text string | source string |
|---|---|
this section, we describe how we identify community pairs, col- lect data, generate instruction-response demonstra- tions I= (xj, yj)for steering models, and build multiple-choice evaluation instances for assessing whether a model steered toward a community C accurately reflects its views. 2In-context learning does not... | https://arxiv.org/abs/2505.20645v1 |
a synthetic data generator. For each topic tk∈T, we sample 50 comments from DAandDB, anonymize subreddit names as “r/A” and “r/B”, and prompt GPT-4o to generate three instructions { xj, xj+1, xj+2}, that elicit contrast- ing viewpoints across the two communities. See Figure 5 in Appendix A for the prompting tem- plate.... | https://arxiv.org/abs/2505.20645v1 |
this task with full awareness that their annotations would only be used to evaluate the performance of GPT-4o’s generation. The evaluation is conducted via Google Forms, with 30 sections per form, each for a randomly sampled topic from a subreddit pair. For each sec- tion, we sample two instructions and one multiple- 5... | https://arxiv.org/abs/2505.20645v1 |
set to 8e-6 for 3B models and 6e-6 for larger ones. Training is con- ducted on 8 NVIDIA H100 GPUs. 5.3 Evaluation Protocol Given a steered f′(via in-context or finetuning), we present it with a set of community-specific multi- choice questions Q={(qk, ak)}. Each question qktargets a topic tkdiscussed by community C, an... | https://arxiv.org/abs/2505.20645v1 |
and Reli- gion emerge as domains where nearly all models perform well, with top models achieving >0.70 ac- curacy. These domains feature clearly articulated community norms (e.g., keto vs. vegan, GetMo- tivated vs. getdisciplined) and strongly polarized rhetoric, which likely facilitates easier identifica- tion of comm... | https://arxiv.org/abs/2505.20645v1 |
result, the benchmark may not generalize to populations that are less active online or are better represented on other platforms, such as X, Weibo, TikTok, or regional forums. Bias in GPT-4o-generated supervision Instruction-response pairs and multiple-choice questions are generated using GPT-4o, which may introduce bi... | https://arxiv.org/abs/2505.20645v1 |
2023. Steering large language models for machine translation with finetuning and in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2023 , pages 11127–11148, Singapore. Association for Computa- tional Linguistics. Reza Bayat, Ali Rahimi-Kalahroudi, Mohammad Pezeshki, Sarath Chandar... | https://arxiv.org/abs/2505.20645v1 |
instruction follow- ing ability of large language models. arXiv preprint arXiv:2404.15846 . Zihao He, Minh Duc Chu, Rebecca Dorn, Siyi Guo, and Kristina Lerman. 2024b. Community-cross- instruct: Unsupervised instruction generation for aligning large language models to online commu- nities. In Proceedings of the 2024 Co... | https://arxiv.org/abs/2505.20645v1 |
the 2016 Conference on Empirical Methods in Natu- ral Language Processing , pages 2383–2392. Yiting Ran, Xintao Wang, Rui Xu, Xinfeng Yuan, Ji- aqing Liang, Yanghua Xiao, and Deqing Yang. 2024. Capturing minds, not just words: Enhancing role- playing language models with personality-indicative data. In Findings of the ... | https://arxiv.org/abs/2505.20645v1 |
for Computational Linguistics: NAACL 2024 , 11 pages 3712–3729, Mexico City, Mexico. Association for Computational Linguistics. Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. 2024. WizardLM: Empow- ering large pre-trained language models to follow comple... | https://arxiv.org/abs/2505.20645v1 |
186,486 personalfinance_wallstreetbets 489,798 apple_Android 425,302 linux_windows 200,278 Parenting_childfree 489,408 keto_vegan 497,936 carnivore_vegetarian 44,210 realmadrid_Barca 496,762 warriors_lakers 499,840 xbox_playstation 495,315 leagueoflegends_DotA2 491,964 simpleliving_UnethicalLifeProTips 499,531 environm... | https://arxiv.org/abs/2505.20645v1 |
toprovide appropriate responses to{domain }-related questions .The instruction - response pairs below demonstrate responses from asubreddit r/{subreddit }.These examples reflect :1.Perspectives common within aspecific ideological community .2.Language and framing typical ofthis particular viewpoint . 3.Arguments and re... | https://arxiv.org/abs/2505.20645v1 |
savings, retirement 1 she, was, relationship, with, friend, did, back, for, it, but 2 age, older, women, gap, young, dating, mature, 30s, olds, attractive 4 single, dating, meet, life, alone, relationships, date, yourself, be, want 5 he, ex, we, relationship, with, ended, back, wasn, time, friend 7 movies, watched, sho... | https://arxiv.org/abs/2505.20645v1 |
shootings, firearms, weapons, amendment, firearm, laws, militia, rifles 3 israel, hamas, palestinians, palestinian, gaza, palestine, genocide, israelis, conflict, hostages 4 religious, god, christianity, christians, bible, religions, commandments, catholic, believe, atheist 5 ukraine, russia, nato, putin, war, ukrainia... | https://arxiv.org/abs/2505.20645v1 |
getting, hours, enjoy, games 4 ssd, storage, 2tb, expansion, hdd, 4tb, heatsink, install, seagate, ps5 5 xbox, gen, console, games, 360, 4k, disc, upgrade, performance, storage 6 ps2, ps1, playstation, nes, owned, sega, memories, n64, atari, snes 7 xbox, exclusives, microsoft, platform, consoles, sony, market, nintendo... | https://arxiv.org/abs/2505.20645v1 |
19 religion, religions, people, control, society, world, power, organized, we, masses Table 4: Top ten keywords for topics across four contrasting subreddit pairs in Gaming, Religion domains. 19 Subreddit Pair Topic Index Topic Keywords exmuslim_islam0 age, aisha, puberty, child, girl, nine, muhammad, pedophilia, marry... | https://arxiv.org/abs/2505.20645v1 |
self-improvement domains. 20 Subreddit Pair Topic Index Topic Keywords realmadrid_Barca0 he, was, ball, injury, player, season, been, goals, as, goal 1 xavi, coach, laporta, manager, season, he, club, team, stay, has 2 ref, var, refs, foul, offside, penalty, referee, referees, negreira, madrid 3 we, half, game, score, ... | https://arxiv.org/abs/2505.20645v1 |
ssds, install, sata, gparted Table 6: Top ten keywords for topics across four contrasting subreddit pairs in sports, Technology domains. 21 Subreddit Pair Topic Index Topic Keywords antiwork_ WorkReform0 rent, housing, homes, property, houses, mortgage, landlords, apartment, income, market 1 insurance, healthcare, medi... | https://arxiv.org/abs/2505.20645v1 |
having, sister 4 cat, pet, pets, animal, love, my, have, are, vet, don 5 names, call, nickname, mama, grandma, his, nicknames, change, mr, use 6 party, birthday, parties, cake, birthdays, friends, invites, family, host, attend 7 she, wants, abortion, if, will, it, decision, is, tell, for 8 gifts, gift, christmas, toys,... | https://arxiv.org/abs/2505.20645v1 |
what 2 air, pan, cook, fryer, steak, sear, oven, sous, iron, cooking 4 milk, dairy, cream, cheese, yogurt, oat, butter, kefir, coffee, raw Teachers_ homeschool4 covid, immune, flu, pandemic, masks, illness, vaccines, wash, air, system 5 math, calculator, memorization, fractions, division, teach, basic, memorize, calcul... | https://arxiv.org/abs/2505.20645v1 |
16 384 238 18 18 Self-improvementGetMotivated vs. getdisciplined 19 456 300 9 9 DecidingToBeBetter vs. howtonotgiveafuck 7 168 112 99 99 Sportsrealmadrid vs. Barca 14 336 224 36 36 warriors vs. lakers 10 240 160 45 45 Technologyapple vs. Android 14 336 219 54 54 linux vs. windows 8 192 128 99 99 social issues antiwork ... | https://arxiv.org/abs/2505.20645v1 |
0.694 Llama-3.2-3B 0.281 0.353 0.360 0.306 0.305 0.372 0.373 0.417 0.473 Llama-3.1-8B 0.500 0.572 0.507 0.520 0.529 0.588 0.633 0.479 0.629 Llama-3.3-70B 0.688 0.678 0.640 0.642 0.597 0.711 0.709 0.500 0.684 Mistral-7B-v0.3 0.500 0.541 0.499 0.505 0.498 0.586 0.595 0.521 0.602 Claude-3.5-Haiku 0.594 0.556 0.496 0.510 0... | https://arxiv.org/abs/2505.20645v1 |
arXiv:2505.20650v1 [cs.CL] 27 May 2025FinTagging: An LLM-ready Benchmark for Extracting and Structuring Financial Information Yan Wang The Fin AI USAYang Ren The Fin AI USA Lingfei Qian The Fin AI USAXueqing Peng The Fin AI USAKeyi Wang Columbia University USAYi Han Georgia Institute of Technology USA Dongji Feng Gusta... | https://arxiv.org/abs/2505.20650v1 |
fine-grained financial facts. These datasets typically cover only 1k+ concepts, leaving most of the 10k+ US-GAAP taxonomy untested. As shown in Table 8, SOTA LLMs such as DeepSeek-V3 [ 13] and GPT-4o [ 11] achieve 0.0 precision, recall, and F1 under this setting. Second, structured data is ignored : existing datasets e... | https://arxiv.org/abs/2505.20650v1 |
state-of-the-art LLMs under a zero-shot setting on three fronts: (1) end-to-end macro-F1 over the full FINTAGGING benchmark, (2) subtask-specific performance on FinNI and FinCL, and (3) ablation of our unified extraction-and-alignment evaluation framework. DeepSeek- V3 [13] and GPT-4o [ 11] achieve the highest macro-F1... | https://arxiv.org/abs/2505.20650v1 |
document understanding tasks. FiNER-ORD [ 22] and FinRED [ 24] focus on entity recognition and relation extraction, respectively. BizBench [ 12] assesses the quantitative- reasoning ability of LLMs for both business and finance. Pixiu [ 33] evaluates LLMs across classifica- tion, QA, and summarization, while FinQA [ 4]... | https://arxiv.org/abs/2505.20650v1 |
defined in the overall task. Formally, we define the mapping as: fFinCL: (e, l, C e,T)7→ˆc (3) where eis a numerical entity identified in the document D= (S, T),l∈Lis its predicted data type from FINNI,Cedenotes the contextual information surrounding einD, andT={c1, c2, . . . , c n}is a financial taxonomy containing nu... | https://arxiv.org/abs/2505.20650v1 |
lacking US-GAAP labels. Then we conducted a statistical analysis of the entity types contained in the data, as summarized in Figure 2, to identify the most frequent types. Figure 2: The statistic of numerical entity type. From Figure 2, we identified 11 numerical entity types across the financial reports of 30 companie... | https://arxiv.org/abs/2505.20650v1 |
a reranking problem, where LLMs are used to disambiguate and select the most appropriate taxonomy concept from a reduced candidate set, avoiding the impracticality of direct multi-thousand-way classification. To this end, we first generate embeddings for each taxonomy concept using text-embedding-3-small7, and retrieve... | https://arxiv.org/abs/2505.20650v1 |
which emphasizes balanced performance across both frequent and rare tags, DeepSeek-V3 and GPT-4o achieve the highest macro-F1 scores (0.0582 and 0.0508), outperforming all fine-tuned PLMs. This highlights the strong generalization of large LLMs and the effectiveness of our task design. DeepSeek-R1-Distill-Qwen-32B also... | https://arxiv.org/abs/2505.20650v1 |
evaluation protocol. By decoupling extraction from concept linking and covering the full 10k+ taxonomy, FinTagging produces meaningful scores and thus offers a far more realistic test bed for future model improvements. 9 Table 8: Performance comparison between w/wo our evaluation framework on the FINTAGGING benchmark d... | https://arxiv.org/abs/2505.20650v1 |
(ICAIF ’24) , 2024. 10 [11] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276 , 2024. [12] Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Loverin... | https://arxiv.org/abs/2505.20650v1 |
nlp, 2025. [30] Yan Wang, Jian Wang, Huiyi Lu, Bing Xu, Yijia Zhang, Santosh Kumar Banbhrani, Hongfei Lin, et al. Conditional probability joint extraction of nested biomedical events: design of a unified extraction framework based on neural networks. JMIR Medical Informatics , 10(6):e37804, 2022. 11 [31] Jason Wei, Xue... | https://arxiv.org/abs/2505.20650v1 |
authors introduced pseudo-token strategies replacing numerals with [NUM] or [SHAPE] tokens to stabilize label assignment across fragmented numeric spans. These strategies, combined with domain- specific pretraining on SEC-BERT, significantly improved tagging performance, reaching 82.1 micro-F1 without the need for comp... | https://arxiv.org/abs/2505.20650v1 |
developed comprehensive benchmarks to assess broader capabilities in information extraction, numerical reasoning, and document understanding. FiNER-ORD [ 22] introduced a high-quality, domain-specific NER dataset annotated over financial news, emphasizing general entity types like persons, organizations, and locations.... | https://arxiv.org/abs/2505.20650v1 |
D Evaluation Metrics To provide a fair evaluation of overall benchmark performance, we adopt a set of metrics, focusing primarily on macro-level and micro-level evaluation strategies inspired by the previous work [ 23].Macro-level evaluation computes precision, recall, and F1 scores independently for each BIO-concept l... | https://arxiv.org/abs/2505.20650v1 |
models in terms of domain understanding and structured output capability. Together, these models offer a comprehensive evaluation spectrum, from general-purpose to domain-specific, encoder-based to decoder-based, and open to closed source, facilitating an in-depth assessment of their perfor- mance across our proposed b... | https://arxiv.org/abs/2505.20650v1 |
with BIO scheme. After constructing the training set, we reconstruct the testing set from the original benchmark dataset. The training settings are detailed below. F.2 Training settings We fine-tune three pretrained models, BERT-large [ 7], FinBERT [ 1], and SECBERT [ 14], on our training set using the HuggingFace Tran... | https://arxiv.org/abs/2505.20650v1 |
setting, likely due to the fragmented and semantically sparse nature of partial table text. 17 When aggregating results across both structures, SAC outperforms FWC by a wide margin at all retrieval depths (e.g., Acc@200 of 0.3163 vs. 0.0608). These results underscore the importance of structure-aware context constructi... | https://arxiv.org/abs/2505.20650v1 |
integerItemType ": Counts of discrete items , such as the number of employees or total transactions . - " monetaryItemType ": Financial amounts expressed in currency , such as revenue , profit , or total assets . - " perShareItemType ": Per - share values , such as earnings per share ( EPS ) or book value per share . -... | https://arxiv.org/abs/2505.20650v1 |
with stronger schema-aware reasoning, a critical step toward trustworthy AI systems in high-stakes financial environments. Through open access to annotated datasets and evaluation code, FINTAGGING also advances reproducible research and supports the broader community in benchmarking financial language models responsibl... | https://arxiv.org/abs/2505.20650v1 |
Chinese Cyberbullying Detection: Dataset, Method, and Validation Yi Zhu1,2,Xin Zou1,Xindong Wu2,3 1School of Information Engineering, Yangzhou University, Yangzhou 225009, China 2Key Laboratory of Knowledge Engineering with Big Data (Hefei University of Technology), Ministry of Education, Hefei 230009, China 3School of... | https://arxiv.org/abs/2505.20654v1 |
such classification does not capture the temporal dynamics or social amplification of cyberbullying incidents, which often escalate rapidly and cause widespread harm before interventions can be deployed. For example, a single offensive comment may not be problematic in isola- tion, but when thousands of similar comment... | https://arxiv.org/abs/2505.20654v1 |
below: (1) We propose a novel annotation approach based on hu- man and machine collaboration to construct a cyberbully- ing dataset that organized by incidents. Our approach pro- vides a good idea for constructing large-scale, high-coverage datasets, especially for tasks involving distinguishing cyber- bullying that ar... | https://arxiv.org/abs/2505.20654v1 |
the MySpace dataset is often employed to develop and evaluate models for cyberbullying detection and to explore gender-based differences in abusive online behavior. Besides English, the German cyberbullying detection dataset originated from the GermEval 2018 and contains around 8,000 instances collected from Twitter, w... | https://arxiv.org/abs/2505.20654v1 |
in images on social media platforms [Al- momani et al. , 2024 ]. By utilizing pre-trained CNN models (e.g., InceptionV3, ResNet50, VGG16) as feature extractors and feeding these features into classifiers such as Logistic Re- gression and SVM, the approach achieves improved detection accuracy. Recently, deep learning me... | https://arxiv.org/abs/2505.20654v1 |
multiple mainstream Chinese social media platforms as data sources, including Douyin1, Weibo2, Xiaohongshu3, and Bili- bili4. These platforms represent the primary arenas of Chi- nese social media, encompassing a broad user base and di- verse forms of discussion. To discover such incidents, we monitored the top-trendin... | https://arxiv.org/abs/2505.20654v1 |
multi-agent-based method generates explanations and introduces a two-layer multi- agent voting strategy for cyberbullying detection. In this ap- proach, a multi-agent system is employed where multiple in- dependent agents collaborate to process the task. Each agent operates based on a distinct prompt template, enhancin... | https://arxiv.org/abs/2505.20654v1 |
likely to reflect potential classifi- cation outcomes, enhancing the reliability and diversity of the labels. 3.3 Manual Annotation To ensure the accuracy of the annotations, we engage mul- tiple annotators for annotation. It is worth mentioning that all the annotators involved in this process are native Chinese underg... | https://arxiv.org/abs/2505.20654v1 |
non-cyberbullying). Agent 3You are an expert in online safety. Review the comment and judge whether it should be classified as cyberbullying. Justify your answer and output the classification result. Agent 4Evaluate the given comment and assess whether it constitutes cyberbullying. Include your reasoning process and co... | https://arxiv.org/abs/2505.20654v1 |
media, was as- signed to assess the accuracy of detecting malicious com- ments within the selected instances. The annotator care- fully examined each instance to determine whether the com- ment was offensive or non-offensive, classifying them ac- cordingly into cyberbullying or non-cyberbullying categories. The accurac... | https://arxiv.org/abs/2505.20654v1 |
”F1-score” as outlined in the SemEval 2007 task. These metrics provide a comprehensive and de- tailed evaluation from multiple perspectives. Implementation Details. Paraphraser-based (Para) and CoT-based (CoT): detections were made directly through a conversational approach with 5-shot training data. For multi- agents-... | https://arxiv.org/abs/2505.20654v1 |
utilizing differ- ent approaches for cyberbullying incidents prediction. The process of validation is illustrated as Figure 5. 6.1 Baseline Methods for Validation Cyberbullying Language Detection. The following nine baselines are selected to validate cyber- bullying language detection on CHNCI, including fine-tuning PL... | https://arxiv.org/abs/2505.20654v1 |
selected to validate cyber- bullying incidents prediction on CHNCI, including the deep learning and transformer-based methods. These models are selected for their strong performance in time series forecast- ing and event trend modeling. Deep Learning Methods: The deep learning methods in- cluding GRU, TCN, and LSTM pro... | https://arxiv.org/abs/2505.20654v1 |
tasks with strong trends and cycles. • FEDformer [Zhou et al. , 2022 ]:A variant of Transformer that combines Fourier transformation and decomposi- tion techniques, enhancing the model’s performance by strengthening frequency domain analysis. It is suitable for complex time series forecasting tasks, such as traffic flo... | https://arxiv.org/abs/2505.20654v1 |
results. 6.3 Validation Results The results of the cyberbullying language detection and cy- berbullying incidents prediction on CHNCI are shown in Ta- ble 6 and Table 7, respectively. It is worth noting that, eachexperiment was conducted three times, and the average and standard deviation were computed. Cyberbullying L... | https://arxiv.org/abs/2505.20654v1 |
( ±0.08) ( ±0.17) HateBert62.59 66.56 68.22 70.29 70.49 71.70 (±0.07) ( ±0.18) ( ±0.31) ( ±0.51) ( ±0.31) ( ±0.52) Conprompt64.97 68.20 72.33 72.23 73.89 73.86 (±0.12) ( ±0.21) ( ±0.44) ( ±0.65) ( ±0.45) ( ±0.63) P-tuning62.90 66.83 68.66 71.49 71.36 73.61 (±0.79) ( ±0.70) ( ±1.30) ( ±0.94) ( ±1.21) ( ±0.95) KPT61.61 6... | https://arxiv.org/abs/2505.20654v1 |
figures are consistent with our cyberbullying incidents prediction. Figure 7 illustrates the differences in word clouds between cyberbullying and non-cyberbullying incidents. In Figure 7(a), high-frequency words such as ”Qingdao prawns,” ”Who do you think you are,” and ”temporary worker” carry strong sarcastic and aggr... | https://arxiv.org/abs/2505.20654v1 |
Almomani, Khalid Nahar, Mohammad Alauthman, Mohammed Azmi Al-Betar, Qus- sai Yaseen, and Brij B Gupta. Image cyberbullying de- tection and recognition using transfer deep machine learn- ing. International Journal of Cognitive Computing in En- gineering , 5:14–26, 2024. [Baiet al. , 2018 ]Shaojie Bai, J Zico Kolter, and... | https://arxiv.org/abs/2505.20654v1 |
Haewoon Kwak, and Jisun An. Chain of explanation: New prompting method to gen- erate quality natural language explanation for implicit hate speech. In Companion Proceedings of the ACM Web Con- ference 2023 , pages 90–93, 2023. [Iwendi et al. , 2023 ]Celestine Iwendi, Gautam Srivastava, Suleman Khan, and Praveen Kumar R... | https://arxiv.org/abs/2505.20654v1 |
Raisi and Bert Huang. Cy- berbullying detection with weakly supervised machine learning. In Proceedings of the 2017 IEEE/ACM Interna- tional Conference on Advances in Social Networks Analy- sis and Mining 2017 , pages 409–416, 2017. [Rajet al. , 2021 ]Chahat Raj, Ayush Agarwal, Gnana Bharathy, Bhuva Narayan, and Mukesh... | https://arxiv.org/abs/2505.20654v1 |
arXiv:2505.20658v1 [cs.CL] 27 May 2025Enhancing Transformation from Natural Language to Signal Temporal Logic Using LLMs with Diverse External Knowledge Yue Fang1, Zhi Jin1, Jie An2, Hongshen Chen3, Xiaohong Chen4,and Naijun Zhan1 1Peking University, Beijing, China 2Institute of Software, Chinese Academy of Sciences, B... | https://arxiv.org/abs/2505.20658v1 |
an inter- mediate representation. Subsequently, by applying a set of predefined rules manually, the intermediate representation is mapped to temporal logic formu- las. These approaches require extensive domain ex- pertise and involve a steep learning curve (Kulkarni et al., 2013). Specifically, they can only be applied... | https://arxiv.org/abs/2505.20658v1 |
in STL transformation tasks. In general, our contributions are as follows: •We develop a dataset, named STL-DivEn, con- taining 16k high-quality NL-STL pairs using LLMs and manual annotation. Compared to the existing DeepSTL dataset, the statistics show that this dataset exhibits significantly greater diversity. •We pr... | https://arxiv.org/abs/2505.20658v1 |
(2024) start with a small set of seed instructions, which are then expanded using in-context learning to generate diverse instruction-response pairs. How- ever, these methods often struggle with ensuring sufficient diversity in the generated data. To ad- dress this, strategies such as iterative generate-filter pipeline... | https://arxiv.org/abs/2505.20658v1 |
formula where temporal operators are applied within the scope of other temporal operators. 4 Approach In this section, we first present our approach for constructing the STL-Diversity-Enhanced (STL- DivEn) dataset, which combines manual annota- tion and LLMs to generate diverse, high-quality data. Second, we introduce ... | https://arxiv.org/abs/2505.20658v1 |
between a new NL-STL pair and all existing seed pairs is below 0.5, the new pair is considered to exhibit sufficient diversity. Next, the NL-STL pairs that pass the rule-based filtering undergo human validation to ensure con- sistency between the natural language and STL specifications. Seven annotators who have been t... | https://arxiv.org/abs/2505.20658v1 |
formulas through randomly sampling from templates and operator distributions, while STL-DivEn is a dataset created using GPT-4 and human annotation. We randomly selected 14,000 samples from each dataset for the training set and 2,000 samples for the test set. Evaluation Measures. To evaluate the results of STL generati... | https://arxiv.org/abs/2505.20658v1 |
AccuracyTemplate AccuracyBLEU DeepSTL 0.2002 0.2916 0.3332 GPT-3.5 0.2145 0.3002 0.2249 GPT-4 0.2262 0.3048 0.2881 DeepSeek 0.2537 0.3254 0.3982 GPT-4+Self-Refine 0.2203 0.3019 0.2682 KGST 0.4538 0.4939 0.5686 (b) DeepSTL Table 1: Metric-based evaluation results. For the DeepSTL dataset, as shown in Table 1b, we also o... | https://arxiv.org/abs/2505.20658v1 |
STL formula indicates that the formulas in theModelSTL Formula AccuracyTemplate AccuracyBLEU KGST 0.5587 0.5627 0.2142 - w/o Fine-tuning 0.5360 0.5390 0.1978 - w/o Refinement 0.4956 0.5007 0.1784 Table 4: Ablation experimental results on STL-DivEn. STL-DivEn dataset have more complex structures. The total word count of... | https://arxiv.org/abs/2505.20658v1 |
generated by GPT-4, F[0,200](z1>1) andG[0,50](z1>1)are used in parallel, but there is no indication of the sequential relationship. The correct logic should specify that z1>1must first occur, followed by its persistence for 50time units. These results confirm that KGST effectively cor- rects errors in the generated STL... | https://arxiv.org/abs/2505.20658v1 |
way, our dataset can be continuously enriched by incorporating human validation to train better models. References Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv prep... | https://arxiv.org/abs/2505.20658v1 |
Effective in- struction tuning with reverse instructions. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models . Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stan- ley, Richárd Nagyfi, et al. 202... | https://arxiv.org/abs/2505.20658v1 |
Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2023. Beyond the imitation game: Quantifying and extrapolating the capabili- ties of language models. Transactions on Machine Learning Research . Zhiqing Sun, Yikang Shen, Qinhong Zhou, Hongxin Zhang, Zhenfang Chen, David Cox, Yimi... | https://arxiv.org/abs/2505.20658v1 |
thespecification. The rules are as follows:1. φ₁U[a,b]φ₂ indicates that there exists a moment t' such that φ₁ is satisfied before t', and φ₂ is satisfied at t', where t' is within a time distance of a to b from the current moment.2. F[a,b]φ indicates that there exists a point within the interval [a, b] where φ is satis... | https://arxiv.org/abs/2505.20658v1 |
shows the prompts used for GPT-4 to generate STL based on the input natural language description and the top KNL-STL pairs retrieved from external knowledge with the highest similarity to the input, which serve as reference pairs in the context. 12 B Evaluation Metrics STL formula accuracy ( AF) and template accuracy (... | https://arxiv.org/abs/2505.20658v1 |
arXiv:2505.20660v1 [cs.CL] 27 May 2025BacktrackAgent: Enhancing GUI Agent with Error Detection and Backtracking Mechanism Qinzhuo Wu, Pengzhi Gao, Wei Liu, Jian Luan MiLM Plus, Xiaomi Inc {wuqinzhuo, gaopengzhi, liuwei40, luanjian}@xiaomi.com Abstract Graphical User Interface (GUI) agents have gained substantial attent... | https://arxiv.org/abs/2505.20660v1 |
P4 P5 P6 P7 P8 P9 P10 P11 I'd like to order a large cup of black tea latte, with extra Tahitian vanilla syrup, delivered to my home. Task 𝑎1 𝑎2 𝑎3 𝑎4 𝑎5 𝑎6 𝑎7 𝑎8 𝑎9 𝑎10 Action Page𝑎11𝑎31𝑎61𝑎62𝑎81 P21P41P71P72P91𝑎91 𝑎41 P51P101 Golden Action Explore Action Backtrack Equivalent Page Wrong: enter the “ord... | https://arxiv.org/abs/2505.20660v1 |
of VLMs like GPT-4o, neglecting whether the action executions align with the overall task goals. Mobile-Agent-E (Wang et al., 2025b) introduces an Action Reflector to verify action outcomes and update the Tips and Shortcuts of the task. ReachAgent (Wu et al., 2025) decomposes the task into subtasks and prioritizes the ... | https://arxiv.org/abs/2505.20660v1 |
correct, the agent considers it the final action at time-step tand proceeds to time- stept+1. Otherwise, the agent goes to the Reflector for error recovery. 4.During error recovery, the Reflector updates the action ai ttoai+1 tbased on all reflected actions at time step t, as well as the pages before and after executin... | https://arxiv.org/abs/2505.20660v1 |
𝑎𝑡𝑖 𝑎𝑡𝑖+1 Figure 4: The action result pages generated by actual execution and simulated execution. where pv t= 1indicates that the action is valid, and pv t= 0indicates that it is not. Judger With page Pt, action at, and page Pt+1, the judger assesses whether executing this action contributes to the successful co... | https://arxiv.org/abs/2505.20660v1 |
a combination of the cross- entropy loss, the verifier loss, and the judger loss: L=Lg+β1Lverifier +β2Ljudger, where β1andβ2are hyperparameters. 4 Dataset Construction 4.1 Datasets We utilize the Mobile3M (Wu et al., 2024, 2025) and Auto-UI (Zhang and Zhang, 2024) datasets. They are two largest public mobile control da... | https://arxiv.org/abs/2505.20660v1 |
Rate Both IoU Text IoU Text GPT-4o FewShot 15.16 55.38 - - - - 19.44 17.06 MobileVLM seperate FewShot 5.99 44.06 - - - - 1.75 10.60 Qwen-VL SFT 16.97 68.75 35.77 20.58 30.13 26.22 73.38 72.14 Auto-UI unified SFT 24.79 75.13 33.40 18.40 29.60 22.20 73.26 70.88 MobileVLM SFT 25.53 77.36 39.78 22.68 34.03 28.43 76.20 74.0... | https://arxiv.org/abs/2505.20660v1 |
the table, we can see that: •The backtracking mechanism improves the task success rate by 5.65% and the accuracy at both task-level and step-level by more than 3.5% and 1.5%, respectively. This is because backtracking helps the agent better align the action execution results with the task goals, enabling the agent to d... | https://arxiv.org/abs/2505.20660v1 |
The speed ratio is the ratio of the time required for a step with the entire agent to the time required with just the Generator. under different parameters, we conducted repeated experiments. For the training phase, we retrained the agent twice from the backbone model with different seeds. For the testing phase, we rep... | https://arxiv.org/abs/2505.20660v1 |
For the 8.48% of wrong actions the model successfullyAccuracyClick Scroll Input Complete IoU Text IoU Text IoU Text Percentage 79.24% 15.10% 4.84% 26.06% ReachAgent 82.09 83.12 71.25 55.07 92.80 88.80 91.82 BacktrackAgent 83.52 84.46 72.40 55.33 91.20 86.80 95.02 ∆ 1.43 1.34 1.15 0.26 -1.60 -2.00 3.20 Table 7: Statisti... | https://arxiv.org/abs/2505.20660v1 |
all training and test data once the paper is accepted. References Antonis Antoniades, Albert Örwall, Kexun Zhang, Yuxi Xie, Anirudh Goyal, and William Wang. 2024. Swe-search: Enhancing software agents with monte carlo tree search and iterative refinement. Preprint , arXiv:2410.20285. Gilles Baechler, Srinivas Sunkara, ... | https://arxiv.org/abs/2505.20660v1 |
Zheng Shou. 2024. Showui: One vision-language-action model for gui visual agent. Preprint , arXiv:2411.17465. Xiao Liu, Bo Qin, Dongzhu Liang, Guang Dong, Hanyu Lai, Hanchen Zhang, Hanlin Zhao, Iat Long Iong, Jiadai Sun, Jiaqi Wang, Junjie Gao, Junjun Shan, Kangning Liu, Shudan Zhang, Shuntian Yao, Siyi Cheng, Wentao Y... | https://arxiv.org/abs/2505.20660v1 |
for math word problems. In Findings of the Association for Computational Linguistics: EMNLP 2021 , pages 2269–2279. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems ... | https://arxiv.org/abs/2505.20660v1 |
Vajipey, Hao Cheng, Michel Galley, Jianfeng Gao, and Zhou Yu. 2025. Exact: Teaching ai agents to explore with reflective-mcts and exploratory learning. Preprint , arXiv:2410.02052. Zhuosheng Zhan and Aston Zhang. 2023. You only look at screens: Multimodal chain-of-action agents. arXiv preprint arXiv:2309.11436 . Chi Zh... | https://arxiv.org/abs/2505.20660v1 |
Button") ****Golden GUI Trajectory:**** Click(box1, "Search Box") Input(box2, "Today’s Gold Price") Click(box3, "Search Button") ————————————————— Task2: Set the display mode to night mode. ****Generate GUI Trajectory:**** Click(box1, "Personal Center") Click(box2, "Setting") Click(box3, "Display Mode") Click(box4, "Ni... | https://arxiv.org/abs/2505.20660v1 |
of reflections for each step is 3. For the SFT version, the Generator, Judge, and Reflector were trained for 2 epochs on the Mobile3M and Auto-UI datasets, respectively. For the RL version, the generator and reflector were further trained for 2 epochs with the new loss function. To ensure fair comparisons, we maintain ... | https://arxiv.org/abs/2505.20660v1 |
B Releated Work In this section, we will discuss other reflection/veri- fier/backtracking mechanisms used in LLM-agents and their similarities and differences with Back- trackAgent. Reflection. Some past works have adopted reflec- tion for self-improvement, improving generation through self-evaluation during reasoning ... | https://arxiv.org/abs/2505.20660v1 |
these operations. WebPliot (Zhang et al., 2025) uses an MCTS-based approach to explore the action space of Web tasks. It uses the maximum backpropagation (MVB) mechanism to prioritize the most promising paths for the MCTS backpropagation step. Our BacktrackAgent adopts a rule-based verifier and a model-based judger to ... | https://arxiv.org/abs/2505.20660v1 |
the execution of the last generated action based on the input of the generator. An example input of the generator. image_path: .../Starbucks0_10_5_2_3_6- screen.png —————————————————– The actions you can use are: click("IngredientButton",[953,637][1068,752]) click("BackButton",[46,150][138,242]) click("StepperReduce",[... | https://arxiv.org/abs/2505.20660v1 |
amount of data that does not require reflection is much larger than the data that needs reflection, we randomly select all negative data and 20% of positive data to construct the reflection dataset. The reflection data formed by the above two judgment examples are as follows: ****Case 1**** Input: X, a<6, ActionSpace( ... | https://arxiv.org/abs/2505.20660v1 |
need to be done. Orange arrows are the actions in the golden flow. Blue arrows are the actions in other GUI trajectories. Both the orange and blue flows can complete the task. 3. On the search page, the agent decides to click the matcha latte button in the recommendation column. P3-> click ("matcha latte") -> P1 4 The ... | https://arxiv.org/abs/2505.20660v1 |
during eval- uation (See Figure 7). We can see that the ReachAgent predicts several steps correctly but if one action is wrong, the agent would fail the task. In contrast, when BacktrackAgent mistakenly enters the "Address Collection" page and browses on the "Browsing History", it can detect the error and recover to th... | https://arxiv.org/abs/2505.20660v1 |
BacktrackAgent 29.72 22.60 23.46 27.33 58.79 15.14 78.04 71.58 75.75 82.11 82.61 74.78 Table 12: Main Result(%) on AutoUI dataset. "separate" means that this baseline is trained on five subsets of Auto-UI, while "unified" means that the baseline is trained on the entire Auto-UI dataset as a whole. - means that Auto-UI ... | https://arxiv.org/abs/2505.20660v1 |
ReachAgent SFT+RL 46.52 29.79 38.75 33.06 83.32 81.77 BacktrackAgent Original Agent 54.11 33.51 43.25 36.67 84.94 83.24 ∆ +7.59 +3.72 +4.50 +3.61 +1.62 +1.47 10 repeated tests, each with 80% of the test dataset Repetition 1 54.30 33.19 43.14 36.45 84.94 83.21 Repetition 2 54.21 33.84 43.61 37.19 84.95 83.24 Repetition ... | https://arxiv.org/abs/2505.20660v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.