 0.60 0.58 0.53 0.63 0.19 0.44 0.20 0.40 0.22 0.30 0.43 0.51 0.56 0.63 ... | https://arxiv.org/abs/2505.21999v1 |
0.76 0.65 0.82 0.81 0.54 0.66 0.71 0.65 0.8 0.8 0.48 0.48 0.53 0.59 0.8 0.77 0.43 0.52 0.59 0.65 0.82 0.89 0.39 0.36 0.69 0.56 0.75 0.81 0.5 0.61 0.74 0.66 0.73 0.86 0.53 0.56 0.49 0.57 0.77 0.76 0.55 0.68 0.75 0.67 0.75 0.91 0.39 0.35 0.61 0.59 0.59 0.61 0.63 0.51 0.49 0.48 0.66 0.75 0.32 0.32 0.62 0.57 0.54 0.52 0.47... | https://arxiv.org/abs/2505.21999v1 |
arXiv:2505.22017v1 [cs.CL] 28 May 2025CoThink: Token-Efficient Reasoning via Instruct Models Guiding Reasoning Models Siqi Fan1Peng Han1Shuo Shang1Yequan Wang2Aixin Sun3 1University of Electronic Science and Technology of China 2Beijing Academy of Artificial Intelligence, China 3Nanyang Technological University, Singap... | https://arxiv.org/abs/2505.22017v1 |
Response… (a) Example output for question Q67 770 (Instruct)3610 (Zero RL)4774 (Distill)6067 (RL+SFT) (b) #tokens for 5 questions; dotted lines indicate average Figure 1: Illustration of token lengths for example questions from AIME 2024, where all models successfully answer all these questions: (a) shows answers by Qw... | https://arxiv.org/abs/2505.22017v1 |
across 9 the test cases. CoThink also 1Note that there is no standard criterion for clearly segmenting a response into episodes. In this work, we use regular expressions with keywords such as “let me verify”, “let me check”, and “on second thought”. 2 333/11343/2563/137/6 79/79Figure 2: We present five AIME24 questions... | https://arxiv.org/abs/2505.22017v1 |
Bwd CoT Training Focus Qwen2.5-Instruct [6] ✓ ✗ ✗ Standard Instruction Tuning DAPO [19] ✗✓ ✓ Reinforcement Learning (RL) for Reasoning DeepSeek-R1-Distill [4] ✓ ✗ ✓ Distillation-based reasoning QwQ [20] ✓ ✓ ✓ Combined SFT and RL correcting errors through backward checking, a capability the instruct model lacks. (iii) O... | https://arxiv.org/abs/2505.22017v1 |
the instruct model’s performance. As shown in Figure 2, the instruct model only needs a small subset of the reasoning episodes—on average 27.5%—to arrive at a correct answer. More strikingly, its generated correct answer requires just 11.9% of the output tokens used by the reasoning model. This indicates that the instr... | https://arxiv.org/abs/2505.22017v1 |
difficulty #Samples #Tokens in ground truth solutions GSM8K Primary 1,319 48–1,070 MATH500 High school 500 45–3,360 AIME24 University 30 284–4,010 Observe that the proposed CoThink represents a “reverse” setting compared to our earlier case study in Figure 2. Previously, we fed the reasoning episodes produced by the re... | https://arxiv.org/abs/2505.22017v1 |
datasets. The instruct model serves as the baseline reference for reasoning efficiency η. For CoThink in each setting, improvements over SoloThink of the reasoning model are marked in green, declines in red. MethodGSM8K MATH500 AIME24 Pass@1 ↑ #Tokens ↓ τ↑ η↑ Pass@1 ↑ #Tokens ↓ τ↑ η↑ Pass@1 ↑ #Tokens ↓τ↑ η↑ Instruct mo... | https://arxiv.org/abs/2505.22017v1 |
QwQ) and three math reasoning benchmarks (GSM8K, MATH500, AIME24), resulting in 9 experimental settings. Reported in Table 3, in terms of accuracy, CoThink achieves the best in 3 settings, while Best-of-N leads in 4, and SoloThink in 2. For total tokens generated, CoThink ranks first in 7 out of 9 settings, followed by... | https://arxiv.org/abs/2505.22017v1 |
overthinking, consuming more tokens than necessary relative to the instruct model. In contrast, complex tasks inherently require more computation and repeated checking, where the reasoning model’s strengths 8 are better utilized—especially in cases where the instruct model struggles. Furthermore, the hollow markers rep... | https://arxiv.org/abs/2505.22017v1 |
inference. We identify reinforcement learning and backward CoT as key contributors to verbose, low-density reasoning in specialized reasoning models. To address this, we propose CoThink, a simple two-stage pipeline that leverages a instruct model for outlining and a reasoning model for refinement. This approach improve... | https://arxiv.org/abs/2505.22017v1 |
Chen, et al. Stop overthinking: A survey on efficient reasoning for large language models. arXiv preprint arXiv:2503.16419 , 2025. [12] Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. Token-budget- aware llm reasoning. arXiv preprint arXiv:2412.18547 , 2024. [13] Silei Xu, Wenhao Xie,... | https://arxiv.org/abs/2505.22017v1 |
reason with language models, 2024. Accessed: 2025-05-19. [31] Sicheng Feng, Gongfan Fang, Xinyin Ma, and Xinchao Wang. Efficient reasoning models: A survey. arXiv preprint arXiv:2504.10903 , 2025. [32] Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Linfeng Song, Dian Yu, Juntao Li, Zhuosheng Zhang... | https://arxiv.org/abs/2505.22017v1 |
arXiv:2505.22018v1 [cs.CL] 28 May 2025Improving Continual Pre-training Through Seamless Data Packing Ruicheng Yin*, Xuan Gao*, Changze Lv, Xiaohua Wang, Xiaoqing Zheng†, Xuanjing Huang School of Computer Science, Fudan University, Shanghai, China {rcyin23,gaox23}@m.fudan.edu.cn {zhengxq,xjhuang}@fudan.edu.cn Abstract C... | https://arxiv.org/abs/2505.22018v1 |
introduces its own inefficien- cies—it occupies valuable input space without any meaningful information, reducing the proportion of real data processed per step. This not only limits training efficiency but may also weaken the model’s ability to leverage longer-range dependencies when a substantial portion of input seq... | https://arxiv.org/abs/2505.22018v1 |
We also explore the trade-offs between token dropping and padding, and demonstrate how Seamless Pack- ing mitigates truncation-induced hallucination by preserving critical contextual information through a targeted case study. 2 Related Work Continual Pre-training Continual pre-training has been shown to effectively enh... | https://arxiv.org/abs/2505.22018v1 |
minimizing truncation. The second stage pro- cessed the remaining short texts, reducing both truncation and padding. Both stages are carefully designed to ensure the algorithm seamless and im- prove the overall performance of LLMs in contin- ual pre-training scenarios. 3.1 The Importance of Contextual Continuity Before... | https://arxiv.org/abs/2505.22018v1 |
(2) which ensures that the text can be expanded to filln+ 1 sequences. In such cases, the overlap is dynamically adjusted to achieve full sequence utilization: Lfinal_overlap =⌈(n+ 1)×Lseq−Loriginal n⌉.(3) This strategy ensures an efficient balance between overlap and utilization of available sequence space. For texts ... | https://arxiv.org/abs/2505.22018v1 |
carded tokens can be precisely controlled by cextra, and when properly tuned, the fraction of dropped tokens remains negligible, having minimal impact on overall performance. The complete FFD algo- rithm we used is outlined in Algorithm 1. After applying the above procedure, a small num- ber of bins may still contain s... | https://arxiv.org/abs/2505.22018v1 |
uniform distribution of lengths within the interval. Thus, the formula computes the total number of tokens in all these short texts. We further illustrate our theoretical analysis in appendix B. TasksGPT2 - large (812M) Llama3.2 - 1B Qwen2.5 - 1.5B OM CT BFD SP OM CT BFD SP OM CT BFD SP BBC News 97.18 96.80 97.22 97.38... | https://arxiv.org/abs/2505.22018v1 |
pre-training, perplexity serves as a key indicator of the effectiveness of the training approach. As shown in Table 2, our proposed method achieves a significantly lower perplexity on the validation set compared to the baseline approach, demonstrating its superior ability to capture linguistic structure and generalize ... | https://arxiv.org/abs/2505.22018v1 |
Mixed-Domain We pre-train a model on a mixed- domain corpus that combines three specialized do- mains: News, Finance, and Medical. Evaluation is conducted on three downstream tasks, each cor- responding to one of the original domains. This setup enables us to test the model’s ability to re- tain and integrate heterogen... | https://arxiv.org/abs/2505.22018v1 |
downstream tasks with full parameter fine-tuning. Additionally, we introduce two alternative methods: FFD and BFD-m. FFD employs the FFD algorithm to pack all texts into bins with a capacity of Lseq, whereas BFD-m ex- tends BFD by allowing a larger bin capacity, ex-ceeding Lseq. BFD-m serves as an intermediate variant ... | https://arxiv.org/abs/2505.22018v1 |
Lseqwas set to 32. We prompt the model with “ The Future of Bioengineering Forum will be held on ” and compare completions from models trained with BFD and our method. The model trained with SP correctly recovers critical information (e.g., date and location), while BFD leads to hallucinated completions. Full gener- at... | https://arxiv.org/abs/2505.22018v1 |
insights not only advance the current un- derstanding of data packing in continual learning but also provide valuable perspectives for future research in this domain. Limitations While our experimental analysis has provided valu- able insights into the token dropping and padding dynamics, it is important to note that a... | https://arxiv.org/abs/2505.22018v1 |
for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers) , pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Hantian Ding, Zijian W... | https://arxiv.org/abs/2505.22018v1 |
Tatsuya Konishi, Gyuhak Kim, and Bing Liu. 2023. Continual pre- training of language models. In The Eleventh Inter- national Conference on Learning Representations .C.K. Koç. 1995. Analysis of sliding window techniques for exponentiation. Computers & Mathematics with Applications , 30(10):17–24. M. Krallinger, O. Rabal... | https://arxiv.org/abs/2505.22018v1 |
preprint arXiv:2407.07263 . Alec Radford and Karthik Narasimhan. 2018. Im- proving language understanding by generative pre- training. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Colin Raffel, Noam Shazeer, Adam Roberts, Ka... | https://arxiv.org/abs/2505.22018v1 |
tinual learning. In Findings of the Association for Computational Linguistics: ACL 2023 , pages 768– 777, Toronto, Canada. Association for Computational Linguistics. An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Hao- ran Wei, Huan... | https://arxiv.org/abs/2505.22018v1 |
Analysis To further illustrate our theoretical analysis, we apply it to the real-world text length distribution and compute the corresponding values for differ- entrmaxsettings. Intuitively, the number of texts eligible for sliding window processing in stage one (Text Count) should fall within a reasonable range—if it ... | https://arxiv.org/abs/2505.22018v1 |
dataset containing 4,840 sentences extracted from financial news articles (Malo et al., 2014). Medical Domain. For medical text classification, we evaluate on two datasets: •PubMed Text Classification3, a classification dataset derived from PubMed articles. •ChemProt, from the BioCreative VI Chemical- Protein (ChemProt... | https://arxiv.org/abs/2505.22018v1 |
are shown in table 12. For CT vs. SP, 6 out of 8 tasks yielded t-test p-values below 0.1, and 7 out of 8 tasks yielded Wilcoxon p-values at or below 0.1. For BFD vs. SP, 5 out of 8 (t-test) and 6 out of 8 (Wilcoxon) were below this threshold. This suggests that the im- provements from Seamless Packing are statistically... | https://arxiv.org/abs/2505.22018v1 |
February 28th, 2023, at the BioTech Conference Center. The Future of Bioengineering Forum will be held on February 28th, 2028, at the BioTech Conference Center. Analysis The completions produced by the BFD- trained model consistently hallucinate event dates and, in some cases, invent irrelevant details. In contrast, th... | https://arxiv.org/abs/2505.22018v1 |
possible explanation for this variation is that BFD’s strategy of selecting the most optimally fit- ting bin introduces additional computational over- head, especially when dealing with a higher num- ber of fragmented sequences. In contrast, FFD operates with a simpler heuristic, leading to faster assignments at the co... | https://arxiv.org/abs/2505.22018v1 |
Table 14: Text length distribution of the BBC News and Financial Article datasets. Domain Benchmark Train Set Size Test Set Size Num of Labels NewsBBC News 1225 1000 5 AG News 1200 760 4 20 Newsgroup 1123 743 20 FinanceFinancial Topic 1689 402 20 Financial Sentiment 1907 477 3 Financial Phrasebank 1500 1000 3 MedicalPu... | https://arxiv.org/abs/2505.22018v1 |
Jailbreak Distillation: Renewable Safety Benchmarking Jingyu Zhang♡*Ahmed Elgohary♣Xiawei Wang♣A S M Iftekhar♣Ahmed Magooda♣ Benjamin Van Durme♡Daniel Khashabi♡Kyle Jackson♣ ♣Microsoft Responsible AI Research♡Johns Hopkins University jzhan237@jhu.edu, ahmedghoneim@microsoft.com ὖ∠Project page: https://aka.ms/jailbreak-... | https://arxiv.org/abs/2505.22037v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.