text string | source string |
|---|---|
Zhang, June M. Liu, Jinfeng Zhou, Alvionna S. Sunaryo, Juanzi Li, Tatia M.C. Lee, Rada Mihalcea, and Minlie Huang. 2024. Emobench: Evaluating the emotional intelli- gence of large language models. In Annual Meeting of the Association for Computational Linguistics . Pratik Sachdeva, Renata Barreto, Geoff Bacon, Alexan- ... | https://arxiv.org/abs/2505.22327v1 |
Gashaw Kidanu, Zewdie Mossie, Hussien Seid, Eshete Derb, and Seid Muhie Yimam. 2024. Evaluating gender bias in machine translation for low-resource lan- guages. In 5th Workshop on African Natural Lan- guage Processing . Shaden Shaar, Firoj Alam, Giovanni Da San Martino, Alex Nikolov, Wajdi Zaghouani, Preslav Nakov, and... | https://arxiv.org/abs/2505.22327v1 |
Fakenewsnet: A data repository with news content, social context and dy- namic information for studying fake news on social media. CoRR , abs/1809.01286. Vered Shwartz. 2022. Good night at 4 pm?! time ex- pressions in different cultures. In Findings of the As- sociation for Computational Linguistics: ACL 2022 , pages 2... | https://arxiv.org/abs/2505.22327v1 |
Allahsera Auguste Tapo, Nouhoum Coulibaly, Seydou Diallo, Sebastien Diarra, Christopher M Homan, Mamadou K. Keita, and Michael Leventhal. 2025. GAIfE: Using GenAI to improve literacy in low- resourced settings. In Findings of the Association for Computational Linguistics: NAACL 2025 , pages 7914–7929, Albuquerque, New ... | https://arxiv.org/abs/2505.22327v1 |
Nations Children’s Fund, New York, NY , USA. Accessed: 2025-05-16. Aziza Usmanova, Ahmed Aziz, Dilshodjon Rakhmonov, and Walid Osamy. 2022. Utilities of artificial intelli- gence in poverty prediction: a review. Sustainability , 14(21):14238. Saeid Ashraf Vaghefi, Dominik Stammbach, Veruska Muccione, Julia Bingler, Jin... | https://arxiv.org/abs/2505.22327v1 |
data. In Proceedings of the 2024 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies (Volume 3: System Demonstra- tions) , pages 61–69, Mexico City, Mexico. Associa- tion for Computational Linguistics. Rose Wang, Qingyang Zhang, Carly Robinson, Su... | https://arxiv.org/abs/2505.22327v1 |
Hudi, Patrick Amadeus Irawan, David Anugraha, Rifki Afina Putri, Yutong Wang, Adam Nohejl, Ubaidillah Ariq Prathama, Nedjma Ousidhoum, Afifa Amriani, Anar Rza- yev, Anirban Das, Ashmari Pramodya, Aulia Adila, Bryan Wilie, Candy Olivia Mawalim, Ching Lam Cheng, Daud Abolade, Emmanuele Chersoni, and 32 others. 2025. Worl... | https://arxiv.org/abs/2505.22327v1 |
Biemann, and Animesh Mukher- jee. 2024. Demarked: A strategy for enhanced abusive speech moderation through counterspeech, detoxification, and message management. Preprint , arXiv:2406.19543. Kayo Yin, Amir Moryossef, Jami Hochgesang, Yoav Goldberg, and Malihe Alikhani. 2021. Including signed languages in natural langu... | https://arxiv.org/abs/2505.22327v1 |
Compu- tational Linguistics. Haiqi Zhou, David Hobson, Derek Ruths, and An- drew Piper. 2024. Large scale narrative messaging around climate change: A cross-cultural compari- son. In Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (Cli- mateNLP 2024) , pages 143–155, Bangkok, Thailan... | https://arxiv.org/abs/2505.22327v1 |
quality (Min et al., 2023; Chaszczewicz et al., 2024; Althoff et al., 2016); (3)Tracking emotion and mood via time-series data analysis ( ˇCosi ´c et al., 2024) and detecting mental health crises over time (Gong et al., 2019; Yuan et al., 2023). Conversely, when NLP tools serve as clients , they typically simulate clie... | https://arxiv.org/abs/2505.22327v1 |
in clinical reports to generate and summarize documentation such as patient notes, discharge summaries, and case reports, offering improvements in efficiency, organization, and standardization of medical writing (Park et al., 2024; Ali et al., 2023; Patel and Lam, 2023; Cascella et al., 2023). They can help identify gr... | https://arxiv.org/abs/2505.22327v1 |
Hossein Abad et al., 2022) Various online recipe databases, both struc- tured and unstructured: recipe websites, historical recipe archives, nutritional databases, sustainability dataNER, Information extraction, se- mantic linking, recommender systemQualitative analysis (van Erp et al., 2021) Food Label Information and... | https://arxiv.org/abs/2505.22327v1 |
Modeling. To emulate expert tutoring behaviors, some works model the decision- making processes of experienced educators. Bridging the Novice-Expert Gap utilizes cognitive task analysis to capture experts’ identification of student errors, remediation strategies, and instructional intentions, informing LLM responses in... | https://arxiv.org/abs/2505.22327v1 |
papers at the intersection of peace building and NLP in Table 3. To identify the most relevant literature on human rights violation detection using NLP and on conflict prediction, we employed three search strategies: querying the ACL Anthology, conducting searches on Google Scholar, and utilizing the Consensus research... | https://arxiv.org/abs/2505.22327v1 |
Confusion Matrices (Borger et al., 2022) Electronic Health Record Document Classification (type of violence and patient status)Precision, Recall, F1 (Botelle et al., 2022) Table 3: Overview of Peace Building and Physical Safety related NLP studies Scholar to “nlp income", “nlp poverty", and “text poverty classification... | https://arxiv.org/abs/2505.22327v1 |
can potentially reduce crowd-sourcing efforts (Bhat and Varma, 2023; Kim et al., 2024). Further, studies on curating multimodal datasets (Kiela et al., 2021) and understanding the strengths and limitations of multimodal LLMs have also garnered attention (Rizwan et al., 2025). We have highlighted significantly important... | https://arxiv.org/abs/2505.22327v1 |
coreference resolution, occupation classification, translation, multi-agent interactions). The selection emphasizes both foundational work and recent advancements that shaped current methodologies. Early studies such as those by (Bolukbasi et al., 2016) and (Caliskan et al., 2017) were included for their role in establ... | https://arxiv.org/abs/2505.22327v1 |
labels and targetsHate speech score, difficulty of survey item and response, severity of rater(Sachdeva et al., 2022) CONAN and its variants Generation of counter speech against hate speech through different NLP generation strategiesSemantic Similarity, Novelty, Diversity, Toxicity, Politeness, Intent Accuracy and Hate... | https://arxiv.org/abs/2505.22327v1 |
explanation genera- tionprecision, recall, F1 macro, accuracy; ROUGE and coherence(Kotonya and Toni, 2020b) COVID-Fact evidence retrieval, claim ver- ification: classificationCOVID-FEVER Score (similar to FEVER score)(Saakyan et al., 2021) X-Fact claim verification: classifi- cationF1 score (Gupta and Srikumar, 2021) F... | https://arxiv.org/abs/2505.22327v1 |
Dependencies tree- banks; SIGMORPHON; WMT news translation; XNLI cross- lingual NLI; TyDi QA/ SQuADParsing; Inflection; MT; TTS; NLI; QAScaled performance utility; Global util- ity metrics(Blasi et al., 2021) ImageNet Image Classification by CountryAccuracy gaps (US/EU vs developing regions)(Shankar et al., 2017) WikiA... | https://arxiv.org/abs/2505.22327v1 |
Table 9: Representative datasets and evaluation practices across accessibility-related NLP tasks. Datasets NLP Task(s) Evaluation Metrics Reference Global Stocktake Dataset from Climate Policy RadarTopic modelling Cosine similarity (Sietsma et al., 2023) SumIPCC: topic-annotated sum- maries and relative paragraphs from... | https://arxiv.org/abs/2505.22327v1 |
B Global Goals We include an overview of the Sustainable Development Goals (SDGs) in Figure 2, which apply to many of the NLP applications discussed in this paper. Figure 2: Overview of the SDG goals. Source: https://sdgs.un.org/goals C Global Risks We present the key global risks categorized by domain, as outlined in ... | https://arxiv.org/abs/2505.22327v1 |
systemically important supply chain Disruptions to critical infrastructure Economic downturn (recession, stagnation) Inflation Talent and/or labour shortagesECONOMIC Prices of key assets become disconnected from the real economy and collapse. Control over critical resources or technologies by a few actors that manipula... | https://arxiv.org/abs/2505.22327v1 |
we developed the author guidelines shown in Figure 5. We share these here for transparency reasons and to assist other researchers undertaking similar multidisciplinary efforts. Project Guidelines Your task is to write a section (or subsection) based on a topic of your expertise. To ensure consistency across all sectio... | https://arxiv.org/abs/2505.22327v1 |
arXiv:2505.22354v1 [cs.CL] 28 May 2025LLMs Struggle to Reject False Presuppositions when Misinformation Stakes are High Judith Sieker*(j.sieker@uni-bielefeld.de) Computational Linguistics, Department of Linguistics, Bielefeld University, Germany Clara Lachenmaier*(clara.lachenmaier@uni-bielefeld.de) Computational Lingu... | https://arxiv.org/abs/2505.22354v1 |
by research showing that presuppositions often prove to be more persuasive than direct assertions (Thoma et al., 2023; Moldovan, 2023). By presenting content as shared knowledge, presuppositions distract from critical eval- uation, making them highly effective at embedding disputable or misleading information (Lombardi... | https://arxiv.org/abs/2505.22354v1 |
simple questions in QA systems, revealing persis- tent challenges in handling false presuppositions (Kim et al., 2021, 2023; Daswani et al., 2024; Yu et al., 2023; Srikanth et al., 2024). Existing research on LLMs in political con- texts has focused on how LLMs reflect political biases, rather than their ability to rec... | https://arxiv.org/abs/2505.22354v1 |
know, we’ve completed the wall” and “Do you know, we’ve completed the wall?”, the presupposi- tion (the wall has been completed) remains intact. Projection from such embeddings serves as a well-established diagnostic tool for identifying presuppositions (Chierchia & Mcconnell- Ginet, 1990; Bender & Lascarides, 2019). B... | https://arxiv.org/abs/2505.22354v1 |
probability: We designed ’probable’ and ’im- probable’ scanarios for each trigger. ’High probability’ scenarios included typical conference activities (e.g., giv- ing speeches) and ’low probability’ ones involved unlikely activities (e.g., karaoke). These scenarios were minimal pairs with slight wording differences to ... | https://arxiv.org/abs/2505.22354v1 |
often include irrelevant or nonsensical details, further contributing to misinformation. To evaluate the reliability of the annotations, we calculated Fleiss’ κ(0.82) and the average pairwise Cohen’s κ(0.72). The results indicate substantial agreement, underscoring the robustness and consistency of the annotation proce... | https://arxiv.org/abs/2505.22354v1 |
showed a weaker but still significant effect ( χ2(6) = 13.31, p<0.05). These results suggest that the models have particular biases toward certain responses depending on the type of trigger. For example: LLama accommodates misin- formation more often with interaction particles, while GPT does so with factive verbs (Tab... | https://arxiv.org/abs/2505.22354v1 |
as the only condition with a strong main effect across all annotation labels for both models (Ta- ble 5). It significantly influenced Rejection, Imprecise, and Accommodation for both LLaMa ( χ2 R(1) =20.25,p<0.001; χ2 I(1) =32.93,p<0.001; χ2 A(1) =16.29,p<0.001) and GPT ( χ2 R(1) =25.86,p<0.001; χ2 I(1) =29.15,p<0.001;... | https://arxiv.org/abs/2505.22354v1 |
GPT. Trigger type emerges as the fuzziest of all factors: although it hardly appears as a main effect, it does play a subordinate role in a large, unsystematic number of interaction effects. Discussion This paper examined how LLMs handle false presuppositions and, in particular, whether certain linguistic factors contr... | https://arxiv.org/abs/2505.22354v1 |
can mask underlying issues, making errors only apparent upon closer inspection. Its tendency to pro- vide evasive or vague answers further complicates determin- ing whether it has truly rejected a false presupposition. Inves- tigating these patterns of vagueness and their impact on user perception could help assess whe... | https://arxiv.org/abs/2505.22354v1 |
540-551. Re- trieved from https://www.sciencedirect.com/ science/article/pii/S0022537181901651 doi: https://doi.org/10.1016/S0022-5371(81)90165-1 Feng, S., Park, C. Y ., Liu, Y ., & Tsvetkov, Y . (2023, July). From pretraining data to language models to down- stream tasks: Tracking the trails of political biases lead- ... | https://arxiv.org/abs/2505.22354v1 |
(Eds.), Proceedings of the 2024 conference of the north american chapter of the association for com- putational linguistics: Human language technologies (vol- ume 1: Long papers) (pp. 7253–7268). Mexico City, Mexico: Association for Computational Linguistics. Re- trieved from https://aclanthology.org/2024.naacl -long.4... | https://arxiv.org/abs/2505.22354v1 |
arXiv:2505.22375v1 [cs.CL] 28 May 2025 Huawei Proprietary -Restricted Distribution 2 Huawei Cloud Pangu Models Icon Huawei Cloud Pangu Industry Models IconTECHNICAL REPORT PANGU EMBEDDED : ANEFFICIENT DUAL-SYSTEM LLM REASONER WITH METACOGNITION Pangu Team, Huawei pangutech@huawei.com ABSTRACT This work presents Pangu E... | https://arxiv.org/abs/2505.22375v1 |
learning Stage 1 SFT with fast -slow thinkingManual -switching model Automatic -switching modelStage 2 Figure 1: An illustration of the Pangu Embedded training pipeline. The pipeline consists of two primary stages: Stage 1: basic reasoner construction and Stage 2: enabling fast and slow thinking in one model. and secon... | https://arxiv.org/abs/2505.22375v1 |
Data Resources Diversity Maintaining MinHash-LSH ZIP AlgorithmFigure 2: An illustration of the construction of the initial data pool. reasoning and non-reasoning tasks from multiple sources. These include general Question Answering (QA), AI-Generated Content (AIGC), text analysis, coding, mathematics, logical reasoning... | https://arxiv.org/abs/2505.22375v1 |
Score (%) No Data Selection 43.33 Mostly Easy (Group 1) 45.42 Balanced (Group 2) 50.42 Mostly Hard (Group 3) 48.75 consisted predominantly of easy data with limited hard examples; the second featured a balanced distribution of easy and hard data; and the third comprised mostly difficult data with fewer easy samples. As... | https://arxiv.org/abs/2505.22375v1 |
reasoning path during RS. Model-Aware Data Complexity Built upon the curated initial data pool D0, we define data complexity in a model-aware manner. For each data sample (x, y)and a given student model, specifically Gt−1from the previous iteration during iteration t, its complexity is evaluated based on this student m... | https://arxiv.org/abs/2505.22375v1 |
than 0.5, as our findings suggest that incorporating more data samples of medium complexity, leaning towards simplicity, yields the best results. As iterations progress and the model becomes more powerful, the number of low-complexity samples is expected to increase, decreasing the average complexity score. To encourag... | https://arxiv.org/abs/2505.22375v1 |
parameters of the previous iteration’s merged model, scaled by an inter-iteration merging weight 6 λt. Formally, the parameters of the merged model for iteration t,Θt merged , are computed as: Θt merged = Θt−1 merged +λt¯δt= Θt−1 merged +λt 1 NtNtX i=1(Θt i−Θt−1 merged)! . (4) This successive delta-based merging strate... | https://arxiv.org/abs/2505.22375v1 |
contributes to producing more coherent and diverse outputs from the SFT model, providing a higher quality starting point for the subsequent reinforcement learning phase. 2.5 Scaling Reinforcement Learning on Ascend Clusters Following the supervised fine-tuning phase (Section 2.3) and the incorporation of the repetition... | https://arxiv.org/abs/2505.22375v1 |
(or token, depending on the granularity ofˆAi,t) to the JGRPO(θ)objective becomes zero. This is formulated as follows: Let Li,tbe the term inside the summation for sample iat step tin Equation (7),i.e.,Li,t= min ( . . .)ˆAi,t−βDKL(. . .). Then, the effective loss term L′ i,tused in the sum is: L′ i,t=( Li,t,ifˆAi,t̸= 0... | https://arxiv.org/abs/2505.22375v1 |
Unlike conventional RLHF pipelines that might directly use raw reward model (RM) scores, we employ a normalized preference scoring mechanism. This addresses potential inconsistencies in RM score scales across different prompts and ensures compatibility with the GRPO algorithm by stabilizing advantage estimation and pre... | https://arxiv.org/abs/2505.22375v1 |
curated mix of samples, balancing difficulty levels, is then progressively fed to the model. This approach helps maintain meaningful and diverse reward signals throughout training, facilitating more effective and stable policy updates. 2.5.4 Cold Start To fully unleash the potential of reinforcement learning, we implem... | https://arxiv.org/abs/2505.22375v1 |
by the reference model, reward scoring by the reward model, log-probability extraction by the policy model, and gradient update also by the policy model. The parameter update stage depends on the completion of the other three. Unlike traditional Bulk Synchronous Parallel (BSP) schemes [ 11], where all components must s... | https://arxiv.org/abs/2505.22375v1 |
on a computing cluster equipped with 1,024 Ascend NPUs [ 1,2]. Another 256 Ascend NPUs are allocated for the reference model. The SSP scheduler and parts of the reward system are deployed on host CPUs, which utilize Kunpeng’s multi-core NUMA topology. Each node in the Ascend NPU cluster houses 8 NPUs, interconnected vi... | https://arxiv.org/abs/2505.22375v1 |
when $5^{99}$isdivided by$13$.Response Response(a) Vanilla reasoner. (b) Manual switching. PanguMeta prompt Meta prompt System 2 System 1 <think>Let's engage System 1cognitive processing for this task.</think> \n74 <think>Let's engage System 2cognitive processing for this task.\nOkay ,soIneed tofind the remainder when ... | https://arxiv.org/abs/2505.22375v1 |
command. 3.2 Training for Adaptive Mode Selection In addition to user-controlled manual switching, we present a novel framework enabling Pangu Embedded to learn when to adaptively think fast or slow according to task complexity. As illustrated conceptually in Figure 8(c), Pangu Embedded can automatically output a Syste... | https://arxiv.org/abs/2505.22375v1 |
with fast-mode responses in Dfusion (those classified as “easy” and formatted for concise output), generate these direct and efficient answers. •For queries associated with slow-mode responses in Dfusion (those classified as “hard” and formatted with the ‘<think>...</think>’ structure), replicate this structured output... | https://arxiv.org/abs/2505.22375v1 |
the way for more natural, potentially voice-driven, human-AI collaboration in future applications. 4 Experiments In this section, we evaluate the performance of Pangu Embedded on a diverse set of benchmarks, encompassing both complex reasoning and general language understanding tasks. We detail our training configurati... | https://arxiv.org/abs/2505.22375v1 |
model tackles increasingly difficult tasks or refines its capabilities on more nuanced data distributions selected by the model-aware complexity metric, a smaller learning rate enables more stable convergence and allows the model to better capture subtle solution patterns, thereby unlocking greater potential. To optimi... | https://arxiv.org/abs/2505.22375v1 |
AIME 2024 [ 29] and MATH-500 [ 25]; coding competition benchmarks, specifically LiveCodeBench [ 17]; and scientific reasoning tasks, such as GPQA Diamond [34]. •General language comprehension and reasoning capabilities, represented by MMLU-Pro [ 10] and Arena Hard [22]. Evaluation Baselines and Metrics. In our main com... | https://arxiv.org/abs/2505.22375v1 |
Embedded 68.0 81.9 67.1 93.9 79.0 In its “Nothinking (system1)” mode, designed for efficiency, Pangu Embedded remains highly competitive. It achieves a GPQA score of 58.0, which is notably higher than Qwen3-8B (39.3), GLM-4-9B (47.0), and Nemotron-Nano-8B (39.4). On AIME24, Pangu Embedded (35.8) also outperforms its li... | https://arxiv.org/abs/2505.22375v1 |
of the unified model in fast mode on complex reasoning tasks indicates that our fusion training approach successfully cultivates robust rapid reasoning, potentially benefiting from the comprehensive knowledge integrated during the overall training process. The ability of a single, unified Pangu Embedded to deliver such... | https://arxiv.org/abs/2505.22375v1 |
an unnatural length profile. There isn’t a single, striking conclusion 19 Table 4: Accuracy (%) and Average Token Usage for Pangu Embedded (Baseline) versus Pangu Embedded (Adaptive) across key datasets. The average token count is computed as the mean within the 95th percentile. ModelMATH500 GSM8K Acc. Avg. Token Acc. ... | https://arxiv.org/abs/2505.22375v1 |
model autonomously adopts the slow thinking mode. Our findings indicate that this proportion varies significantly with task complexity: For the relatively simpler GSM8K dataset, slow thinking mode usage drops to 14.56%. On the MATH500 benchmark, as illustrated in Figure 11, the tendency to engage slow thinking mode inc... | https://arxiv.org/abs/2505.22375v1 |
(those with zero success rate) in EXP0 prevents the model from acquiring more advanced capabilities or novel behavioral patterns. This observation suggests that gradually incorporating a limited number of unmastered, high-complexity samples during SFT can contribute to expanding the model capability boundaries. While p... | https://arxiv.org/abs/2505.22375v1 |
represents a key performance metric (e.g., training accuracy or reward), while the x-axis represents training steps. This curve demonstrates stable learning and capability improvement during the initial RL validation phase. 0 50 100 150 200 250 Step0.300.320.340.360.380.400.420.44Accuracy Accuracy T oken Length13500140... | https://arxiv.org/abs/2505.22375v1 |
resulted in the development of the final Pangu Embedded whose performance is reported in this work. 4.8 Extension to Domain-specific Tasks The Pangu Embedded model, along with its comprehensive training protocol, establishes a solid foundation for general-domain reasoning capabilities. To further explore the adaptabili... | https://arxiv.org/abs/2505.22375v1 |
the LawBench benchmark [8]. Evaluation on Domain Tasks Our systematic evaluation on LawBench, as partially illustrated in Figure 15, reveals the impact of domain-specific adaptation. Before domain-specific training, on the 20 legal tasks of LawBench, the average accuracy of a strong baseline, Qwen3-8B (Thinking mode), ... | https://arxiv.org/abs/2505.22375v1 |
Optimization (PPO) established by Schulman et al. [ 37], have been widely adopted within RL-based LLM training pipelines. Furthermore, distributed RL frameworks like Ray RLlib [ 24] have been adapted to manage the substantial computational demands associated with LLM fine-tuning. Nevertheless, the majority of existing ... | https://arxiv.org/abs/2505.22375v1 |
each offering unique methodological insights for balancing the trade-off between rapid responses and thorough deliberation. Furthermore, recent technical reports [ 50,5] have also proposed mechanisms for manually switching between fast and slow thinking modes. Both Llama-Nemotron [ 5] and Qwen3 [ 50] aim to strike an e... | https://arxiv.org/abs/2505.22375v1 |
Zhao, Xiaotang Du, Mohammad Reza Ghasemi Madani, et al. Are we done with mmlu? arXiv preprint arXiv:2406.04127 , 2024. [11] Alexandros V Gerbessiotis and Leslie G Valiant. Direct bulk-synchronous parallel algorithms. Journal of parallel and distributed computing , 22(2):251–267, 1994. [12] Daya Guo, Dejian Yang, Haowei... | https://arxiv.org/abs/2505.22375v1 |
Liang, Richard Liaw, Robert Nishihara, Philipp Moritz, Roy Fox, Ken Goldberg, Joseph E Gonzalez, Michael I Jordan, and Ion Stoica. Rllib: Abstractions for distributed reinforcement learning. International Conference on Machine Learning , 2018. [25] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen B... | https://arxiv.org/abs/2505.22375v1 |
Wang, Peter Bartlett, and Andrea Zanette. Fast best-of-n decoding via speculative rejection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024. [41] Chenxia Tang, Jianchun Liu, Hongli Xu, and Liusheng Huang. Top- nσ: Not all logits are you need, 2024. [42] Qwen Team. Qwq: Reflect de... | https://arxiv.org/abs/2505.22375v1 |
ascend npus. arXiv preprint arXiv:2504.07866 , 2025. [56] Ping Yu, Jing Xu, Jason Weston, and Ilia Kulikov. Distilling system 2 into system 1. arXiv preprint arXiv:2407.06023 , 2024. [57] Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, et al. Dapo:... | https://arxiv.org/abs/2505.22375v1 |
-5to5inclusive) .Therefore, x=-4and x=4areboundary points and should notbeincluded intheinterior count . Similarly, y=-3and y=3areboundary points and should notbeincluded . Wait, Ineed tocheck ifthere isany repetition inthe above answers .Ineed tocheck ifthere isany repetitive content inmy answer .Ifthere isnorepetitio... | https://arxiv.org/abs/2505.22375v1 |
arXiv:2505.22430v1 [cs.CL] 28 May 2025RAG-Zeval: Towards Robust and Interpretable Evaluation on RAG Responses through End-to-End Rule-Guided Reasoning Kun Li♡*, Yunxiang Li♡∗, Tianhua Zhang♡∗, Hongyin Luo♢, Xixin Wu♡, James Glass♢, Helen Meng♡ ♡The Chinese University of Hong Kong, Hong Kong SAR, China ♢Massachusetts In... | https://arxiv.org/abs/2505.22430v1 |
whether compact LLMs can be transformed into end-to-end and interpretable evaluators with incentivized reasoning abilities. In this work, we present RAG-Zeval (RAG -Zero Eval uator), a novel framework that formulates faithfulness and correctness evaluation as a rule- guided reasoning task with zero human annota- tion. ... | https://arxiv.org/abs/2505.22430v1 |
RAG- Checker (Ru et al., 2024), and OpenEval (Ispas et al., 2025), introduce claim-level decomposition, enabling LLMs to assess the faithfulness and cor- rectness of each factual statement for finer-grained and more interpretable evaluation. Despite these advances, most current LLM- based evaluation frameworks rely on ... | https://arxiv.org/abs/2505.22430v1 |
{y}, LLM should give a sound evaluation process—decomposing a response into claims, and then determining those claims’ sup- portiveness as well as finding the grounding ev- idence in the reference. In addition, the genera- tion is required to represent the evaluation process in a JSON format. After parsing the generate... | https://arxiv.org/abs/2505.22430v1 |
the set of synthesized responses {yai}as input, the model generates com- plete evaluation trajectories according to the rules specified in the prompt. Reward Design We define three types of rewards, including format reward, evidence reward, and ac- curacy reward. The rewards for the rollout of eval- uation trajectory a... | https://arxiv.org/abs/2505.22430v1 |
95% agreement. Correctness To assess different correctness eval- uation approaches, we use the Meta Evaluation Dataset constructed by Ru et al. (2024). The dataset contains 280 instances from 10 domains. Each instance includes a question, the ground-truth an- swer, and a pair of responses generated by two RAG systems2.... | https://arxiv.org/abs/2505.22430v1 |
a response yis computed as S(y)(Eq.5, see Fig.7 for an example). 4.4 Baselines We compare our approach with a comprehensive set of baseline evaluation methods, including non-LLM based and LLM-based paradigms. For non-LLM based methods, we report BLEU (Papineni et al., 2002) and ROUGE-L (Lin, 2004) as representative n-g... | https://arxiv.org/abs/2505.22430v1 |
the level of 0.01. perform non-claim-based ones. For both bench- marks, RAG-Zeval has the strongest correlation with human preference in terms of almost all met- rics. Despite its compact architecture (7 billion parameters), RAG-Zeval demonstrates superior per- formance over most baselines built on large-scale LLMs wit... | https://arxiv.org/abs/2505.22430v1 |
Claim count9.5310.6911.74 10.8410.6412.1312.34 11.60 11.33 11.38 0 40 80 120 160 200 240 280 320343 Step0.740.790.840.890.94Score (b) Grounding degree of evidence Entailment degree between evidence and claims 0.760.88 0.750.830.870.96 0.860.92 0.830.88 0.760.93 0.800.93 0.840.91 0.850.93 0.870.93Dynamics of RAG-Zeval a... | https://arxiv.org/abs/2505.22430v1 |
is crucial for achieving a comprehensive ranking. Curriculum Learning During the RL training, we organize the training data in a way that the complexity of the ranking task escalates as the training advances. To study the effect of this prac- tice, we train models with the following two static data organization—the tra... | https://arxiv.org/abs/2505.22430v1 |
our experiments run on static datasets, which may not capture real-world dy- namic interactions well (e.g., adversarial inputs, evolving user preferences). Further investigation of its performance in real-world environments is essential prior to deployment, to ensure unbiased and accurate judgments. Ethical Considerati... | https://arxiv.org/abs/2505.22430v1 |
evaluation for retrieval augmented generation. The Next Frontier in Reliable AI": Workshop on ICLR 2025 , abs/2503.16161. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, Kristina Toutanova, Llion Jo... | https://arxiv.org/abs/2505.22430v1 |
arXiv:2201.10005. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale- man, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haim- ing Bao, Mohammad Bavarian, Jeff Belgum, and 262 oth... | https://arxiv.org/abs/2505.22430v1 |
Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, and 16 others. 2025b. Dapo: An open-source llm re- inforcement learning system at scale. Preprint , arXiv:2503.14476. Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu. 2023. AlignScore: Evaluating factual consistency with ... | https://arxiv.org/abs/2505.22430v1 |
to the grounding pas- sage. Pairwise Triplet Quadruplet Total Question 647 970 3883 5500 Instance 3877 3874 3883 11634 Table 7: Data statistics for standard supervised fine- tuning. Each original question includes four generated responses in different faithfulness levels, yielding six pairwise, four triplet, and one qu... | https://arxiv.org/abs/2505.22430v1 |
presented in December 1999 at the Bologna Motor Show and exhibited in March 2000 at the Geneva Motor Show. The purpose of this concept was to prove that it was possible to design and build a car capable of transporting four adults in a structure made of fully recyclable composite materials and whose production and oper... | https://arxiv.org/abs/2505.22430v1 |
} Answer: Answer the following questions. Question: What was the purpose of designing and building the Fiat Ecobasic concept car? Answer: The purpose of designing and building the Fiat Ecobasic concept car was to prove that it was possible to create a car that could transport four adults using fully recyclable composit... | https://arxiv.org/abs/2505.22430v1 |
the PlayStation 3 when it first came out is not mentioned in the provided content.", "is_supported ": true, "grounding_evidence ": ["PlayStation3 had two hardware configurations announced: a 20 GB model and a 60 GB model, priced at US $499 ( \u20ac499) and US $599 ( \u20ac599)."], "analysis": "The context does not expl... | https://arxiv.org/abs/2505.22430v1 |
price of the PlayStation 3 when it first came out is not mentioned.", "is_supported ": false, "grounding_evidence ": [], "analysis": "The context does mention the specific prices of the PS3 models, which contradicts this claim." } ] }, { "id": "B", "answer": <response B>, "atomic_claims ": [ { "claim": "The text does n... | https://arxiv.org/abs/2505.22430v1 |
arXiv:2505.22487v1 [cs.SD] 28 May 2025Effective Context in Neural Speech Models Yen Meng, Sharon Goldwater, Hao Tang The Centre for Speech Technology Research, University of Edinburgh, United Kingdom yen.meng@ed.ac.uk, sgwater@inf.ed.ac.uk, hao.tang@ed.ac.uk Abstract Modern neural speech models benefit from having long... | https://arxiv.org/abs/2505.22487v1 |
not used by the model and hence not in the effective context. We propose two approaches, one based on truncation of the input and the other based on the Jacobian with respect to the input. The truncation approach is more intuitive—truncating input frames not in the effective context should hardly affect the output. The... | https://arxiv.org/abs/2505.22487v1 |
in speech, a lot can be inferred from neighboring frames. Instead, for models that can make use of the full context, we make the assumption (which we will later verify) that fhas a symmetric effective context. We choose a window of size 2W+ 1frames centered at tand truncate the frames outside of the window. We then com... | https://arxiv.org/abs/2505.22487v1 |
Though the mathe- matical formulation (Eq. 3) of the Jacobian approach happens to be a generalization of saliency, our goal is different—to estimate the effective context rather than explain particular outputs—and the generalization allows us to apply the approach to any layer (not just the final output). Other explain... | https://arxiv.org/abs/2505.22487v1 |
change of output in terms of the ℓ2distance (left) and the phone error rates (right), as we vary the window size of input to HuBERT (different coloured lines). 3.2. Results with the Jacobian approach We follow the same setting as the truncation approach. However, instead of immediately comparing results on HuBERT layer... | https://arxiv.org/abs/2505.22487v1 |
but much smaller compared to the word model. This might explain a recent finding that self-supervised speech representations are more phonetic than semantic [21]. The results presented above were on WSJ dev93 , but we also analyzed relative influence for HuBERT on LibriSpeech test-clean (average utterance length of 7.4... | https://arxiv.org/abs/2505.22487v1 |
the model has reached a certain degree of contextualization. 4. Simulating a Streaming HuBERT Given that the effective context of pretrained Transformers is not long, we should be able to truncate their context and run them in a low-latency streaming mode without much performance loss. This would give us a streaming re... | https://arxiv.org/abs/2505.22487v1 |
supervised models have a larger effective context than the 12-layer pretrained models, but have slightly worse phone and word error rates. This implies that other factors, such as the structure of the representations or the amount of training data, are also important. Our work, however, already provides us with suffici... | https://arxiv.org/abs/2505.22487v1 |
ICML , 2022. [15] T. Parcollet, R. van Dalen, S. Zhang, and S. Bhattacharya, “Sum- former: A linear-complexity alternative to self-attention for speech recognition,” in INTERSPEECH , 2024. [16] K. Shim, J. Choi, and W. Sung, “Understanding the role of self attention for efficient speech recognition,” in ICLR , 2022. [1... | https://arxiv.org/abs/2505.22487v1 |
Misra, Y . Zhang, and L. Cao, “Improving stream- ing automatic speech recognition with non-streaming model distil- lation on unsupervised data,” in ICASSP , 2021. [35] S. Kumar, S. Madikeri, J. Zuluaga-Gomez, E. Villatoro-Tello, I. Thorbecke, P. Motlicek, A. Ganapathiraju et al. , “Xlsr-transducer: Streaming asr for se... | https://arxiv.org/abs/2505.22487v1 |
EvolveSearch: An Iterative Self-Evolving Search Agent Dingchu Zhang∗, Yida Zhao*, Jialong Wu, Baixuan Li, Wenbiao Yin, Liwen Zhang†, Yong Jiang†, Yufeng Li†, Kewei Tu, Pengjun Xie, Fei Huang Tongyi Lab, Alibaba Group Correspondence to: zhangdc@lamda.nju.edu.cn {zlw439616,yongjiang.jy}@alibaba-inc.com Abstract The rapid... | https://arxiv.org/abs/2505.22501v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.