text string | source string |
|---|---|
response includes ‘screenshot’ action, please disregard it and just take the action of sleeping for 1s. Return one line or multiple lines of python code to perform the action each time, be time efficient. When predicting multiple lines of code, make some small sleep like ‘time.sleep(0.5);’ interval so that the machine ... | https://arxiv.org/abs/2505.21936v1 |
arXiv:2505.21937v1 [cs.CL] 28 May 2025Graph-Assisted Culturally Adaptable Idiomatic Translation for Indic Languages Pratik Rakesh Singh, Kritarth Prasad, Mohammadi Zaki and Pankaj Wasnik Media Analysis Group, Sony Research India {pratik.singh, kritarth.prasad, mohammadi.zaki, pankaj.wasnik}@sony.com Abstract Translatin... | https://arxiv.org/abs/2505.21937v1 |
the word or phrase level rather than capturing an idiom’s holistic meaning. This often leads to literal translations, distorting the intended meaning of the source text (Baziotis et al., 2023; Raunak et al., 2023; Dankers et al., 2022). Recent efforts to address idiomatic translation have primarily relied on (1) idiom ... | https://arxiv.org/abs/2505.21937v1 |
possible pairs of languages. 2 Related work and Motivation 2.1 Related Works Idiomatic Text Translation. Previous studies have explored various strategies to enhance NMT per- formance for idiomatic translation. (Salton et al., 2014) introduced a substitution-based method, where source-side idioms are replaced with thei... | https://arxiv.org/abs/2505.21937v1 |
multiple meanings, increas- ing complexity. GNNs disambiguate meanings us- ing context, leveraging neighborhood information and learned representations to differentiate between senses based on connectivity. 3 Methodology In this section, we first present the problem state- ment followed by the training and inference of... | https://arxiv.org/abs/2505.21937v1 |
/ h1 Features Link PredictionFigure 2: Overall training process of IdiomCE : (a) GNN training – illustrating the creation of a Knowledge Graph using source and target idioms, specifically for en-hi , leveraging LaBSE embeddings and training a GNN for the Link Prediction (LP) task; (b) Node Duplication – demonstrating h... | https://arxiv.org/abs/2505.21937v1 |
then insert edges from vto these duplicated source nodes, as illustrated in Figure 2. In this way, we obtain an augmented graph G′with these newly created nodes and edges added to the origi- nal graph. This approach differs from (Guo et al., 2024), where the authors duplicate source nodes directly based on their degree... | https://arxiv.org/abs/2505.21937v1 |
the one-to- many nature of idioms where a single target idiom may correspond to multiple source idioms convey- ing the same figurative meaning, it is crucial to connect the unseen node to the most similar target idiom neighbors. To achieve this, we propose training a BERT-based encoder (denoted as BCL(·)) in a contrast... | https://arxiv.org/abs/2505.21937v1 |
For unseen nodes, completely isolated idioms would yield no meaningful results. To address this, we make the following assumption about the train- ing dataset D. Assumption. For any unseen node u,∃v∈ D such thatcos(BCL(u),BCL(v))⩾τ, where τ∈[0,1]. For our experiments, we choose τto be 0.75. To infer on unseen nodes, we... | https://arxiv.org/abs/2505.21937v1 |
2023a) as our primary metric, as it is an LLM- based approach specifically designed for assessing idiomatic translations we call it here LLM-eval and use WMT22-CometKiwi-DA as a supplementary evaluation metric. Models. We test the effectiveness of our approach by using base LLMs of varying sizes like Gemma2 9B (Team et... | https://arxiv.org/abs/2505.21937v1 |
1.25 0.57 1.3 0.54 Direct 1.27 0.68 1.23 0.67 1.16 0.62 1.12 0.59 GPT-4oIdiomCE 2.39 0.70 2.25 0.69 1.87 0.67 1.83 0.66 Direct 2.14 0.73 1.99 0.764 1.741 0.72 1.67 0.71 Table 1: Performance Metrics of Various Models on Mixed Dataset; COMET range [0,1]. Model Methodshi-xx bn-xx ta-xx te-xx LLM-eval COMET LLM-eval COMET ... | https://arxiv.org/abs/2505.21937v1 |
observed errors into three distinct types, as outlined below: •Morphological Issues. In some cases, Llama 3.1 8B and Llama 3.2 3B directly replaced an idiom without adapting its morphology, leading to unnatural phrasing in the target language. This suggests that smaller mod- els struggle with idiom adaptation, whereas ... | https://arxiv.org/abs/2505.21937v1 |
data used, which can limit the model’s performance. References Ruchit Agrawal, Vighnesh Chenthil Kumar, Vignesh- waran Muralidharan, and Dipti Sharma. 2018. No more beating about the bush : A step towards idiom handling for Indian language NLP. In Proceedings of the Eleventh International Conference on Language Resourc... | https://arxiv.org/abs/2505.21937v1 |
Hao, Jing Zhang, Hongzhi Yin, Cuiping Li, and Hong Chen. 2020. Pre-training graph neural net- works for cold-start users and items representation. Preprint , arXiv:2012.07064. Kazuma Hashimoto and Yoshimasa Tsuruoka. 2016. Adaptive joint learning of compositional and non- compositional phrase embeddings. In Proceedings... | https://arxiv.org/abs/2505.21937v1 |
et al. Rei. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT) , pages 634–645, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics. Sara Rezaeimanesh, Faezeh Hosseini, and Yadollah Ya... | https://arxiv.org/abs/2505.21937v1 |
linguistic researcher on idioms and good at {tgt_lang} and {src_lang}. Choose the best {tgt_lang} idiom matching the {src_lang} idiom and Context of Source Sentence in which it is used in. Only Provide Best macthing {tgt_lang} Idiom Do not provide any explaination. {src_lang} idiom:{en_idm} Source Sentence:{sent} Optio... | https://arxiv.org/abs/2505.21937v1 |
- 4,479, Tamil - 4,179, Hindi - 4,722, and English – 4,500. When this data is transformed into a graphical structure, the training dataset—prior to applying the Node Duplication Augmentation strategy the Training Composition for GNN expands to: • English and Tamil: 7,646 • English and Telugu: 7,988 • English and Bengal... | https://arxiv.org/abs/2505.21937v1 |
বািড়েয় নওয়া নই , ক ােবাচন বেলিছল, এবং ু হািস িদেয়। IdiomCE: আমরা িহেপ যু নই , ক ােবাচন বেলিছল , এবং ু হািস িদেয়। Source: After all, charity begins at home. Direct: অবশ ই, দান িনেজর ঘেরর দখা তারপর পেরর ঘর। IdiomCE: অবেশেষ, দান বািড়র থেক হয়। Source: Youd both have got on like a house on fire. Direct: আপনার... | https://arxiv.org/abs/2505.21937v1 |
1.26 0.76 1.21 0.74 LLama-3.2-3BIdiomCE 1.42 0.58 1.26 0.59 1.15 0.52 1.24 0.51 Direct 1.12 0.62 1.06 0.60 1.03 0.51 1.09 0.54 IdiomKB 1.25 0.61 1.05 0.59 1.07 0.52 1.11 0.52 LIA 1.13 0.565 1.01 0.5702 0.97 0.510 1.06 0.491 SIA 1.18 0.57 1.15 0.58 1.10 0.53 1.09 0.48 Gemma2-9b-itIdiomCE 2.08 0.69 1.84 0.69 1.76 0.68 1.... | https://arxiv.org/abs/2505.21937v1 |
arXiv:2505.21940v1 [cs.CL] 28 May 2025RISE: Reasoning Enhancement via Iterative Self-Exploration in Multi-hop Question Answering Bolei He1,2*Xinran He2*Mengke Chen2Xianwei Xue2 Ying Zhu2Zhen-Hua Ling1† 1University of Science and Technology of China, Hefei, China 2Baidu Inc., Beijing, China hebl@mail.ustc.edu.cn ,zhling... | https://arxiv.org/abs/2505.21940v1 |
(Wei et al., 2022b; Wang et al., 2023a; Yu et al., 2023), are employed to address MHQA by split complex problems into smaller, thereby harnessing the reasoning poten- tial of LLMs. However, these methods often lack external knowledge, resulting in key evidence be- ing overlooked and generate hallucinations (Rawte et al... | https://arxiv.org/abs/2505.21940v1 |
lower usage costs. Our main contributions are as follows: •We propose RISE, which combines RAG and self-iteration to address two key challenges in MHQA tasks: Evidence Aggregation Errors and Reasoning Decomposition Errors. •We design self-exploration mechanism, con- verts MHQA in RAG into multi-objective op- timization... | https://arxiv.org/abs/2505.21940v1 |
and relationships betweenquestions and sub-questions, thereby improving its ability to decompose complex problems. Retrieve-then-Read. This task follows the stan- dard RAG paradigm to provide evidence-based an- swers for sub-questions. At the t-th exploration node, a retriever obtains relevant fragments rt based on the... | https://arxiv.org/abs/2505.21940v1 |
pand and collect the three types of dataset. Subse- quently, we employ self-exploration mechanism to automatically expand and collect Dd,Dr, andDc datasets for subsequent model training. Multi-Objective Optimization. These three datasets, Dd,Dr, andDc, are interconnected, with sample sizes ranging from 2k to 8k (detail... | https://arxiv.org/abs/2505.21940v1 |
et al., 2023a) and GenRead (Yu et al., 2023), while the retrieval-based methods consist of Naive RAG, Self-Ask (Press et al., 2023), WebGLM (Liu et al., 2023), Self-RAG (Asai et al., 2023), RRR (Ma et al., 2023), and GenGround (Shi et al., 2024a). In 10203040Accuracy (%) (a): Accuracy per Iteration round0 round1 round2... | https://arxiv.org/abs/2505.21940v1 |
RISE with mainstream MHQA methods. Second, we conduct an in-depth analysis of the performance of question decomposi- tion, retrieve-then-read, and self-critique using ob- jective metrics and AI-based evaluations. Finally, we conduct ablation studies to verify the impor- tance of different tasks in enhancing performance... | https://arxiv.org/abs/2505.21940v1 |
models often struggle to integrate logical information from extensive evidence, espe- cially in filtering irrelevant content. To evaluate the changes in the model’s summarization capability over iterations, we disable the decomposition func- tionality and instead allow model to perform single- round retrieval and direc... | https://arxiv.org/abs/2505.21940v1 |
read, and self-critique tasks. Compared to joint training (RISE), the accuracy of separate training is consistently lower across all datasets. 5 Related Works Multi-hop Question Answering: MHQA tasks ad- dress questions that require integrating information from multiple sources and performing multi-step reasoning to pr... | https://arxiv.org/abs/2505.21940v1 |
Our method addresses this gap by integrating self- exploration into RAG to generate diverse training data, enabling continuous model evolution and en- hancing performance in complex tasks. 6 Conclusion We propose RISE, a framework that addresses two key errors in MHQA tasks: Evidence Aggrega- tion and Reasoning Decompo... | https://arxiv.org/abs/2505.21940v1 |
Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv e-prints , pages arXiv–2407. Zhangyin Feng, Xiaocheng Feng, Dezhi Zhao, Maojin Yang, and Bing Qin. 2024. Retrieval-generation syn- ergy augmented large language models. In ICASSP 2024 - 2024 IEEE International Conference on Acou... | https://arxiv.org/abs/2505.21940v1 |
the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 4549–4560. Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jian- guang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wiz- ardmath: Empowering mathematical reasoning for large language models via reinforced ev... | https://arxiv.org/abs/2505.21940v1 |
single-hop question composition. Transactions of the Association for Computational Linguistics , 10:539–554. Jonathan Uesato, Nate Kushman, Ramana Kumar, Fran- cis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. Solving math word problems with process- and outcome- based feedba... | https://arxiv.org/abs/2505.21940v1 |
Wang, Yichong Xu, Mingxuan Ju, S Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. 2023. Generate rather than retrieve: Large language models are strong context 11 generators. In International Conference on Learning Representations . Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi,... | https://arxiv.org/abs/2505.21940v1 |
the main problem? Does it avoid redundant issues that deviate from the primary objective? A.2 Experiment detail A.2.1 Implementation Details We conduct all experiments on a server equipped with four NVIDIA A800 80G GPUs. For the ex- perimental setup, we use the following hyperpa- rameters: learning rate of 1×10−4, batc... | https://arxiv.org/abs/2505.21940v1 |
each approach. RISE demonstrates a higher input token consumption due to its multi-step reasoning, but maintains efficient reasoning performance. improvements observed across multiple datasets after each iteration. This confirms that RISE main- tains strong generalization capabilities and stable performance even when a... | https://arxiv.org/abs/2505.21940v1 |
28.65% 7.58% 23.43% 40.33% Table 10: Alignment between model self-judgment and ground truth correctness on initial question set Q0. Method ModelMHQA SHQA 2WIKI Hotpot MSQ NQ WebQ Trival w/o retrieval Naive LLM LLaMA-3.1-8B 0.00 0.30 0.00 0.25 0.00 2.25 GPT-3.5-turbo 0.50 5.50 0.20 2.50 0.75 26.75 CoT (Wei et al., 2022b... | https://arxiv.org/abs/2505.21940v1 |
this format: ’Follow up: question’ - Ensure each follow-up question is direct and structured to be easily searchable, focusing on key information for efficient search engine retrieval. - For each answer to a follow-up question, use exactly this format: ’Intermediate answer: answer’ - Do not repeat or alter any previous... | https://arxiv.org/abs/2505.21940v1 |
return ’flag = False’ as this information is redundant. - Step 2: Assess Relevance. If the information is not a duplicate, analyze its relevance to the main question. Determine whether it provides new, relevant information that helps move closer to solving the main question, even if it only provides indirect context or... | https://arxiv.org/abs/2505.21940v1 |
Results 1 Decomposition Results 2 Tie" } # Problem: {problem} # Decomposition Results to Compare: - Decomposition Results 1: {result1} - Decomposition Results 2: {result2} # Output: Figure 9: GPT-4o decomposition prompt template. 20 Multi-Hop Question Generation Prompt #Multi-Hop-Question-Generation-in-{Task}# Instruct... | https://arxiv.org/abs/2505.21940v1 |
final answer. Ensure all follow-up questions are optimized for search engine queries, making each question concise, direct, and easily searchable. Avoid modifying or repeating any existing content. — Question: Are the directors of both films Inter Nos and La Bandera (Film) from the same country? Are follow up questions... | https://arxiv.org/abs/2505.21940v1 |
if it only provides indirect context or background.Note that information does not need to directly answer the main question to be considered relevant; it can also support understanding or provide necessary context. Mark it as “flag = True. - Step 3: Based on your analysis, provide a final judgment in the following form... | https://arxiv.org/abs/2505.21940v1 |
arXiv:2505.21941v1 [cs.CL] 28 May 2025Test-Time Scaling with Repeated Sampling Improves Multilingual Text Generation Ashim Gupta Vivek Srikumar Kahlert School of Computing University of Utah ashim@cs.utah.edu Abstract Inference-time scaling via repeated sampling has shown promise in reasoning tasks, but its ef- fective... | https://arxiv.org/abs/2505.21941v1 |
response alignment (via reward models), we aim to identify outputs that are more consistent with human judgments or benchmark standards. We show empirically that repeated sampling with verification improves multilingual text genera- tion. Using two perplexity- and reward-based veri- fiers each, we observe that output q... | https://arxiv.org/abs/2505.21941v1 |
report the difference ( delta ) between win and loss rates, averaged across all lan- guages. All evaluations use gemini-2.0-flash as the judge, which we choose because of its cost efficiency and strong multilingual performance. Our baseline consists of single-sample genera- tions without repeated sampling, using temper... | https://arxiv.org/abs/2505.21941v1 |
are typically optimized to align with either their own reward signals or their internal fluency preferences, suggesting they might better self-evaluate. Nonetheless, Llama verifiers appear well-suited to identifying strong outputs from re- lated model families. 3.Gemma-based reward model delivers the highest gains. The... | https://arxiv.org/abs/2505.21941v1 |
performance gains are robust across models, verifier types, and languages, with notable improvements achievable even for small- scale verifiers. Crucially, we find that verifier ef- fectiveness can be task and language-dependent: perplexity-based methods work well for open- ended prompts but underperform on domains re-... | https://arxiv.org/abs/2505.21941v1 |
llms via reinforcement learning. arXiv preprint arXiv:2501.12948 . Srishti Gureja, Lester James V Miranda, Shayekh Bin Islam, Rishabh Maheshwary, Drishti Sharma, Gusti Winata, Nathan Lambert, Sebastian Ruder, Sara Hooker, and Marzieh Fadaee. 2024. M-rewardbench: Evaluating reward models in multilingual settings. arXiv ... | https://arxiv.org/abs/2505.21941v1 |
(ru), Telugu ( te), Turkish ( tr), and Chinese ( zh). To get the Punjabi subset, we translate Hindi exam- ples to Punjabi using Google Translate followed by manual post-editing. Second dataset is the m- ArenaHard dataset which is the multilingual variant of ArenaHard-Auto (Li et al., 2024). This dataset was released by... | https://arxiv.org/abs/2505.21941v1 |
model for calculating win/loss rates. For a subset of our experiments on Aya Evaluation Suite with Llama-based verifiers, we found that on average the delta scores from GPT-4o and Gemini were within 3.0% of each other. Gemini model is 25x cheaper than the GPT-4o model and therefore considering the scale of our evaluati... | https://arxiv.org/abs/2505.21941v1 |
32.79 -2.00 -3.20 19.15 28.40 25.60 25.60 8.54 13.57 27.13 16.00 11.60 39.65 37.60 -3.61 13.60 -2.01 13.00 3.61 13.50 10.00 20.35 14.06Aya Expanse 8B ar en hi pa pt ru te tr zh Language19.60 15.60 -3.00 0.00 32.40 4.00 -10.40 35.02 35.20 28.40 24.00 0.00 1.00 32.66 4.50 0.80 36.44 35.20 17.60 32.13 1.50 29.15 29.20 16.... | https://arxiv.org/abs/2505.21941v1 |
1 Resolving Knowledge Conflicts in Domain-specific Data Selection: A Case Study on Medical Instruction-tuning Qihuang Zhong, Member, IEEE, Liang Ding, Member, IEEE, Fei Liao, Juhua Liu, Member, IEEE, Bo Du, Senior Member, IEEE, and Dacheng Tao, Fellow, IEEE Abstract —Domain-specific instruction-tuning has become the de... | https://arxiv.org/abs/2505.21958v1 |
deita” are from DEITA [ 10], and the metrics in red are ours. The y-axis denotes the average performance of tuned LLaMA models on several medical benchmarks, where the details are shown in Section IV. domain, they often struggle to handle the domain-specific tasks, e.g., medical question answering [ 6]. To enhance the ... | https://arxiv.org/abs/2505.21958v1 |
representative domain-specific application ( i.e., medical instruction-tuning) as the testbed, and evaluate the LLaMA3 [ 2] and Qwen2.5 [ 3] models tuned with KDS on a variety of medical question-answering benchmarks. Extensive results show that our KDS not only surpasses the other DS methods by a clear margin, but als... | https://arxiv.org/abs/2505.21958v1 |
high-quality and desired subset appears to be crucial in domain-specific instruction-tuning. B. Data Selection for Instruction-tuning In the general-domain instruction-tuning, many data-centric DS methods [ 11], [10], [9] have been proposed, which aim to select the high-quality and diverse data via heuristic methods (e... | https://arxiv.org/abs/2505.21958v1 |
proposed conflict detection methods are simply based on ICL or prompt engineering, which is sensitive to few-shot examples and might introduce bias into the results [ 39], [40]. On the other hand, the method [ 16] mainly focuses on the multiple-choice QA settings and might fall short in free-style generation tasks. To ... | https://arxiv.org/abs/2505.21958v1 |
. . High Knowledge Alignment Question: Which cranial nerve provides innervation to the levator palpebrae superioris muscle? Answer: The levator palpebrae superioris muscle is innervated by the oculomotornerve (cranialnerve III). Multiple Responses: 1.Truthfully... The cranial nerve that provides innervation to the leva... | https://arxiv.org/abs/2505.21958v1 |
shows a high uncertainty and may yield divergent responses. That is, higher uncertainty between multiple responses generally refers to larger knowledge conflicts. Here, to quantitatively evaluate the uncertainty, we propose a cluster-based knowledge consistency metric. Let pij=1 mbe the uniform probability for j-th res... | https://arxiv.org/abs/2505.21958v1 |
training dataset D={Q,A}, base LLM Mintial , data budget k, quality filter threshold τ, diversity filter threshold λ 2:Output: The selected subset S 3: Initialize Empty Dataset S 4:forEach sample (q, a)∈ D do 5: Obtaining multiple responses of Mintial forq 6: Calculating ScoreKAin Eq. 1 or ScoreKCin Eq. 2 7:end for 8:S... | https://arxiv.org/abs/2505.21958v1 |
, and Efficiency . The statistics of all evaluation datasets in the main experiments are shown in Table II, where each task is described as follows: •MedMCQA [48]: It consists of 4-option multiple-choice QA samples from the Indian medical entrance examina- tions (AIIMS/NEET). This dataset covers 2.4K healthcare topics ... | https://arxiv.org/abs/2505.21958v1 |
: We randomly sample 5K data from the instruction-tuning training dataset and fine-tune the LLMs with this data. This baseline is used as the vanilla DS. •IFD [9]: Following the original paper [ 9], we first calculate the Instruction Following Difficulty (IFD) scores for each data point of the instruction-tuning traini... | https://arxiv.org/abs/2505.21958v1 |
51.07 KDS-KC 35.17 56.42 62.53 75.00 71.11 78.49 84.72 68.21 87.00 79.04 51.20 KDS-KA+KC 35.30 56.04 62.84 76.20 74.07 78.11 85.42 68.21 86.00 76.47 51.40 Compared Results upon Qwen-2.5-14B-Instruct Base 24.03 63.61 69.84 78.20 75.30 83.77 89.58 75.72 88.00 83.46 53.05 Full-SFT 36.63 62.90 69.05 76.60 72.59 83.02 90.97... | https://arxiv.org/abs/2505.21958v1 |
tuned models on the long-form medical QA bench- mark [ 52] and illustrate the results in Figure 4. Specifically, we show the winning rates of our method ( KDS-KA+KC) against other baseline DS methods in the figure. The LLaMA3-8B- Instruct is used as the base model in this experiment. As seen, compared to other baseline... | https://arxiv.org/abs/2505.21958v1 |
role in the KDS method; 2) removing each data filtering strategy will lead to performance degradation and the full KDS performs best. Specifically, when using KAmetric in KDS, “-w/o quality ” and “ -w/o diversity ” lead to 0.23 and 0.29 average performance drops, respectively. These results prove the effectiveness of t... | https://arxiv.org/abs/2505.21958v1 |
to performance degradation, as many helpful samples might be ignored. KDS performs best with τ= 3, thus leaving as our default experimental settings. 4) Impact of diversity threshold λ:The factor λ, which is used to control the data diversity, also needs to be investigated. Here, we conduct contrastive experiments to a... | https://arxiv.org/abs/2505.21958v1 |
adaptation, thus resulting in better generalization. To verify it, we further analyze the effect of KDS from the following aspects: 1) Multilingual Generalization: We evaluate the medical LLMs trained with different methods on the popular multilin- gual medical QA benchmarks, i.e., MMedBench [ 57], and illustrate the c... | https://arxiv.org/abs/2505.21958v1 |
field, which mainly assesses the knowledge and reasoning abilities of LLMs in financial scenarios. More specifically, FinanceIQ covers 10 major financial categories and 36 minor financial categories, with a total of 7,173 multiple-choice questions. Table VI shows the comparative results of LLaMA3-8B models tuned with d... | https://arxiv.org/abs/2505.21958v1 |
them as NLI models in KDS. Task xsmall base large NLI accuracy 79% 85% 89% TABLE VIII: Performance of NLI models with varied model sizes on the medical-domain test set . We manually label 100 pairs of answers and model responses as the test set. Notably, since DeBERTa-v3-large-mnli has achieved remarkable performance a... | https://arxiv.org/abs/2505.21958v1 |
more realistic estimate of what a person can expect to live to based on current trends and data. Case 2 Question: What is the average maximum muscle weight percentage for a 6-foot-tall male that weighs 210 pounds? GPT-4o: . . . Muscle mass percentage: For a very muscular person (such as a competitive bodybuilder), musc... | https://arxiv.org/abs/2505.21958v1 |
of our KDS, LLM can achieve more effective domain adaptation and output more professional and accurate responses. These results continue proving the effectiveness and superiority of our KDS method. VI. C ONCLUSION In this paper, we focus on the knowledge conflict problem in the domain-specific instruction-tuning, which... | https://arxiv.org/abs/2505.21958v1 |
Linguistics: Human Language Technologies (Volume 1: Long Papers) , 2024. [10] W. Liu, W. Zeng, K. He, Y . Jiang, and J. He, “What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning,” in The Twelfth International Conference on Learning Representations , 2024. [11] L. C... | https://arxiv.org/abs/2505.21958v1 |
L. Bing, and W. Lam, “A survey on aspect-based sentiment analysis: Tasks, methods, and challenges,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 11, pp. 11 019–11 038, 2022. [24] Q. Zhong, L. Ding, J. Liu, B. Du, H. Jin, and D. Tao, “Knowledge graph augmented network towards multiview representati... | https://arxiv.org/abs/2505.21958v1 |
2024) , 2024. [36] Q. Yu, J. Merullo, and E. Pavlick, “Characterizing mechanisms for factual recall in language models,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , 2023. [37] H. Ding, Y . Fang, R. Zhu, X. Jiang, J. Zhang, Y . Xu, X. Chu, J. Zhao, and Y . Wang, “3ds: Deco... | https://arxiv.org/abs/2505.21958v1 |
International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , 2019. [51] D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt, “Measuring massive multitask language understanding,” in International Conference on Learning Representations , 2020. [52] P. Hosseini, J. M. Sin, ... | https://arxiv.org/abs/2505.21958v1 |
arXiv:2505.21964v1 [cs.HC] 28 May 2025UI-Evol: Automatic Knowledge Evolving for Computer Use Agents Ziyun Zhang* 1Xinyi Liu* 1Xiaoyi Zhang2Jun Wang2Gang Chen2Yan Lu2 Abstract External knowledge has played a crucial role in the recent development of computer use agents. We identify a critical knowledge-execution gap: re... | https://arxiv.org/abs/2505.21964v1 |
can significantly improve agent performance. Nevertheless, our analysis of Agent S2 (Agashe et al., 2025) based on GPT-4o reveals a persistent gap between the avail- ability of correct knowledge and the knowledge can be effec- tively consumed by agent for task execution. Specifically, in our sampling survey, even when ... | https://arxiv.org/abs/2505.21964v1 |
consume for practical task completion. To address the aforementioned problem, we propose UI- Evol, a plug-and-play module that can be seamlessly inte- grated into existing computer use agent systems by introduc- ing an autonomous GUI knowledge evolution mechanism aimed at improving knowledge through realistic interac- ... | https://arxiv.org/abs/2505.21964v1 |
edge and real task execution, and propose UI-Evol, a plug-and-play module that effectively bridges this gap by autonomously evolving GUI task knowledge. •We are the first to systematically identify and analyze the previously overlooked instability issue in contemporary computer use agents, and develop a highly parallel... | https://arxiv.org/abs/2505.21964v1 |
of knowledge (Chen et al., 2023; Zelikman et al., 2024). More commonly, refine- ment is driven by an additional critique process, including critiques independently generated by LLM itself (Lu et al., 2023; Madaan et al., 2023), or critiques produced during interactions between LLM and external tools such as code interp... | https://arxiv.org/abs/2505.21964v1 |
the two stages separately. 4.1. Retrace Stage Due to the intrinsic hallucination tendency and limited UI perception capabilities of Large Multimodal Models, exist- ing computer use agents often generate infeasible actions or incorrectly interpret the current computer state. Moreover, the inherent complexity of the comp... | https://arxiv.org/abs/2505.21964v1 |
Daily Office Professional Workflow Avg. SR. Web Search GPT-4o 51.39 23.98 14.27 32.43 9.11 19.5 Evolved from 4o Traj. GPT-4o 47.22 27.83 17.98 35.61 10.43 22.0 Evolved from o3 Traj. GPT-4o 48.61 26.12 23.09 31.33 8.78 22.4 liminary sampling analysis has demonstrated that the re- trieved web-based knowledge generally re... | https://arxiv.org/abs/2505.21964v1 |
as the original knowledge, but is systematically refined and enhanced. Ultimately, upon completing these carefully designed rea- soning steps, the LLM generates an updated and refined knowledge representation that explicitly incorporates the identified corrections, clarifications, and supplementary al- ternative strate... | https://arxiv.org/abs/2505.21964v1 |
the be- ginning of each run, we capture and freeze a snapshot of the entire knowledge base before the first experiment while set all hyperparameters as constant value including temperature as 0. This eliminates variability caused by dynamic web. Stability. During experiments, we observe that even with static precapture... | https://arxiv.org/abs/2505.21964v1 |
UI- Evol can effectively enhance both agent performance and stability. Notably, when integrating UI-Evol, the standard deviation of OpenAI-o3 drops to as low as 0.26, approxi- mately 4.19 times the reduction observed with GPT-4o. This suggests that models with stronger reasoning capabilities are better at understanding... | https://arxiv.org/abs/2505.21964v1 |
the trajectory is fed into UI- Evol, the system first enters the Retrace Stage, where each step is distilled into two textual elements: Action and Result. In the first Retrace step, UI-Evol analyzes the pre-execution screenshot and summarizes the initial state, correctly noting that the agent selected only the section ... | https://arxiv.org/abs/2505.21964v1 |
Evaluating multi-modal os agents at scale. arXiv preprint arXiv:2409.08264 , 2024. Chen, X., Lin, M., Schaerli, N., and Zhou, D. Teaching large language models to self-debug. In The 61st Annual Meet- ing Of The Association For Computational Linguistics , 2023. Cheng, K., Sun, Q., Chu, Y ., Xu, F., YanTao, L., Zhang, J.... | https://arxiv.org/abs/2505.21964v1 |
arXiv preprint arXiv:2504.11257 , 2025a. Liu, Y ., Li, P., Wei, Z., Xie, C., Hu, X., Xu, X., Zhang, S., Han, X., Yang, H., and Wu, F. Infiguiagent: A multimodal generalist gui agent with native reasoning and reflection. arXiv preprint arXiv:2501.04575 , 2025b. Liu, Y ., Li, P., Xie, C., Hu, X., Han, X., Zhang, S., Yang... | https://arxiv.org/abs/2505.21964v1 |
C., Zhong, V ., and Yu, T. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. Xu, Y ., Wang, Z., Wang, J., Lu, D., Xie, T., Saha, A., Sahoo, D., Yu, T., and Xiong, C. Aguvis: Unified pure vision agents for autonomous gui interaction. arXiv preprint arXiv:2412.04454 , 2024.... | https://arxiv.org/abs/2505.21964v1 |
25.12 21.80 9.22 20.64 30.43 10.43 47.22 44.44 20.27 47.83 22.02 + UI-Evol w/Comp. Select 27.38 17.95 11.35 18.52 43.47 11.43 44.45 44.44 26.14 40.58 22.74 + UI-Evol w/knowledge evolved from o3 21.02 19.23 16.31 26.29 30.43 8.78 48.61 44.45 23.76 42.03 22.38 Table 5. Performance (%) of different settings across individ... | https://arxiv.org/abs/2505.21964v1 |
replacing all 3 occurrences of "foo" with "bar" in the document ,→ <END_EXAMPLE> <BEGIN_EXAMPLE> # Only the clock changed BEFORE: Desktop 10:01 AFTER : Desktop 10:02 OUTPUT: [A] BEFORE Desktop environment showing wallpaper and system clock reading 10:01. [B] OPERATIONS - No operations performed. <END_EXAMPLE> <BEGIN_EX... | https://arxiv.org/abs/2505.21964v1 |
Evolving for Computer Use Agents For every Root Cause from SECTION B, describe the preventive or corrective idea and mention which forthcoming step embodies it. ,→ Example: c) Syntax error →Add \lint before run" check (handled in Step 2). d) Permission →Verify sudo rights before executing installer (Step 5). f) Invalid... | https://arxiv.org/abs/2505.21964v1 |
arXiv:2505.21967v1 [cs.CL] 28 May 2025Seeing the Threat: Vulnerabilities in Vision-Language Models to Adversarial Attack Juan Ren, Mark Dras, Usman Naseem School of Computing, Macquarie University, Australia, ada.ren@hdr.mq.edu.au, {mark.dras,usman.naseem}@mq.edu.au Abstract Large Vision-Language Models (LVLMs) have sh... | https://arxiv.org/abs/2505.21967v1 |
benchmarks, a critical gap remains: Why do cross-modal adversarial attacks succeed, and how should their success be rigorously evaluated? Many recent works report high attack success rates without a consistent definition of what constitutes success, complicating the comparison of methods and undermining reproducibility... | https://arxiv.org/abs/2505.21967v1 |
when jointly interpreted, convey malicious intent. Type IV: Implicit Harmful Intent via Modal- ity Interaction. Some attacks hide unsafe se- mantics in subtle interactions between modalities. SIUO (Wang et al., 2025) and MSSBench (Zhou et al., 2025) show that benign inputs—when strate- gically combined—can lead LVLMs t... | https://arxiv.org/abs/2505.21967v1 |
helpful—but bounded—responses, often with disclaimers. These models aim to inform re- sponsibly rather than categorically refuse, particu- larly in domains such as law, medicine, and finance. This mismatch complicates the interpretation of safety benchmarks. To address these gaps, we propose a structured evaluation pro... | https://arxiv.org/abs/2505.21967v1 |
the harm in successful attacks, we introduce a Likert-based scoring metric. Each response is rated on a 5-point scale and normalized between 0 and 1: ASR score=Norm (MajorityV ote (LikertScale )) Refusal responses are further categorized into: (1) Hard Refusal : Clear and explicit denials; (2) Soft Refusal : Indirect o... | https://arxiv.org/abs/2505.21967v1 |
which are implicitly related to fraction informa- tion from the iamge ize the outcome in Figure 3. Interestingly, token projections frequently con- tain semantically meaningful fragments—such as “method,” “illegal,” and “organ”—despite no direct text recognition. This indicates that while LLaV A fails to extract litera... | https://arxiv.org/abs/2505.21967v1 |
for final deci- sions. Instruction-Following Behavior. LLaMA 3.2 Vision displays notable deficiencies in instruction- following, with a failure rate of 76%. This is often due to misclassifying benign prompts as identity- related, leading to generic refusals. LLaV A simi- larly exhibits limited instruction-following, li... | https://arxiv.org/abs/2505.21967v1 |
benchmarks such as MMS AFETY andFIG- STEPexpect blanket refusals, commercial deploy- ments favor contextual answers with warnings. As shown in Figure 8, removing "professional advice" examples from scoring reduces both ASR and ASR Quality scores, highlighting these as ma- jor sources of evaluative disagreement. Ideal R... | https://arxiv.org/abs/2505.21967v1 |
https://www. anthropic.com/news/claudes-constitution . Accessed: 2025-05-20. Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wen- bin Ge, Sibo Song, Kai Dang, Peng Wang, Shi- jie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, and 8 oth- ... | https://arxiv.org/abs/2505.21967v1 |
Jeff Wu, Xu Jiang, Diogo Almeida, Car- roll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow i... | https://arxiv.org/abs/2505.21967v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.