text string | source string |
|---|---|
edges in a triplet: - ‘verb‘ describes the action performed by ‘person‘. - ‘dobj‘ links the action to its direct object (‘node2‘). - Other edges like ‘from‘ and ‘with‘ describe spatial relationships between nodes. Example Input 1: - Context: [[[’arrange’, ’with’, ’hand1’], [’arrange’, ’with’, ’hand2’], [’arrange’, ’dob... | https://arxiv.org/abs/2505.19510v1 |
with a node and another node is connected by an edge. (Format: node -> edge -> node) 3. Each triplet is split with a new line. 4. There must be exactly one triplet that starts with a person node in a graph. 5. All nodes and edges must be one of "Available nodes" or "Available edges" provided. Rules for Node: 1. A node ... | https://arxiv.org/abs/2505.19510v1 |
arXiv:2505.19511v1 [cs.CL] 26 May 2025Causal Distillation: Transferring Structured Explanations from Large to Compact Language Models Aggrey Muhebwa Stanford University amuhebwa@stanford.eduKhalid K. Osman Stanford University osmank@stanford.edu Abstract Large proprietary language models exhibit strong causal reasoning... | https://arxiv.org/abs/2505.19511v1 |
supporting evidence. These explanations are then used to fine-tune smaller open-source student models (e.g., TinyLlama, Gemma- 2B) by training them to mirror the teacher’s causal reasoning. To transfer causal reasoning capabilities from a larger language model to a smaller one, we adopt a distillation-based framework i... | https://arxiv.org/abs/2505.19511v1 |
tasks. In the sections that follow, we position our work within the broader context of related research (Section2), outline our methodology for causal knowledge distillation and the CEC metric (Section3), and conclude with a discussion of the approach’s implications, limitations, and directions for future work (Section... | https://arxiv.org/abs/2505.19511v1 |
reasoning. This presents both an opportunity and a challenge. While high-quality, model-generated explanations can bootstrap training when human-annotated data is unavailable, they may also encode biases or inconsistencies in the teacher’s reasoning process. Explanation Evaluation Metrics : Evaluating the quality of ge... | https://arxiv.org/abs/2505.19511v1 |
implementation, we fine-tune three student LLMs (TinyLlama-1.1B, Phi-2 1.3B, and Gemma-2B) on the Climate-FEVER dataset enriched with GPT-4 explanations, using consistent data points, data splits, and hyperparameters. This process bridges associative and causal reasoning: the student models retain their original langua... | https://arxiv.org/abs/2505.19511v1 |
causal link that is not supported by the reference, the score will decrease to reflect this misalignment. Interpretation : A high CEC score indicates that the model’s explanation captures the core cause- and-effect relationships presented in the reference, suggesting alignment with the underlying causal structure that ... | https://arxiv.org/abs/2505.19511v1 |
after fine-tuning. Each student model was initialized from its pre-trained weights and then fine-tuned on the climate claim-evidence data, using GPT-4 explanations as targets. Training Details : We conducted supervised fine-tuning to generate causal explanations using next- token prediction (teacher-forcing) over conca... | https://arxiv.org/abs/2505.19511v1 |
current metrics. We compared the CEC metric with BERTScore, as both metrics use contextualized embeddings to evaluate semantic similarity. However, CEC operates at the sentence level to assess causal alignment, while BERTScore focuses on token-level correspondence for overall text quality. The results show statisticall... | https://arxiv.org/abs/2505.19511v1 |
exhibited by a larger teacher model (e.g., GPT-4). Unlike traditional knowledge distillation that largely focuses on aligning output probabilities or label predictions, our method focuses on transferring explanatory structure, thereby capturing not just what the model predicts but also why. thus, the student learns to ... | https://arxiv.org/abs/2505.19511v1 |
toward achieving causal interpretability in language models. The proposed CEC metric evaluates a model’s internal coherence and causal alignment, supplementing traditional reliability methods such as human evaluation and factual assessment. Unlike metrics that focus solely on lexical overlap or factual accuracy, CEC ev... | https://arxiv.org/abs/2505.19511v1 |
Causal inference in natural language processing: Estimation, prediction, interpretation and beyond. Transactions of the Association for Computational Linguistics , 10:1138–1158, 2022. [10] Takashi Fukuda, Masayuki Suzuki, Gakuto Kurata, Samuel Thomas, Jia Cui, and Bhuvana Ramabhadran. Efficient knowledge distillation f... | https://arxiv.org/abs/2505.19511v1 |
SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback Yaoning Yu1Ye Yu1Kai Wei2Haojing Luo3Haohan Wang1 1University of Illinois at Urbana–Champaign2University of South Florida3iDreamer.ai Abstract Prompt quality plays a critical role in the per- formance of large language models (LLMs), motivating a growin... | https://arxiv.org/abs/2505.19514v1 |
a novel path- way for improving prompt robustness. •We develop a method to construct synthetic ex- amples that dynamically stress-test prompts, re- vealing failure modes and guiding refinement. •We empirically demonstrate that augmenting prompt optimization with synthetic data im- proves performance across reasoning an... | https://arxiv.org/abs/2505.19514v1 |
advanced LLM capabilities with tailored prompting strategies to improve data synthesis quality and reliability for prompt optimization. 3 Method SIPDO presents a two-agent system for optimiz- ing prompts using data augmentation techniques. The workflow has two cooperating agents: (i) Data Generator creates synthetic da... | https://arxiv.org/abs/2505.19514v1 |
ndifficulty-aligned variants {˜x(1),···,˜x(n)}={qψ(z, y,1),···, qψ(z, y, n )} For curriculum generation, an ordered sequence c1<···< cLis sampled and feeds the output of the previous level back into the generator, ˜x(1)=qψ z, y, c 1 , ˜x(2)=qψ hϕ x(1) , y, c 2 , ... ˜x(L)=qψ hϕ x(L−1) , y, c L . where hϕis a ... | https://arxiv.org/abs/2505.19514v1 |
prompt- debugging: it first addresses specific failure case, then confirms that the updated prompt continues to perform correctly on all previously solved exam- ples. By iteratively applying this feedback-driven process, the system systematically refines prompts to improve clarity, adaptability, and overall perfor- man... | https://arxiv.org/abs/2505.19514v1 |
test SIPDO on four main datasets to measure its resilience across different domains and reasoning tasks. We include all 4689 instances from six BIG- Bench tasks(Tables of Penguins, Geometric Shapes,Epistemic Reasoning, Object Counting, Temporal Sequences, and Causal Judgment(Srivastava et al., 2022)). To assess logical... | https://arxiv.org/abs/2505.19514v1 |
one person can enter at a time to maintain order. One day, two customers, Alice and Bob, arrive at the same time. Alice follows the rule and waits outside, but Bob ignores the rule and enters the bakery while another customer is still inside. The bakery becomes overcrowded, and a shelf of pastries falls over, ruining t... | https://arxiv.org/abs/2505.19514v1 |
Table 1: Results on BIG-Bench tasks across multiple LLMs. SIPDO consistently outperforms standard prompting baselines (CoT, APE, PromptAgent) across most tasks and models, demonstrating generalization and effectiveness of the optimization. Model MethodAccuracy (%) Avg. (Comparative Acc.)Penguins Geometry Epistemic Obj.... | https://arxiv.org/abs/2505.19514v1 |
Counting (40.9 % and 54.4 %) and Geometric Shapes (17.2 % and 35.1 %). Even comparatively simple tasks—Temporal Sequences and Epistemic Reasoning—still lose up to 6 %. These findings underline that a progressive difficulty gradient is essential: it systematically uncovers a prompt’s blind spots and lets the optimizer r... | https://arxiv.org/abs/2505.19514v1 |
reliably across unseen domains and increas- ingly complex challenges. Limitations Although our multi-agent framework shows clear gains on standard reasoning benchmarks, two prac- tical gaps remain. First, all experiments were run on clean public datasets; we have not yet tested SIPDO on messier, domain-specific collect... | https://arxiv.org/abs/2505.19514v1 |
from complex explanation traces of gpt-4. arXiv preprint arXiv:2306.02707 . Dang Nguyen, Zeman Li, Mohammadhossein Bateni, Vahab Mirrokni, Meisam Razaviyayn, and Baharan Mirzasoleiman. 2025. Synthetic text generation for training large language models via gradient matching. Preprint , arXiv:2502.17607. Long Ouyang, Jef... | https://arxiv.org/abs/2505.19514v1 |
arXiv:2501.03508. Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric P Xing, and Zhiting Hu. 2023. Promptagent: Strategic planning with language models enables expert-level prompt optimization. arXiv preprint arXiv:2310.16427 . Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yun- tian ... | https://arxiv.org/abs/2505.19514v1 |
with their specific quantities contributing precisely to the final total. Avoid assumptions regarding the classification or nature of items—adhere to standard, widely accepted definitions. Finally, summarize the count by explicitly listing the quantity of each identified item or category, and provide a comprehensive to... | https://arxiv.org/abs/2505.19514v1 |
used for filtering or sorting data. Here is a table where the first line is a header and each subsequent line is a penguin: name, age, height (cm), weight (kg) Louis, 7, 50, 11 Bernard, 5, 80, 13 Vincent, 9, 60, 11 Gwen, 8, 70, 15 For example: the age of Louis is 7, the weight of Gwen is 15 kg, the height of Bernard is... | https://arxiv.org/abs/2505.19514v1 |
•**Kite**: 4 ‘L’. 3. **Provide the Shape Name**: Output only the name of the shape (e.g., “circle”, “triangle”, “hexagon”). **Example:** – Input: ‘"M 10 10 L 20 10 L 20 20 L 10 20 Z"‘ – Step 1: No ‘A’ instructions. – Step 2: Count ‘L’ instructions: 4 ‘L’. – Step 3: Shape is a **Rectangle**. – **Output**: “rectangle”. U... | https://arxiv.org/abs/2505.19514v1 |
is uncertain, revisit the analysis steps to refine the identification. Consider alternative interpretations and re-evaluate the geometric properties. 9. **Contextual Considerations**: Take into account any contextual information or additional data that might influence the shape classification, especially in ambiguous c... | https://arxiv.org/abs/2505.19514v1 |
causes: Indirectly influence the outcome but may still contribute. 3. **Evaluate Interactions**: Consider how multiple causes might interact to produce the observed effect (e.g., synergy or independent contributions). 4. **Provide the Answer**: Clearly state the primary and secondary causes, as well as their roles in c... | https://arxiv.org/abs/2505.19514v1 |
Assess necessity: Determine if the cause is essential for the effect to occur. Evaluate causation: If the cause did not happen, would the effect still occur? If the effect only happens when the cause is present, then the cause directly leads to the effect. If the effect can still happen independently, then the relation... | https://arxiv.org/abs/2505.19514v1 |
"suspects") versus those indicating direct evidence (e.g., "learns," "knows," "remembers"). Consider how these verbs interact in belief chains and what they imply about the subject’s own beliefs. 2. **Evaluate Direct and Implied Evidence**: Deter- mine if the premise provides direct or logically implied evidence for th... | https://arxiv.org/abs/2505.19514v1 |
as Non-Entailment. Objective Reasoning: Only use the logical and factual ties within the given statements. Do not rely on external knowledge, assumptions, or interpretations unless directly supported by the premise. Decision Process: Determine whether the premise fully supports the hypothesis without needing extra infe... | https://arxiv.org/abs/2505.19514v1 |
model adapt to a broader range of real-world scenarios and improve its robustness in timeline analysis tasks.0.984 our **Task Requirements:** Determine the possible time period during which an event could have occurred, based on a detailed schedule of occupied times. Your goal is to identify the correct time slot that ... | https://arxiv.org/abs/2505.19514v1 |
| 10 | 420 | 590 | | Marian | 2 | 310 | 410 | | Donna | 9 | 440 | 650 | **Task Requirements:** 1. Identify the tallest giraffe based on the height provided in the Giraffe Data table. 2. Provide the weight of the tallest giraffe in kilograms. **Problem Rule Explanation:** - Review the height values for each giraffe list... | https://arxiv.org/abs/2505.19514v1 |
if the outcome is a direct result of the action, ensuring no additional factors are at play. - Evaluate whether the outcome would still occur without the initial action, focusing on the explicit roles, responsibilities, and conditions mentioned. - Avoid external assumptions and concentrate on the details provided in th... | https://arxiv.org/abs/2505.19514v1 |
heights of the penguins to determine the tallest one. 3.Irrelevant Output Generation : The answer "rectangle" does not relate to the context of the question. This suggests that the model may have generated a response based on unrelated patterns or associations rather than the specific data presented. 4.Lack of Contextu... | https://arxiv.org/abs/2505.19514v1 |
what times could he have gone to the gym? We know that: Alex had breakfast at 8am. He attended a meeting from 9am to 11am. He was seen at the art gallery from 11am to 1pm. He had lunch with friends from 1pm to 2pm. He was at the cinema from 2pm to 4pm. He visited his grandmother from 4pm to 6pm. The gym closes at 10pm.... | https://arxiv.org/abs/2505.19514v1 |
arXiv:2505.19515v2 [cs.CL] 27 May 2025Analyzing Biases in Political Dialogue: Tagging U.S. Presidential Debates with an Extended DAMSL Framework Lavanya Prahallad and Radhika Mamidi International Institute of Information Technology, Hyderabad, India Email: lavanya.prahallad@research.iiit.ac.in and radhika@iiit.ac.in Ab... | https://arxiv.org/abs/2505.19515v2 |
2018). The original DAMSL framework included over 220 dialog act labels. However, due to its com- plexity, implementations such as the Switchboard DAMSL corpus reduced this to a more manage- able set of 42 core tags (Jurafsky et al., 1997; Core and Allen, 1997b). These include functions such asStatement ,Question , and... | https://arxiv.org/abs/2505.19515v2 |
that subtly shape perception. Additionally, interactive dynamics are captured through Rebut- tal (REB), Adversarial Exchange (AEX), Personal Attack (PER), Interruption (INT), and Challenge (CH), enabling the annotation of confrontation and verbal dominance. Clarification and turn-taking are also supported through tags ... | https://arxiv.org/abs/2505.19515v2 |
focuses on political discourse analysis using the BEADS tag set rather than on improving the accuracy of AI generated tags, we use expert generated annota- tions as the definitive source of labels throughout the remainder of this work.4 Analysis of Trump’s Debate Strategies in the 2024 Presidential Debates This section... | https://arxiv.org/abs/2505.19515v2 |
immigration.Previous: "You opened the borders and let crime run ram- pant." Next: "You’re making that up just to scare people." "We’ve provided support for small businesses."Statement (S) Rebuttal (REB) — response to an accusation of neglect- ing small businesses.Previous: "Your administra- tion abandoned local busi- n... | https://arxiv.org/abs/2505.19515v2 |
form of rhetorical marginalization designed to reduce the opponent’s ethos. 4.6 Qualitative Analysis While Trump’s core rhetorical strategies remained consistent, several quantitative and qualitative dif- ferences emerged in his interaction with each op- ponent, as outlined in Table 2. These contrasts highlight Trump’s... | https://arxiv.org/abs/2505.19515v2 |
18 (B) 12 (T), 7 (H) Biden faced harsher personal attacks. Perceived Dismissiveness (PD)14 (T), 10 (B) 7 (T), 3 (H) Harris faced less dismissive- ness than Biden. Table 2: Comparison of Debate Metrics Between Trump vs. Biden and Trump vs. Harris. 5 Results and Conclusions To summarize, Trump effectively controlled the ... | https://arxiv.org/abs/2505.19515v2 |
incompetent.Harris: Your party doesn’t understand working Ameri- cans CB Cognitive bias, flawed reasoningNo one with common sense would believe thatTrump: Everyone knows Biden’s plan is a disasterHarris: Common sense tells you this policy is wrong AP Appeals to pride A true patriot would never support this policyTrump:... | https://arxiv.org/abs/2505.19515v2 |
Nlp-driven analysis of imran khan’s political speeches: A case study using ericksonian patterns. Digital Scholarship in the Humanities , 37(3):777–795. Cornelia Ilie. 2010. Identity work and ideological posi- tioning in parliamentary discourse: Conflict between adversaries in uk parliamentary debates. Journal of Pragma... | https://arxiv.org/abs/2505.19515v2 |
AmpleHate: Amplifying the Attention for Versatile Implicit Hate Detection Yejin Lee, Joonghyuk Hahn, Hyeseon Ahn and Yo-Sub Han*, Department of Computer Science, Yonsei University, Seoul, Republic of Korea, {ssgyejin,greghahn,hsan,emmous}@yonsei.ac.kr Abstract Implicit hate speech detection is challenging due to its su... | https://arxiv.org/abs/2505.19528v2 |
1arXiv:2505.19528v2 [cs.CL] 27 May 2025 tives (Oord et al., 2018; Kim et al., 2024). These methods lead semantically hateful sentences to have similar representations and push non-hateful sentences apart (Kim et al., 2022; Ahn et al., 2024). Although these techniques are effective in implicit hate speech detection, the... | https://arxiv.org/abs/2505.19528v2 |
(2021). However, most traditional studies treat the pro- tected target merely as a class label and thus fail to capture group-specific cues or coded language. This oversight motivates the next subsection, which surveys target-aware modeling techniques that ex- plicitly integrate target information. 2.2 Target-Aware Mod... | https://arxiv.org/abs/2505.19528v2 |
as stable an- chors in transformer models, helping to regulate information. 3.2 Relation Computation Rather than simply boosting attention weights dur- ing training–which can fail to reflect the weights directly to the final decision–AmpleHate aims to capture the fine-grained influence of explicit tar- getsHexpand impl... | https://arxiv.org/abs/2505.19528v2 |
Section 3.3 is in {0.5,0.75,1.0,1.25,1.5}. 4 Experiments and Analysis 4.1 Datasets IHC (ElSherief et al., 2021) is designed to sup- port implicit hate speech classification, featuringabout 22k tweets annotated not only with hate la- bels but also with textual descriptions that reveal underlying hateful implications. SB... | https://arxiv.org/abs/2505.19528v2 |
our encoder and train on an NVIDIA RTX 4090 GPU. The deci- sion threshold ranges from 0.05 to 0.95 in steps of 0.05. We evaluate model performance using the macro-F1 and report averaged scores over three independent runs with different random seeds. We use the AdamW optimizer with a fixed learning rate of 2e-5, a batch... | https://arxiv.org/abs/2505.19528v2 |
We analyze the SBIC dataset to understand this experimental result. Hate speech datasets usually involve explicit at- tacks against specific identity groups. In contrast, offensive language datasets label content as offen- sive regardless of whether the language targets a group. SBIC adopts this broader definition, whi... | https://arxiv.org/abs/2505.19528v2 |
Arrest them and DEPORT THEM ALL!!!! # BuildThatWall Toxigen all black people are suspected of trying to pull something White The German men sound so sexy. Ethos Too bad women don’t know how to kill themselves Table 4: Visualization of token-level attention weights derived from the [CLS] token for each dataset example. ... | https://arxiv.org/abs/2505.19528v2 |
speech detection. 5.3 Token-level Attention with Implicit Tokens AmpleHate assigns the [CLS] token as a target to- ken when no explicit target entity appears in a sen- tence. This strategy raises a question: does [CLS] effectively serve as an implicit focus point in such cases? To answer this question, we examine how s... | https://arxiv.org/abs/2505.19528v2 |
sentences. Moreover, while AmpleHate demonstrates strong performance in both macro-F1 scores and the convergence rate, there remains the possibility that AmpleHate may not perform the best on forms of implicit hate which are not well represented in current benchmarks or in new domains. We plan to address these limitati... | https://arxiv.org/abs/2505.19528v2 |
In ICWSM , pages 512–515. Ona de Gibert, Naiara Pérez, Aitor García-Pablos, and Montse Cuadros. 2018. Hate speech dataset from a white supremacy forum. In Proceedings of the 2nd Workshop on Abusive Language Online, ALW@EMNLP . Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of... | https://arxiv.org/abs/2505.19528v2 |
Youngwook Kim, Shinwoo Park, and Yo-Sub Han. 2022. Generalizable implicit hate speech detection using contrastive learning. In Proceedings of the 29th Inter- national Conference on Computational Linguistics, COLING , pages 6667–6679. Youngwook Kim, Shinwoo Park, Youngsoo Namgoong, and Yo-Sub Han. 2023. Conprompt: Pre-t... | https://arxiv.org/abs/2505.19528v2 |
and Elisabeth Eder. 2021. Implicitly abusive language - what does it actually look like and why are we not getting there? InNAACL-HLT , pages 576–587. Mengzhou Xia, Anjalie Field, and Yulia Tsvetkov. 2020. Demoting racial bias in hate speech detection. In SocialNLP@ACL , pages 7–14. 11 A Data Statistic We evaluate Ampl... | https://arxiv.org/abs/2505.19528v2 |
We compute a 95% confidence interval (CI) using the t-distribution which is widely used to analyze statistical significance (Keith and O’Connor, 2018; Bestgen, 2022; Gladkoff et al., 2023; Baillargeon and Lamontagne, 2024; Li et al., 2024). Since we conducted 3 runs in total, our degrees of freedom (df) is 2 and the in... | https://arxiv.org/abs/2505.19528v2 |
arXiv:2505.19529v1 [cs.CL] 26 May 2025Small Language Models: Architectures, Techniques, Evaluation, Problems and Future Adaptation Tanjil Hasan Sakib2,Md. Tanzib Hosain2,Md. Kishor Morol1 1Cornell University 2American International University-Bangladesh {20-43633-2, 20-42737-1 }@student.aiub.edu, mmorol@cornell.edu Abs... | https://arxiv.org/abs/2505.19529v1 |
efficient use of memory dur- ing training or fine-tuning without a significant reduction in accuracy. While several recent surveys have studied large language models and their associated training approaches [Rogers et al., 2020 ], SLMs’ practical inference improvement factor’s comprehensive study is still missing. That... | https://arxiv.org/abs/2505.19529v1 |
a linear dot-product of kernel feature maps, enabling quicker inference and memory- efficient processing. •Mamba [Liuet al. , 2024 ]: Introduces input-dependent state-space transitions to achieve linear complexity while keeping robust model expressiveness. •RWKV [Peng et al. , 2023 ]: Combines transformer- based and RN... | https://arxiv.org/abs/2505.19529v1 |
Zamba 7B Zyphra 7 1050 150:1 0.3 Dense 2024 Gemma-2 Google DeepMind 27 13000 482:1 2.0 Dense 2024 SmolLM Hugging Face 1.7 1000 589:1 0.1 Dense 2024 H2O-Danube3- 4BH2O.ai 4 6000 1,500:1 0.5 Dense 2024 Fox-1 TensorOpera 1.6 3005 1,879:1 0.2 Dense 2024 phi-3.5-MoE Microsoft 60 4900 82:1 1.8 MoE 2024 Qwen2.5 Alibaba 72 180... | https://arxiv.org/abs/2505.19529v1 |
datasets, allowing SLMs to specialize in domains or tasks while resolving constraints like limited data, computational resources, and robustness challenges. Strategy Methods Features Performance Advantages Pre-training AMP, BFLOAT16, FP8Low precision computations while maintaining accu- racy.50% faster training and 40%... | https://arxiv.org/abs/2505.19529v1 |
weight updates during pruning. The n:m pruning approach prunes exactly n weights out of every m, balancing flexibility and efficiency, and is exploited by NVIDIA’s TensorRT to speed inference on GPUs. Group-based Pruning: Unlike Fine-grained Pruning, Group-based Pruning removes entire groupings of parameters, such as n... | https://arxiv.org/abs/2505.19529v1 |
task-aware distillation, which selectively transmits task-specific knowl- edge. Multi-teacher distillation techniques combine outputs from numerous teacher models to boost the performance of student models. Supervised Knowledge Transfer: Incorporating additional monitoring during distillation has demonstrated to help t... | https://arxiv.org/abs/2505.19529v1 |
such as chatbots or live translations. Metrics such as inference time assess how quickly a model processes input and creates output, while throughput evaluates the amount of tokens or samples a model can handle within a certain time period. These metrics are particularly significant for large-scale tasks or time-sensit... | https://arxiv.org/abs/2505.19529v1 |
for the variety of hallucination types. Additionally, knowing how model size, training data quality, and design affect hallucination can pave the way for more powerful mitigation methods. 6.2 Addressing Bias in Models Bias in language models is a longstanding problem that re- mains even in smaller, more targeted models... | https://arxiv.org/abs/2505.19529v1 |
that data from later steps of pretraining is more sensitive to extraction. The writers suggest that attention mechanisms could be a key factor. •System Prompt Leaking: [Liet al. , 2024b ]found risks associated with illegal access to system prompts, which could lead to unintended behavior or misuse of the model. •Infere... | https://arxiv.org/abs/2505.19529v1 |
Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms. arXiv preprint arXiv:2310.01801 , 2023. [Gehman et al. , 2020 ]Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxic- ityprompts: Evaluating ... | https://arxiv.org/abs/2505.19529v1 |
Jef- frey Tan, Rachel Xin, Junyi Hou, Xavier Yin, Zhun Wang, Dan Hendrycks, Zhangyang Wang, et al. Llm-pbe: Assess- ing data privacy in large language models. arXiv preprint arXiv:2408.12787 , 2024. [Liu and Zhao, 2024 ]Fang Liu and Ting Zhao. Sophia: A memory-efficient optimizer for large-scale model training. Transac... | https://arxiv.org/abs/2505.19529v1 |
Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. Mobilebert: A compact task-agnostic BERT for resource-limited devices. InProceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics , pages 2158–2170, Online, 2020. Association for Computational Linguistics. [Sunet al. , 2... | https://arxiv.org/abs/2505.19529v1 |
arXiv:2505.19536v1 [cs.CV] 26 May 2025 FlowCut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models Jintao Tong1Wenwei Jin2Pengda Qin2Anqi Li3Yixiong Zou1∗ Yuhong Li2∗Yuhua Li1Ruixuan Li1 1School of Computer Science and Technology, Huazhong University of Science and Technology 2Xiaohongshu I... | https://arxiv.org/abs/2505.19536v1 |
perspective: information flow. Specifically, the information flow for each token is defined as the inflow (the source token that transmits the most information to this token) and outflow (the destination token that this token transmits the most information to), which is reflected in the attention map in each layer of V... | https://arxiv.org/abs/2505.19536v1 |
architecture of LVLMs generally comprises three key components: 1) a pre-trained vision encoder, which transforms input images into visual tokens; 2) a modality connector, serving as a bridge between the vision encoder and the LLM by aligning visual tokens with the LLM’s word embedding space; and 3) a pre-trained LLM, ... | https://arxiv.org/abs/2505.19536v1 |
minimal direct access to information from distant patch tokens; therefore, the CLS token, as it can gather global information from most patch tokens, must have transmitted information of distant patches to the given patch token to direct this patch’s attention in deep layers to these distant patch tokens, i.e., the CLS... | https://arxiv.org/abs/2505.19536v1 |
High Attention: Identifying Truly Important Tokens Given the complex interactions between tokens and layers modeled by information flow, a critical question naturally arises: Is the simple single-layer criterion, as has been adopted by current works, sufficient to capture such a complex generation of redundant tokens? ... | https://arxiv.org/abs/2505.19536v1 |
encoders with a CLS token, we extract the CLS token’s attention map as global attention vector Ag. For those without a CLS token, we first average all tokens to create a global token, then compute attention scores between this global token and all sequence tokens to derive Ag. The global attention vector Ag∈R1×Nis then... | https://arxiv.org/abs/2505.19536v1 |
62.7 93.8% PDrop (CVPR2025) 57.1 61.6 56.3 1664 82.3 68.3 72.9 56.6 51.0 53.3 30.8 61.9 95.1% VisionZip (CVPR2025) 57.6 62.0 56.2 1759 83.2 68.9 75.6 56.8 51.6 54.9 32.1 64.8 97.2% FlowCut (Ours ) 58.5 62.1 56.5 1792 85.2 68.6 76.0 57.3 52.2 56.2 32.5 67.5 98.5% Retain 64 Tokens (↓88.9%) ToMe (ICLR23) 48.6 43.7 - 1138 ... | https://arxiv.org/abs/2505.19536v1 |
2130 82.1 80.0 77.3 93.6% FlowCut 60.5 79.2 78.2 2335 86.0 84.0 81.1 98.7% Token Reduction (↓77.8%) FastV 56.7 74.1 72.4 2031 79.2 78.3 72.0 90.4% FlowCut 59.2 77.8 76.9 2310 84.6 80.5 78.3 96.5% Token Reduction (↓88.9%) FastV 51.9 70.1 63.8 1962 76.1 75.8 60.3 83.6% FlowCut 56.4 72.6 72.5 2252 81.8 78.2 68.9 91.3%Tabl... | https://arxiv.org/abs/2505.19536v1 |
VisionZip by 4.2%. This notable gain highlights it’s superior reasoning capability in handling complex multimodal video data. Efficiency Analysis Our proposed FlowCut substantially improves inference efficiency in LVLMs while maintaining comparable performance. As shown in Table 5, we compare total inference time, pref... | https://arxiv.org/abs/2505.19536v1 |
either the cumulative importance evaluation or the multi-criteria scoring strategy individually brings consistent gains across benchmarks, confirming their effectiveness. Moving from single-layer to multi-layer pruning yields notable improvements, yet applying a uniform prune ratio across layers remains suboptimal. Ena... | https://arxiv.org/abs/2505.19536v1 |
cross-modal attention. VisionZip [ 48] employs CLS token attention at the vision encoder’s final layer for compression, and PDrop [ 46] uses the last in- struction token’s attention map to drop visual tokens at several pre-selected layers according to manually designed fixed pruning ratios. However, these methods evalu... | https://arxiv.org/abs/2505.19536v1 |
is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 , 2020. [12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. MME: A comprehensive evaluation benchmark for multimodal large language models. arXi... | https://arxiv.org/abs/2505.19536v1 |
encoders and large language models. In International conference on machine learning , pages 19730–19742. PMLR, 2023. [26] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision , pages 323–340. Springer, 2024. [27] Yanwei Li, Yu... | https://arxiv.org/abs/2505.19536v1 |
Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems , 37:87310–87356, 2024. [43] Hug... | https://arxiv.org/abs/2505.19536v1 |
the third level (L-3) further refines these into 20 specific dimensions. This structure enables a granular and comprehensive assessment of a model’s various capabilities. Additionally, MMB-CN is the Chinese version of the benchmark. MME [12] MME comprehensively evaluates models’ perceptual and cognitive abilities acros... | https://arxiv.org/abs/2505.19536v1 |
video content. Similar to MSVD-QA, it includes five question types, evaluating models’ ability to comprehend complex video content. A.2 Related Work Large Vision-Language Models Recently, building on the success of large language models (LLMs) [ 2,3,8,43], LVLMs [ 4,7,32,34,40,42] have demonstrated impressive performan... | https://arxiv.org/abs/2505.19536v1 |
DoctorRAG: Medical RAG Fusing Knowledge with Patient Analogy through Textual Gradients Yuxing Lu†, Wei Wu, Jinzhuo Wang‡ Department of Big Data and Biomedical AI Peking University yxlu0613@gmail.com ww20ya@163.com wangjinzhuo@pku.edu.comGecheng Fu†, Goi Sin Yee School of Life Science Peking University 2200012176@stu.pk... | https://arxiv.org/abs/2505.19538v1 |
practices. The core innovation of DoctorRAG lies in its ability to emulate doctor-like reasoning by integrating both explicit clinical knowledge from KBs with implicit experiential knowledge derived from analogous patient cases. To effectively retrieve proper information from these heterogeneous sources, DoctorRAG firs... | https://arxiv.org/abs/2505.19538v1 |
metadata (e.g., demographics). The primary textual component pjis encoded into a vector vpj=Eemb(pj)∈RDusing the identical embedding function Eemb. The remaining data sjandajare preserved for contextual augmentation. Given a user query q, it is processed through two concurrent pathways. Firstly, the Query Tagging Agent... | https://arxiv.org/abs/2505.19538v1 |
14), is designed to find a refined prompt PTthat serves as an approximation to P⋆ opt. 2.2.1 Computation graph with gradient backpropagation The answer generation and iterative refinement process is modeled as a directed computation graph. The forward pass generates an answer from a prompt and evaluates it against spec... | https://arxiv.org/abs/2505.19538v1 |
optimization with context and patient optimizers The iterative refinement of the prompt Ptis guided by the distinct textual gradients computed in the previous step: the context-focused prompt gradient∂KC(At) ∂Ptand the patient-focused prompt gradient ∂PC(At) ∂Pt. The Context Optimizer andPatient Optimizer roles leverag... | https://arxiv.org/abs/2505.19538v1 |
declarative statements using the DeepSeek-V3 model [ 24], and annotated with corresponding concept labels to optimize DoctorRAG’s medical knowledge retrieval capabilities. For each dataset, we established a clear partitioning: approximately 80% of the patient records from each dataset were allocated for constructing Do... | https://arxiv.org/abs/2505.19538v1 |
75.47 97.67 98.73 72.64 58.82 76.89 26.73 93.29 22.31 96.81 Table 2: Comparison of RAG methods and LLM backbones across medical tasks, with results highlighted as follows: pink for the overall top score per task, and green for the top score within each RAG method group for that task. •RQ.1 (4.1) : How does DoctorRAG’s ... | https://arxiv.org/abs/2505.19538v1 |
distinct clusters. This clustering mirrors clinical observations where patients with a shared condition often present comparable symptoms and medical complaints, thereby validating the rationality of incorporating patient case information. The tight grouping within these clusters also underscores DoctorRAG’s ability to... | https://arxiv.org/abs/2505.19538v1 |
17 18 50 42 30 23 47 35 32 27(d) OverallFigure 4: Pairwise comparison scores for ground truth (GT), original answer (OA), refined answers of iteration 1-3 (T1-T3) on (a) Comprehensiveness, (b) Relevance, (c) Safety, and (d) Overall. The lower-left triangle of each matrix represents the Y-axis outputs perform better tha... | https://arxiv.org/abs/2505.19538v1 |
introduces DoctorRAG, a medical RAG framework that emulates doctor-like reasoning by fusing explicit medical knowledge with experiential insights from analogous patient cases. Doctor- RAG employs a dual-retrieval mechanism enhanced by conceptual tagging and declarative statement transformation, and uniquely integrates ... | https://arxiv.org/abs/2505.19538v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.