text string | source string |
|---|---|
to be deputy chairperson. I am a {person, woman, man} We then compute the probabilities of each comple- tion (person, woman, man) according to the LLM llama-3.1-8B (Grattafiori et al., 2024).1 We use these probabilities to compute how femi- nine each stimulus sentence sis, as: context_fem (s) =p(woman |s) p(woman |s)+p... | https://arxiv.org/abs/2505.21378v1 |
is explicitly declared, as opposed to implicit in pronoun us- age(H3a). Similarly, usage context more generally affects application of language reforms (Silverstein, 1985; Watson et al., 2023b). We assess whether gender 4 revised ∼ original_masc + original_fem + H1a prompt_masc + prompt_fem + original_mask:prompt_fem +... | https://arxiv.org/abs/2505.21378v1 |
for each hypothesis, referring to regression results in Table 4, and descriptive statistics of the revisions in Figure 2. Hypothesis H1a: The results support the pre- dicted strategy of overall neutralization. Sig- nificant positive effects of original_masc and original_fem indicate that gendered role nouns are more of... | https://arxiv.org/abs/2505.21378v1 |
However, we also find that explicit prompts increase rates of revision to gendered variants, suggesting that the LLMs’ tendency towards neutralization may be overruled by (more) explicit information about gender, which has implications for prompt based value alignment strategies. Hypothesis H4a: Finally, LLMs are more ... | https://arxiv.org/abs/2505.21378v1 |
For H2b, we contrast masculine vs. fem- inine vs. nonbinary Gender Declaration prompts, assessing what justifications LLMs present to mo- tivate these neutralizations. We expect the theme ofinclusivity to be used more when the referent belongs to a group the reforms intend to include, i.e., women and nonbinary people. ... | https://arxiv.org/abs/2505.21378v1 |
effect of professionalism in the opposite direction being the exception to this trend). This pattern indicates that the justifications represent contrasting views on language reform, leading to inconsistencies in the values they com- municate (cf. Watson et al., 2025). We also find that different themes are em- phasize... | https://arxiv.org/abs/2505.21378v1 |
value alignment for English over other languages, for which the relationship between lin- guistic forms and values may be different. For example, in languages with grammatical gender, feminization – using feminine terms to make fem- inine referents visible – is a common strategy for feminist language reforms (Sczesny e... | https://arxiv.org/abs/2505.21378v1 |
et al., 2024; Gemma license; 9B parameters), Mistral-Nemo-Instruct- 2407 (Mistral AI Team, 2024; Apache 2.0 License; 12B parameters), and gpt-4o (Hurst et al., 2024; parameters unknown). All models were used in a way that is consistent with their terms of use. We queried gpt-4o through the OpenAI API. For the other mod... | https://arxiv.org/abs/2505.21378v1 |
ing. Proceedings of the Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) . Susan Ehrlich and Ruth King. 1992. Gender-based lan- guage reform and the social construction of meaning. Discourse & Society , 3(2):151–16... | https://arxiv.org/abs/2505.21378v1 |
through augmentation. In Proceed- ings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023) , pages 148–162, Toronto, Canada. Association for Computational Linguistics. Alonzo Martinez. 2023. An employer’s guide to inclu- sive language. Forbes Magazine . Mistral AI Team. 2024. ... | https://arxiv.org/abs/2505.21378v1 |
considered are: Neutral Feminine Masculine alderperson alderwoman alderman anchor anchorwoman anchorman assemblyperson assemblywoman assemblyman ball person ballgirl ballboy bartender bargirl barman businessperson businesswoman businessman camera operator camerawoman cameraman caveperson cavewoman caveman chairperson c... | https://arxiv.org/abs/2505.21378v1 |
tences and 93% in exactly identifying justifications. See accuracy per model in Table 7. 12 starting variant percentage gender-neutral neutral 95 feminine 96 masculine 96 Overall 95.7 Table 8: Rates of gender-neutral alternative wordings, by starting variant C Alternative wording revisions In Sections 4 and 5, we split... | https://arxiv.org/abs/2505.21378v1 |
profession (e.g., firefighter →work in firefighting orbusinessperson → career in business ). 4.Verb Phrase : The intended meaning of the original role noun is conveyed through a verb phrase describing associated actions or respon- sibilities, rather than naming the role directly (e.g., revising to replace outdoorsperso... | https://arxiv.org/abs/2505.21378v1 |
was mentioned. Based on 13,609responses, where role nouns were revised to alternative wordings. role noun variants with [MASK] tokens, to limit the influence of specific occupations on these represen- tations. For adjectives that corresponded to multi- ple wordpiece tokens, we averaged the wordpiece contextual embeddin... | https://arxiv.org/abs/2505.21378v1 |
arXiv:2505.21380v1 [cs.CL] 27 May 2025PHISH in MESH : Korean Adversarial Phonetic Substitution and Phonetic-Semantic Feature Integration Defense Byungjun Kim, Minju Kim, Hyeonchu Park, and Bugeun Kim Department of Artificial Intelligence, Chung-Ang University, Republic of Korea {k36769, minjunim, phchu0429, bgnkim}@cau... | https://arxiv.org/abs/2505.21380v1 |
to enhance the ro- bustness of detectors against phonetic perturbation by augmenting phonetic information. Specifically, our methods adopt cross-attention mechanism to incorporate semantic and phonetic information. To examine the effectiveness of both our pro- posed attack and defense methods, we conducted experiments ... | https://arxiv.org/abs/2505.21380v1 |
a sequence of syllables Ti, a perturbation ratior, and the degree of attack m. Here, the de- grees mof ‘single’ and ‘dual’ refer to single and dual-jamo attacks, respetively. PHISH consists of two main phases: (1) Index searching and (2) Substitution. In index searching, PHISH identifies the target indices to be pertur... | https://arxiv.org/abs/2505.21380v1 |
if 5:Decompose sybl into a list of jamos J. 6:Shuffle list J. 7:foreach jamo jinJdo 8: ifD[j]̸=∅then 9: Substitute jwith random jamo in D[j] 10: nsttd←nsttd+ 1 11: end if 12: ifnsttd=nattkthen 13: break 14: end if 15:end for 16:Recompose sybl with substituted jamos J 17:return sybl 14). After the substitution, the algo... | https://arxiv.org/abs/2505.21380v1 |
process of detectors, it can severely disrupt the semantic struc- ture of the text (Yu et al., 2024). This disruption can interfere with the detectors’ semantic under- standing of the text. As phonetic information can provide a hint for reconstructing the unknown word, we believe that augmenting phonetic information ca... | https://arxiv.org/abs/2505.21380v1 |
sets. UsingPHISH, we derived different test sets with differ- ent settings, including attack ratios and degrees. Specifically, we conducted attacks under three per- turbation ratios (10, 20, and 30%) and two degrees of attack (single-jamo and dual-jamo). Note that we did not alter training set; all methods are trained ... | https://arxiv.org/abs/2505.21380v1 |
±0.9 66.4±1.5 -8.7 ±1.6 RoBERTa 72.6 ±1.6 71.6±1.7 -1.0 ±2.3 66.6±3.0 -6.0 ±3.4 60.8±5.8 -11.8 ±6.0 KCBERT 77.5 ±0.4 76.7±0.6 -0.8 ±0.7 75.4±0.7 -2.1 ±0.8 72.6±1.3 -4.9 ±1.4 BERT dir-MESH 75.9±0.5 75.0±0.7 -0.9 ±0.9 73.0±0.6 -2.9 ±0.8 71.6±0.8 -4.3 ±0.9 RoBERTa dir-MESH 75.9±0.7 75.1±0.6 -0.8 ±0.9 73.6±0.4 -2.3 ±0.8 71... | https://arxiv.org/abs/2505.21380v1 |
±0.6 71.9±0.4 -7.0 ±0.6 69.6±0.8 -9.3 ±0.9 RoBERTa seq-MESH 74.6±0.6 72.8±0.6 -1.8 ±0.8 69.7±0.8 -4.9 ±1.0 67.7±0.9 -6.9 ±1.1 KCBERT seq-MESH 80.8±0.2 77.7±0.3 -3.1±0.4 73.8±0.4 -7.0±0.4 71.6±0.7 -9.2±0.7 Table 3: Detection performance on K-HATERS dataset with dual-jamo attack Attack Ratio 0% 10% 20% 30% F1 F1 ∆F1 F1 ∆... | https://arxiv.org/abs/2505.21380v1 |
to account for language-specific constraints. For example, visual substitution strategies are not applicable to the Korean language because Uni- code encoding does not support the replacement of Hangul jamo with visually-similar non-Hangul characters. So, researchers have investigated more language-specific adversarial... | https://arxiv.org/abs/2505.21380v1 |
its adversarial potential. Furthermore, detectors equipped with seq-MESH or dir-MESH consistently outperformed their base models across both perturbed and original test sets, suggesting that our defense methods not only improve robust- ness but also can be generalized to real-world data where phonetic substitutions may... | https://arxiv.org/abs/2505.21380v1 |
Victor HP Van Daal, Nicoletta Polyzoe, MARIA-LOUISA TSIPA, and Michalis Petalas. 2004. The effects of orthographic depth on learning to read alphabetic, syllabic, and logographic scripts. Reading research quarterly , 39(4):438–468. Younghoon Jeong, Juhyun Oh, Jongwon Lee, Jaimeen Ahn, Jihyung Moon, Sungjoon Park, and A... | https://arxiv.org/abs/2505.21380v1 |
casino? transposed-letter similarity effects with nonadjacent letter positions. Journal of memory and language , 51(2):231–246. Edwin Puertas and Juan Carlos Martinez-Santos. 2021. Phonetic detection for hate speech spreaders on twit- ter. Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. 2020. Leveraging pre-trained... | https://arxiv.org/abs/2505.21380v1 |
due to PHISH’s phonetic perturbations. By providing phonetic information to the detectors, we could mitigate this loss. Second, the statistics may explain why KCBERT consistently outperforms the other two detectors. As KCBERT showed fewer unknown tokens, it is highly likely that the model suffers less from semantic los... | https://arxiv.org/abs/2505.21380v1 |
AutoJudger : An Agent-Driven Framework for Efficient Benchmarking of MLLMs Xuanwen Ding1,3∗, Chengjun Pan1∗, Zejun Li1∗, Jiwen Zhang1∗, Siyuan Wang2, Zhongyu Wei1,3† 1Fudan University, Shanghai, China 2University of Southern California, Los Angeles, USA 3Shanghai Innovation Institute, Shanghai, China dxwpika@gmail.com ... | https://arxiv.org/abs/2505.21389v1 |
evaluation results, respectively. (b) compares several efficient benchmarking methods on MMT-Bench. AutoJudger achieves 92% rank consistency using only 4% of the data (125 samples). for text-only benchmarks, performing stratified sampling based on question categories [ 71] and difficulty levels [ 99,72] to construct su... | https://arxiv.org/abs/2505.21389v1 |
diversity, we equip Au- toJudger with a semantic-aware retrieval mechanism supported by Item Response Theory (IRT), ensuring that selected questions are representative. To further enhance the adaptivity, we incor- porate a dynamic memory that maintains contextual statistics of previously evaluated questions, enabling c... | https://arxiv.org/abs/2505.21389v1 |
Other lines of work approach difficulty estimation from content-based features [ 36,89], step-level reasoning complexity [ 14,86], or LLM-based prediction models [ 26,42]. In our work, we incorporate IRT into the adaptive benchmark construction process, where question difficulty is pre-estimated from historical model r... | https://arxiv.org/abs/2505.21389v1 |
select appropriate samples, AutoJudger should acquire a clear understanding of the difficulty of questions during evaluation. To this end, we first leverage Item Response Theory (IRT) [ 9] to characterize the difficulty diof each question qibefore the evaluation (§ 4.1). We then employ an intelligent agent powered by a... | https://arxiv.org/abs/2505.21389v1 |
directly select questions from the entire question pool Q. To address this challenge, we design a retrieval strategy to provide the agent with a candidate set C∗ kwith feasible size |C∗ k| ≪ |Q|. We aim to select questions that are both appropriate in difficulty and semantically distinct from those previously attempted... | https://arxiv.org/abs/2505.21389v1 |
model responses, grouped by semantically inferred cate- gories as a markdown table. Since many benchmarks lack predefined class labels or contain noisy annotations, AutoJudger assigns categories based on semantic features and dynamically expands the category table as new topics emerge. For each category, the memory tra... | https://arxiv.org/abs/2505.21389v1 |
a shared question pool for all models, including Random Sampling (Random) , which selects a fixed number of questions at random; Stratified Random Sampling (Stratified) [ 71] also selects a fixed number of questions randomly, but applies weighted sampling based on the number of categories within each benchmark; andClus... | https://arxiv.org/abs/2505.21389v1 |
advantages. We believe the reason why AutoJudger shows less improvement on SEEDBench is that the dataset itself is relatively large—about four times the size of the others—so even 5% already includes a substantial amount of data for baselines to converge. Therefore, we further explore smaller compression rates and find... | https://arxiv.org/abs/2505.21389v1 |
in Table 2, removing visual information consistently harms the performance, indicating that the semantics of images are crucial for ensuring the diversity of selected questions in multimodal benchmarks. Necessity of Dynamic Memory M.To understand the contribution of the proposed dynamic memory Mk, we analyze the impact... | https://arxiv.org/abs/2505.21389v1 |
ability to adaptively select semantically diverse and ability-matched questions may lead to more fair and comprehensive assessments of model capabilities. Limitations AutoJudger reduces evaluation cost by focusing on informative samples based on estimated question difficulty, avoiding the need to exhaustively test all ... | https://arxiv.org/abs/2505.21389v1 |
procedure enables efficient and stable estimation of real-time model ability during evaluation, while keeping the question difficulties {di}fixed. 11 C Implementation Details of Efficient Benchmarking Methods C.1 Baselines We detail the implementation of the baselines below. •Random Sampling (Random) : We uniformly sam... | https://arxiv.org/abs/2505.21389v1 |
set models used for IRT-based question difficulty assessment. These models span a range of sizes and include both open-source and proprietary models. Models Open-source # Params (B) Date InternVL2-1B [13] Yes 0.9 2024.11 llava-onevision-qwen2-0.5B-ov [43] Yes 0.9 2024.07 llava-onevision-qwen2-0.5B-si [43] Yes 0.9 2024.... | https://arxiv.org/abs/2505.21389v1 |
61.81 58.00 7.75 multi-mean 72.61 70.57 55.07 55.92 10.00 image 69.80 64.72 66.07 65.45 7.50 text 81.31 73.02 60.88 75.59 3.75 CLIP ViT-L/14multi-concat 67.23 75.53 52.00 61.61 9.50 multi-mean 75.70 75.79 67.00 54.88 5.25 image 73.24 74.52 62.30 64.01 6.00 text 77.45 66.38 66.15 62.62 6.25 Qwen2.5-VL-7Btext&image 78.12... | https://arxiv.org/abs/2505.21389v1 |
provide limited information. To validate this argument, we conduct an experiment to assess top-performing models (top 50% in terms of average ranks), either with their personalized questions picked by AutoJudger or simple questions that are selected to evaluate the worst model. Results are provided in Table 7. Consider... | https://arxiv.org/abs/2505.21389v1 |
questions. Benchmark # CandidateCompression Ratio 1% 2% 3% 4% 5% AI2D TEST5 85.29 89.71 93.38 92.65 94.85 7 88.24 92.65 93.38 93.38 94.85 10 83.82 86.03 89.71 91.18 91.91 MMMU DEV VAL5 77.21 82.54 84.74 85.85 87.94 7 77.21 82.35 80.88 83.09 83.82 10 77.21 82.35 83.82 85.29 85.29 MMT-Bench5 85.66 87.87 89.34 91.91 92.06... | https://arxiv.org/abs/2505.21389v1 |
JSON object with the following format: { "<Question_ID_1>": "<Category_Name_1>", "<Question_ID_2>": "<Category_Name_2>", ... } - Keys are question IDs (index) from the input data. - Values are descriptive category names that you assign. - ONLY return the JSON object; do not include any other text or explanation. 17 Pro... | https://arxiv.org/abs/2505.21389v1 |
and performance metrics. This selection strikes a balance between providing an appropriate challenge and ensuring the student encounters a diverse range of categories based on historical performance, which appears to be lacking in the provided categories.","QuestionID": 9086 AutoJudgerCountMax Diff.…Acc.Statistics155.3... | https://arxiv.org/abs/2505.21389v1 |
chairsD.Four chairsB.One chairC.Three chairsOptions:A.On a flat surfaceD.On a steep slopeB.On a curvy pathC.On a downhill slopeOptions:A.11D.13B.10C.12Figure 9: Response Examples from AutoJudger on SeedBench. 20 References [1]Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit ... | https://arxiv.org/abs/2505.21389v1 |
generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 24185–24198, 2024. [14] Yi Cheng, Siyao Li, Bang Liu, Ruihui Zhao, Sujian Li, Chenghua Lin, and Yefeng Zheng. Guiding the growth: Difficulty-controllable question generation through step-by-step... | https://arxiv.org/abs/2505.21389v1 |
but not perceive. In European Conference on Computer Vision , pages 148–166. Springer, 2024. [25] Shaikat Galib, Shanshan Wang, Guanshuo Xu, Pascal Pfeiffer, Ryan Chesler, Mark Landry, and Sri Satish Ambati. H2ovl-mississippi vision language models technical report. arXiv preprint arXiv:2410.13611 , 2024. [26] Yifan Ga... | https://arxiv.org/abs/2505.21389v1 |
Yun, Kyoungsoo Park, YoungHoon Jung, Damji Stratton, and Hyeoncheol Kim. Difficulty-focused contrastive learning for knowledge tracing with a large language model-based difficulty prediction. arXiv preprint arXiv:2312.11890 , 2023. [43] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Pe... | https://arxiv.org/abs/2505.21389v1 |
, 2024. [59] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek- vl: Towards real-world vision-language understanding, 2024. [60] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong... | https://arxiv.org/abs/2505.21389v1 |
Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pages 8748–8763. PmLR, 2021. [75] Georg Rasch. Probabilistic models for some intelligence and attainment tests. ERIC, 1993. [76] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashre... | https://arxiv.org/abs/2505.21389v1 |
In Findings of the Association for Computational Linguistics: NAACL 2025 , pages 6396–6418, 2025. [92] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 , 2024. [93] Yuan Yao, Tianyu... | https://arxiv.org/abs/2505.21389v1 |
arXiv:2505.21397v1 [cs.CL] 27 May 2025DecisionFlow: Advancing Large Language Model as Principled Decision Maker Xiusi Chen1*, Shanyong Wang1∗, Cheng Qian1∗, Hongru Wang1∗, Peixuan Han1, Heng Ji1 1University of Illinois Urbana-Champaign {xiusic, hengji}@illinois.edu Abstract In high-stakes domains such as healthcare and... | https://arxiv.org/abs/2505.21397v1 |
em- bed this deliberative process into models, ensuring that decisions and explanations emerge from the same principled, traceable reasoning framework. To address these limitations, we introduce Deci- sionFlow , a step-by-step decision modeling frame- work that transforms natural language scenar- ios into structured, u... | https://arxiv.org/abs/2505.21397v1 |
bridging the gap between statistical learning and principled decision science. We anticipate that this line of work will catalyze further advances in build- ing AI systems capable of transparent, accountable, and high-stakes decision support in critical real- world applications. 2 Related Work LLMs for Decision Making.... | https://arxiv.org/abs/2505.21397v1 |
max 1≤i≤nP(ai| S,C), (3)Step Input Output S1 Scenario S, Actions A Attributes P={p1, . . . , p m} Relevance matrix R∈Rn×m S2 (A, P, R ), Constraints CWeights w∈Rn×m Filtered matrix R′=w′◦R S3R′, Constraints C Objective function O(A) S4O(A), Feasible A Final decision a⋆ Explanation rational Table 1: Data flow in the Dec... | https://arxiv.org/abs/2505.21397v1 |
social iden- tity or age. In this case, prioritizing fairness is the constraint and injury severity is an attribute More- over, humans usually have auxiliary constraints that prevent them from picking certain options. As a result, certain actions should be ruled out to ac- count for the constraints. To explicitly refle... | https://arxiv.org/abs/2505.21397v1 |
cal decision-makers and how their decisions can be aligned to different DMAs. To address limita- tions of the original 62-instance dataset, including its small scale and frequent alignment errors, we revise the original examples and expand it into a 200-instance human-verified examples, supple- mented with formalized d... | https://arxiv.org/abs/2505.21397v1 |
69.25 48.33 27.50 37.92 DeLLMa-Pairs - - - 27.50 34.17 30.84 DeLLMa-Top1 - - - 44.17 26.67 35.42 DecisionFlow 90.50±2.29 68.00 ±2.13 79.25 ↑10.00±1.18 76.67±3.94 64.72 ±1.27 70.70 ↑30.28±2.40 Table 2: The performance of different methods in three datasets (MTA, DeLLMa Agriculture, DeLLMa Stocks). the results are compar... | https://arxiv.org/abs/2505.21397v1 |
in detail in Section 6.2. We note that our method can alleviate this issue by achieving more balanced performance between the high and low-settings. DeLLMa. Most of baselines such as Zero-shot and CoT performs better on Agriculture against Stock dataset. We attribute this to the exclusively numerical nature of the Stoc... | https://arxiv.org/abs/2505.21397v1 |
alignment to a particular DMA (Hu et al., 2024a). Table 4 shows the results. There are several observations: 1) It is clear that different models have the inherent bias for decision-making and this problem does not alleviate when model size increases; 2) CoT 7 Dimensions Task Description Response Case: Scenario: … Pati... | https://arxiv.org/abs/2505.21397v1 |
modeling explanations. 6.4 Human Evaluation We first randomly sample 80 cases across three domains: Medical (40 samples, including 20 high- DMA and 20 low-DMA cases), Agriculture (20 samples), and Finance (20 samples). We then ask three well-educated annotators to indicate whethereach step accurately capture the inform... | https://arxiv.org/abs/2505.21397v1 |
enhance the credibility of harmful or deceptive systems. Such misuse in areas like finance, healthcare, or political discourse poses significant risks. We caution against these applications and em- phasize the importance of responsible use with appropriate human oversight, particularly in high- stakes settings.Acknowle... | https://arxiv.org/abs/2505.21397v1 |
Dani Yogatama, and Willie Neiswanger. 2024. Dellma: Decision making under uncertainty with large language models. Preprint , arXiv:2402.02392. Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. 2024. A language agent for autonomous driving. Preprint , arXiv:2311.10813. Allen Nie, Ching-An Cheng, Andrey Kolo... | https://arxiv.org/abs/2505.21397v1 |
medical triage scenario, when presented with a multiple-choice question related to medical conditions or symp- toms, the decision-making process prioritizes a high protocol focus. This approach entails strict adherence to established rules, protocols, or guide- lines, regardless of specific contextual factors that migh... | https://arxiv.org/abs/2505.21397v1 |
This means it may prefer a response that offers a superior ex- pected outcome, even if the total expected value is lower or if it commits to a singular course of action. High Continuing Care In a military medical triage setting, when presented with multiple-choice questions about medical scenarios, the decision- making... | https://arxiv.org/abs/2505.21397v1 |
low utilitarianism approach. This means it may allocate limited resources based on personal considerations toward patients or other values, such as kindness, fairness, respect, or loy- alty, rather than solely aiming to save the most people or maximize overall benefit, even if some individuals are detrimentally affecte... | https://arxiv.org/abs/2505.21397v1 |
the Scoring and Filter modules. The full model achieves the best performance, with 79.25% average accuracy on Medical and 70.70% on DeLLMa. Removing the Filter causes a sharp drop—Medical-Low drops from 68.00% to 30.50%, and DeLLMa average drops to 61.95%. Without Scoring, Medical aver- age further declines to 53.00%. ... | https://arxiv.org/abs/2505.21397v1 |
certain domains, though overly aggressive filtering (e.g., "top 1") can harm overall performance by discarding useful informa- tion, especially in the Agriculture setting. Overall, both experiments highlight the impor- tance of careful hyperparameter tuning. Moderate values for ϵtend to provide a good trade-off be- twe... | https://arxiv.org/abs/2505.21397v1 |
with the utilitarian principle. Thus, the optimal choice is to treat the bomber (Patient 2). Table 10: Illustration of decision-making under high utilitarian principles in a triage scenario. 15 EXAMPLE ZERO-SHOT PROMPT FOR MEDICAL TRIAGE ALIGNMENT You are an algorithm assisting with medical triage in a military setting... | https://arxiv.org/abs/2505.21397v1 |
A mapping that associates a variable with its relevant characteristics, states, or functions. - Target Bias: A specific evaluative perspective (e.g., "high moral deservingness") that should guide what types of attributes are most important to extract and emphasize. # Instructions: Input: - variable: Either two patients... | https://arxiv.org/abs/2505.21397v1 |
ethical stance leads to the preferential allocation of resources toward patients who have demonstrated moral merit, such as those injured while assisting others or those blameless for their injuries. Conversely, resources may be conserved from individuals injured during nefarious activities or due to personal neglect."... | https://arxiv.org/abs/2505.21397v1 |
Product Summary: apple production is forecasted to be up 3 percent from 2020/21 but down 5 percent from 2019/20. washington state’s crop is expected to be larger, but there is concern over heat damage. export markets may remain sluggish due to high tariffs and shipping challenges, potentially pushing more apples into t... | https://arxiv.org/abs/2505.21397v1 |
new york and michigan, which account for a significant portion of processed apples. - California Price and Yield Statistics: the average apple yield is 19,000 LB / ACRE and the average price per unit is 0.244 $ / LB. - Avocado: - Product Summary: california avocado production has decreased, with wildfires and water res... | https://arxiv.org/abs/2505.21397v1 |
2023-06: 24.38, 2023-07: 23.04, 2023-08: 19.12, 2023-09: 17.66, 2023-10: 14.33, 2023-11: 13.15. I’m a trader planning my next move. I would like to maximize my profit with 10000 dollars. Below are the actions I can take: Action 1. AMD: 10000 dollars Action 2. GME: 10000 dollars I would like to know which action I shoul... | https://arxiv.org/abs/2505.21397v1 |
arXiv:2505.21411v2 [cs.CL] 28 May 2025 Huawei Proprietary -Restricted Distribution 2 Huawei Cloud Pangu Models Icon Huawei Cloud Pangu Industry Models IconTECHNICAL REPORT PANGU PROMOE: M IXTURE OF GROUPED EXPERTS FOR EFFICIENT SPARSITY Pangu Team, Huawei pangutech@huawei.com ABSTRACT The surgence of Mixture of Experts... | https://arxiv.org/abs/2505.21411v2 |
experts for a token, we divide the experts into equal groups and then choose experts from each of the groups. Each group has an identical number of activated experts. In typical distributed deployments, the experts are assigned to the devices according to the group ID. MoGE effectively balances the computational load a... | https://arxiv.org/abs/2505.21411v2 |
its associated routing mechanisms, and finally discuss the auxiliary losses designed to optimize its performance. 2 (a) The distribution of activated experts for a single token. 0.00 0.25 0.50 0.75 1.00 1.25 1.50 Imbalance Score0.00.20.40.60.81.0ProbabilityMoE MoGE (b) Distribution of imbalance score. Figure 2: Compari... | https://arxiv.org/abs/2505.21411v2 |
e.g., a set of tokens) X. For this batch, let Ti(X)be the total number of expert computations (or tokens routed to experts) handled by device i. The Imbalance Score is then defined as the difference between 3 the maximum and minimum load across all Mdevices, normalized by the batch size |X|: IS(X) =1 |X| max i∈{1,···,... | https://arxiv.org/abs/2505.21411v2 |
for all Nexperts using the input token hand a router weight matrix W∈Rd×n. Instead of directly applying a global Top-K, we first apply a global softmax to these raw scores to obtain normalized probabilities or affinities, S, for all experts: S=Softmax W⊤h (5) Here, Sis a vector of length N, where Siis the initial sco... | https://arxiv.org/abs/2505.21411v2 |
domain expertise is used to shortlist potential models, narrowing the design space; third, the performance of candidate models is evaluated using an operator-level simulator that correlates system hardware parameters, such as TFLOPS, memory access bandwidth, memory capacity, and interconnection topology, and automatica... | https://arxiv.org/abs/2505.21411v2 |
in the first phase, then to improve reasoning skills of the model in the second phase, and to further refine model knowledge and behavior in the third phase. Apart from the general data from various sources, we particularly involve a lot of high-quality data from multiple industrial domains in the first general phase. ... | https://arxiv.org/abs/2505.21411v2 |
cosine learning rate schedule is employed throughout, encompassing three progressive phases. In the general phase, the learning rate decays from 3×10−4to3×10−5with a batch size of 4 million tokens. This is followed by the reasoning phase, during which the learning rate further decreases from 3×10−5to1×10−5, and the bat... | https://arxiv.org/abs/2505.21411v2 |
toward stepwise problem solving. Second, the staged finetuning allows the model to consolidate general linguistic capabilities before being pushed toward more cognitively demanding tasks. Together, this formulation enables the model to balance fluency with depth, yielding improved performance on benchmarks that require... | https://arxiv.org/abs/2505.21411v2 |
Rewards: For tasks with verifiable ground-truth, such as mathematics or coding, correctness-based rewards are assigned. Mathematical problems are assessed by a hybrid system combining rule-based verifiers for standard formats and LLM-based verifiers for more nuanced interpretations. Code responses undergo a multi-stage... | https://arxiv.org/abs/2505.21411v2 |
2 to minimize EP communication volume when memory capacity allows. The model contains 48 transformer layers, and to achieve better load balancing across pipeline stages, 2 additional no-op layers are appended, increasing the total number of layers to 50. These layers are then evenly partitioned across 5 pipeline stages... | https://arxiv.org/abs/2505.21411v2 |
a hybrid DP2+TP4 parallelism strategy is used to reduce cross-CPU communication overhead on the 300I Duo NPU, where four chips are controlled by one CPU. Requests are grouped along the batch dimension to balance computation between CPU domains. For expert modules, a combination of Tensor Parallelism (TP) and Expert Par... | https://arxiv.org/abs/2505.21411v2 |
quantization-induced logit perturbations. Even minor deviations in gate scores can disrupt the Top-K expert assignment logic, degrading model performance due to misrouted tokens. Third, expert activation sparsity creates calibration bottlenecks: rarely activated experts receive insufficient data coverage during paramet... | https://arxiv.org/abs/2505.21411v2 |
pipeline with a pingpong scheduler is introduced for KV processing to enhance MTE2 utilization, as indicated by steps (2)(3)(4) and (6)(7)(8). To address this challenge, we propose MulAttention , a fused attention operator optimized for Ascend hardware, specifically designed for the decoding stage in LLM inference. Spe... | https://arxiv.org/abs/2505.21411v2 |
Prefill stage, only the Top-8 experts are activated per token for the MoE architecture, which effectively reduces the model size to an equivalent 16B dense model. This sparse activation mechanism significantly reduces computational cost and communication overhead. Moreover, the adoption of a minimal-card deployment str... | https://arxiv.org/abs/2505.21411v2 |
achievements are attained through a computationally efficient MoE design. When benchmarked against contemporary base models including Qwen3-32B-base [ 43], GLM4-32B-base [ 11], Gemma3-27B-base [ 35], and Llama-4-Scout- baset [1], Pangu Pro MoE demonstrates consistent performance advantages. 5.2 Instruct Model Benefitin... | https://arxiv.org/abs/2505.21411v2 |
dataset, we employed a referee model to conduct scoring, ensuring a fair and objective assessment. For the remaining datasets, we adopted matching and exact matching techniques to evaluate the model’s performance. For LiveCodeBench, we use versions 24.8.1-25.1.1, which cover the data and problem sets between this time ... | https://arxiv.org/abs/2505.21411v2 |
models like GLM-Z1-32B (52.6). Remarkably, Pangu Pro MoE matches 32B-scale state-of-the-art models’ reasoning capabilities using only 16B activated parameters. This efficiency stems from the innovative MoGE architecture, which enhances inference speed while maintaining reasoning accuracy. 5.3 Inference Efficiency Perfo... | https://arxiv.org/abs/2505.21411v2 |
accelerator enables efficient and cost-effective inference for billion-scale MoE models. Pangu Pro MoE is quantized under the W8A8 configuration during inference. In the Prefill stage, by employing two Ascend 300I Duo accelerators with a batch size of 2, 72BA16B MoE achieves 1.94s latency for 2k-length input sequences,... | https://arxiv.org/abs/2505.21411v2 |
in Layer 23, which in turn surpass those in Layer 0. This progressive trend suggests that expert specialization intensifies with network depth. Furthermore, for tasks that primarily assess general language understanding—such as C-Eval, and MMLU—the distribution of expert activations tends to be more balanced across the... | https://arxiv.org/abs/2505.21411v2 |
between experts from different groups remain consistently low across layers, suggesting that inter-group interactions are minimal. This observation supports the hypothesis that our model achieves a low degree of expert redundancy and encourages specialization, where different experts are responsible for distinct aspect... | https://arxiv.org/abs/2505.21411v2 |
suggesting that deeper layers may adaptively modulate expert usage to capture more task-specific or abstract representations. Global Expert Distribution The balance of expert load in MoE architectures remains a critical topic, as more uniform expert activation is generally associated with improved resource efficiency a... | https://arxiv.org/abs/2505.21411v2 |
2021. [6]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 , 2021. [7]Yiming Cui, Ting Liu, Wanxiang Che, Li Xiao, Zhipeng Chen, Wen... | https://arxiv.org/abs/2505.21411v2 |
al. Mixtral of experts. arXiv preprint arXiv:2401.04088 , 2024. [17] Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard H. Hovy. Race: Large-scale reading comprehension dataset from examinations. ArXiv , abs/1704.04683, 2017. [18] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping ... | https://arxiv.org/abs/2505.21411v2 |
Annual Meeting of the Association for Computational Linguistics , 2022. 23 [34] Yehui Tang, Yichun Yin, Yaoyuan Wang, Hang Zhou, Yu Pan, Wei Guo, Ziyang Zhang, Miao Rang, Fangcheng Liu, Naifu Zhang, Binghan Li, Yonghan Dong, Xiaojun Meng, Yasheng Wang, Dong Li, Yin Li, Dandan Tu, Can Chen, Youliang Yan, Fisher Yu, Ruim... | https://arxiv.org/abs/2505.21411v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.