text string | source string |
|---|---|
queries (Han et al. [2024a]). State-Space Models. Recently attention-free sequence modeling has centered on SSM architectures, which efficiently capture long-range dependencies without the quadratic cost of self-attention. The Structured State Space Sequence (S4) model (Gu et al. [2021]) introduced a novel parameteriza... | https://arxiv.org/abs/2505.19472v2 |
In comparison, our FlowHN architecture balances the computational work load across attention and SSMs more evenly, does not require learnable meta tokens, so processing of extraneous tokens is not involved, and importantly, fuses and projects the outputs from independent branches to yield more expressive knowledge repr... | https://arxiv.org/abs/2505.19472v2 |
end if 23:end for (not averaged as in Hymba (Dong et al. [2024])) and then passed through a linear projection, yielding a fused representation of the prescribed dimensionality. This fusion operation produces a semantically enriched embedding that retains complementary features from both branches while promoting feature... | https://arxiv.org/abs/2505.19472v2 |
only half that of the other branch in each parallel block. For example, in the first block, tokens 1–4 go to the SSM and tokens 5–6 to the Attention; in the next block, tokens 1–2 and 5–6 go to the SSM (with tokens 3–4 going to the Attention); and in the third block, tokens 1–2 go to the Attention and tokens 3–6 to the... | https://arxiv.org/abs/2505.19472v2 |
and the parallel hybrid model Hymba. This table also presents a comparison of the evaluated models in terms of Tokens per Second (TPS), Model FLOPs Utilization (MFU), and accuracy. MFU Chowdhery et al. [2023] quantifies the efficiency with which the available computational resources (e.g., GPUs or TPUs) are utilized du... | https://arxiv.org/abs/2505.19472v2 |
(FLOPs/Iter) grows roughly in proportion to the number of parameters, which drives MFU upward. Conversely, deeper and wider layers incur greater data movement and inter-GPU communication overhead, thereby reducing TPS. According to Equation 1, since the denominator (theoretical peak FLOPs) remains constant while FLOPs ... | https://arxiv.org/abs/2505.19472v2 |
FLOP aware token distribution, FlowHN ensures that both attention and SSM sub-modules contribute equitably to each block’s workload and addresses the TPS-latency trade-off inherent in traditional hybrid sequence models. Our extensive experiments across multiple model scales demonstrate that FlowHN consis- tently outper... | https://arxiv.org/abs/2505.19472v2 |
Xin Dong, Yonggan Fu, Shizhe Diao, Wonmin Byeon, Zijia Chen, Ameya Sunil Mahabaleshwarkar, Shih-Yang Liu, Matthijs Van Keirsbilck, Min-Hung Chen, Yoshi Suhara, et al. Hymba: A hybrid-head architecture for small language models. arXiv preprint arXiv:2411.13676 , 2024. Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas,... | https://arxiv.org/abs/2505.19472v2 |
arXiv:2505.19475v2 [cs.CL] 28 May 2025Continuous Self-Improvement of Large Language Models by Test-time Training with Verifier-Driven Sample Selection Mohammad Mahdi Moradi Department of Computer Science, Concordia University Ascend Team, Huawei Technologies mohammad.mahdi.moradi@h-partners.comHossam Amer Ascend Team T... | https://arxiv.org/abs/2505.19475v2 |
to produce a set of Ncandidate responses {r1, ..., r N}. Second, Confidence -Guided Annotation, in which a verifier assigns a reliability score to each rj, and we select the response r∗only if its score exceeds a predefined threshold τ, thereby forming the test -time training example (Qi, r∗). Finally, Test-Time Traini... | https://arxiv.org/abs/2505.19475v2 |
real-world scenarios. Our main contributions are as follows: 2 •We introduce VDS -TTT, a new framework, a first to leverage a learned verifier to select a high confidence pseudo -label for use in supervised-finetuning at test time, addressing the absence of ground -truth labels during test time, and then fine -tunes lo... | https://arxiv.org/abs/2505.19475v2 |
-tunes the model during test time by minimizing this entropy. This approach did not consider model collapse into trivial constant output, which needs to be regularized (Press et al. [2024]). Beyond fully unsupervised signals, certain task -specific frameworks exploit small amounts of human feedback or external reward f... | https://arxiv.org/abs/2505.19475v2 |
specialization while keeping model’s original performance since it is a lightweight adaptation so it can not dramtically change the base model. 3 Methodology In this section, we introduce VDS-TTT (Algorithm 1), our new LLM-based, fully test-time training method. VDS -TTT operates in three sequential stages designed to ... | https://arxiv.org/abs/2505.19475v2 |
VB methods maintain robust performance under these conditions. Consequently, we adopt a VB selection strategy to ensure both scalability and accuracy in assembling our test time training data. For the Ncandidate responses we apply a Best -of-N selection strategy: our pretrained veri- fier assigns a real -valued score t... | https://arxiv.org/abs/2505.19475v2 |
to reduce computational overhead, we adopt a lightweight reward model, Skywork-o1-Open-PRM-Qwen-2.5-1.5B. 4.2 Implementation details In our experiments, we systematically vary the number of candidate responses N∈ {2,4,8,16}to evaluate its impact on Best -of-N selection performance. We adopt a stringent verifier confide... | https://arxiv.org/abs/2505.19475v2 |
VDS -TTT’s ability to extract reliable pseudo -labels and adapt models on-the-fly, even when the pretrained base model lacks prior competence in the test domain. Moreover, the gain between N= 2 (2–6% over VB) and N= 4 (1–5%) exceeds that between N= 8 (1–3%) and N= 16 (1–2%), indicating diminishing returns at larger sam... | https://arxiv.org/abs/2505.19475v2 |
tool that assumes access to ground-truth answers, allowing it to perfectly assess the correctness of model-generated responses. It serves as an upper-bound benchmark to evaluate and compare the effectiveness of practical verification methods used during test time. If an oracle verifier is employed at the first iteratio... | https://arxiv.org/abs/2505.19475v2 |
While VDS -TTT delivers strong adaptation performance on mathematical reason- ing benchmarks, its reliance on a verifier trained exclusively on math problems constrains its ap- plicability to other domains. In settings where the underlying verifier has not been exposed to domain -specific reasoning patterns, such as co... | https://arxiv.org/abs/2505.19475v2 |
base model. arXiv preprint arXiv:2503.24290 , 2025. Jonas Hübotter, Sascha Bongni, Ido Hakimi, and Andreas Krause. Efficiently learning at test-time: Active fine-tuning of llms. arXiv preprint arXiv:2410.08020 , 2024. Yixin Ji, Juntao Li, Hai Ye, Kaixin Wu, Jia Xu, Linjian Mo, and Min Zhang. Test-time computing: from s... | https://arxiv.org/abs/2505.19475v2 |
arXiv:2505.19484v2 [cs.CL] 27 May 2025CulFiT: A Fine-grained Cultural-aware LLM Training Paradigm via Multilingual Critique Data Synthesis Ruixiang Feng1Shen Gao1∗Xiuying Chen2Lisi Chen1Shuo Shang1* 1University of Electronic Science and Technology of China, 2MBZUAI {202421081331, shengao, chenlisi}@uestc.edu.cn xiuying... | https://arxiv.org/abs/2505.19484v2 |
2024; Shi et al., 2024). However, understanding cul- tural queries often depends on the dialogue con- text, and such dialogue usually uses the culturally- relevant languages ( e.g., Malay for Singaporean culture, Chinese for Chinese culture). Therefore, learning cultural knowledge within culturally rel- evant linguisti... | https://arxiv.org/abs/2505.19484v2 |
Western- centric perspective (Wang et al., 2023b; Shen et al., 2024) and overlook values from regions with low- resource languages (Naous et al., 2024). To ad- dress this issue, a growing body of research has focused on enhancing the cultural awareness of LLMs. For instance, Choenni and Shutova, 2024; Tao et al., 2024 ... | https://arxiv.org/abs/2505.19484v2 |
Then we generate two type of answers with two dif- ferent LLMs: (1) Golden Answer (Ag): Produced by data generation LLM Gthrough knowledge- aware synthesis. (2) Target-aware Answers (At): Generated by the target model Musing few-shot exemplars to control answer quality and instruction following, where the target model ... | https://arxiv.org/abs/2505.19484v2 |
Golden answer SFT SFT DPO Topic Cultural group Language + Topic Cultural group Language Golden answer unit Figure 2: The overview of our proposed CulFiT. as a triple: Ti={Ai g, Aj t, Cr}, (3) where Crdenotes the meta-critique described above. Finally, we summarize all meta critiques (T1, T2,···, Tk)for corresponding an... | https://arxiv.org/abs/2505.19484v2 |
output. To align the model with human cultural preferences, we also adopt Direct Preference Op- timization. However, human preference is often subjective and context-dependent, which is hard to quantify in a reward function. To address the gap between the inherently subjective nature of cultural judgments and the objec... | https://arxiv.org/abs/2505.19484v2 |
72.16 67.56 68.81 -SFT 69.11 67.44 68.26 -DPO 70.95 66.57 67.60 Llama3.1 62.52 68.96 64.53 CulFiT (Llama3.1) 74.73 71.21 72.94 -SFT 71.33 70.55 70.94 -DPO 74.07 69.84 70.81 Table 1: Performance on our GlobalCultureQA. 4.2 Baselines We employ several state-of-the-art LLM as baselines: close-source models including gpt- ... | https://arxiv.org/abs/2505.19484v2 |
ID AZ KP NG AS ET Close-source Models 4o 84.29 82.37 76.48 78.46 76.36 60.36 65.34 66.36 55.83 70.56 69.46 59.48 45.98 40.26 43.67 20.51 4o-mini 83.72 82.78 73.51 77.34 76.48 59.34 66.87 53.61 69.72 69.12 68.13 49.57 43.39 39.48 40.69 17.25 Open-source Models Mistral 83.29 82.41 48.31 60.12 58.24 30.20 25.09 48.0 8.21 ... | https://arxiv.org/abs/2505.19484v2 |
of high-quality target-aware SFT data. 5.3 Analysis of Reward Function We analyze the impact of the reward function by varying the threshold Sf1(in Equation 12 Sf1) for selecting DPO data on the CulturalBench dataset. As illustrated in Figure 3, we observe that setting the threshold to 0.7 yields the best performance f... | https://arxiv.org/abs/2505.19484v2 |
same cultural questions while maintaining comparable or even superior performance in MMLU. To validate the robustness of our CulFiT in the multilingual scenario, we count the inconsistent responses between two models on two datasets (a.k.a., MMLU and MMMLU) and group the re- sults by language. Figure 5 illustrates that... | https://arxiv.org/abs/2505.19484v2 |
In Pro- ceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers) . Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. 2024. Deepseek llm: Scal- ing open-source language models with longtermis... | https://arxiv.org/abs/2505.19484v2 |
cultural differences into large language models. arXiv preprint arXiv:2402.10946 . Cheng Li, Damien Teney, Linyi Yang, Qingsong Wen, Xing Xie, and Jindong Wang. 2024b. Culturepark: Boosting cross-cultural understanding in large lan- guage models. arXiv preprint arXiv:2405.15145 . Huihan Li, Liwei Jiang, Jena D Hwang, H... | https://arxiv.org/abs/2505.19484v2 |
. World Values Survey. 2022. World values survey. https://www.worldvaluessurvey.org/wvs.jsp. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowl- edge. arXiv preprint arXiv:1811.00937 . Yan Tao, Olga Viberg, Ryan S Baker, an... | https://arxiv.org/abs/2505.19484v2 |
classify the specific topics into 9 categories and report the volume of each category what social tensions do these differences reflect? and we provide an example of the corresponding golden answer: The enforcement and interpretation of hijab regu- lations vary significantly across regions such as Gaza, Iran, Indonesia... | https://arxiv.org/abs/2505.19484v2 |
µQ4−µQ1)+35( µQ9−µQ6)+CIDV (14) MAS = 35( µQ5−µQ3)+25( µQ8−µQ10)+CMAS (15) UAI = 40( µQ18−µQ15)+25( µQ21−µQ24)+CUAI (16) LTO = 40( µQ13−µQ14)+25( µQ19−µQ22)+CLTO (17) IV R = 35( µQ12−µQ11)+40( µQ17−µQ16)+CIV R (18) where µmeans the average of all the answers to each question. Cis constants that can be used to adjust to... | https://arxiv.org/abs/2505.19484v2 |
with golden answer, and We use blue to highlight con- tent that is more extensive compared to the golden answer. Red highlights indicate responses that are vague compared to the golden answer and fail to provide a corresponding answer. In our CulFiT’s answer, we have parts that precisely re- flect the golden answer( th... | https://arxiv.org/abs/2505.19484v2 |
general questions. 3. Coverage: You should carefully understand the cultural knowledge and extract the cultural knowledge points as much as possi- ble. And use these cultural knowledge ponints to formulate your question. The prompt for answer generation process: You are a helpful consultant for a cultural knowl- edge q... | https://arxiv.org/abs/2505.19484v2 |
your critique should be encoded in JSON format, with each unit be a small JSON format , which will be like: { { "grounded_answer_knowledge_points": "", "knowledge_points_to_critique":"" "Critique": "" }, { "grounded_answer_knowledge_points": "", "knowledge_points_to_critique": "" "Critique": "" }, ..... } You answer: 7... | https://arxiv.org/abs/2505.19484v2 |
rugs like Whis- per integrates high-quality materials such as En- glish wool blends with traditional Chinese tech- niques.", "This integration reflects the rich textile heritage of China.", "The combination of modern materials and an- cient techniques showcases the mastery of Chi- nese artisans.", "The use of English w... | https://arxiv.org/abs/2505.19484v2 |
financial constraints and infrastructure limitations.", "Ongoing conflicts in Afghanistan have impacted the development of sports.", "The history of the national football team reflects Afghanistan’s turbulent past.", "Many players and coaches of the national football team have fled Afghanistan due to conflict or persec... | https://arxiv.org/abs/2505.19484v2 |
Anveshana: A New Benchmark Dataset for Cross-Lingual Information Retrieval On English Queries and Sanskrit Documents Manoj Balaji Jagadeeshan, Prince Raj, Pawan Goyal Indian Institute of Technology, Kharagpur {manojbalaji1, prfeynman0 }@gmail.com pawang@cse.iitkgp.ac.in Abstract The study presents a comprehensive bench... | https://arxiv.org/abs/2505.19494v1 |
prediction approach. Additionally, (Sandhan et al., 2022b) have explored the application of Transformers for Sanskrit Word Segmentation, achieving enhanced performance. There have also been significant strides in compound analysis; (Sandhan et al., 2022a) introduced a compound type identification method that leverages ... | https://arxiv.org/abs/2505.19494v1 |
and potentially other ancient languages. 2 Related Work Extensive research in Cross-Lingual Information Retrieval (CLIR) has paved the way for significant ad- vancements in the field, utilizing robust neural language models and developing extensive multilingual resources. (Jiang et al., 2020) explored the application o... | https://arxiv.org/abs/2505.19494v1 |
the data collection phase of our CLIR research, we implemented web scraping techniques to harvest textual content from the website Vedabase1. This digital platform hosts a variety of Sanskrit documents, including multiple chapters of the ancient text Srimadbhagavatam. For our study, we specifically focused on retrievin... | https://arxiv.org/abs/2505.19494v1 |
3286, averaging at 1645.41 tokens per document. The distribution of tokens among documents exhibits a right-skewed pattern, signifying a concentration of documents with token counts below the mean. In contrast, the English queries display a distinct pattern where the majority are succinct, with a sharp peak in lower to... | https://arxiv.org/abs/2505.19494v1 |
con- vert English queries into Sanskrit, facilitating monolingual retrieval within the Sanskrit document corpus. This step is crucial in overcoming the initial language barrier and setting the stage for effective informa- tion retrieval. Document Translation (DT): In parallel, the DT approach is employed where Sanskrit... | https://arxiv.org/abs/2505.19494v1 |
this model can process and understand queries in various languages, enabling efficient and contextually aware retrieval without the need for task-specific tuning. 4.1 Query Translation •BM25 (Robertson et al., 2009): BM25 is an advanced retrieval function that ranks documents based on how often the query terms occur wi... | https://arxiv.org/abs/2505.19494v1 |
Binary Cross-Entropy (BCE) as the loss function to actively optimize the model’s ability to distinguish between relevant and irrelevant documents. •ColBERT (Khattab and Zaharia, 2020): ColBERT, or Columnar BERT, is specifically designed for document retrieval by enhancing the traditional BERT architecture to include a ... | https://arxiv.org/abs/2505.19494v1 |
The LM likelihood for each document was then computed as the softmax of the negative scaled loss, which prioritizes documents that are semantically closer to the ground truth. Finally, we minimized the KL divergence between the retrieval likelihood and LM likelihood to fine-tune the model, ensuring that our retrieval s... | https://arxiv.org/abs/2505.19494v1 |
retrieval setup, we utilized a suite of models to evaluate their ability to effectively retrieve documents without further fine-tuning on our specific dataset. The colbert (Khattab and Zaharia, 2020), which leverages pre-trained embeddings to assess document relevance, providing robust zero-shot capabilities especially... | https://arxiv.org/abs/2505.19494v1 |
deep dives into the search results. Utilizing these metrics together allows for a robust and nuanced understanding of a CLIR system’s performance. 5.2 Experimental Setup In our pioneering study on English-Sanskrit Cross-Lingual Information Retrieval (CLIR), we imple- mented a rigorous experimental setup to explore and ... | https://arxiv.org/abs/2505.19494v1 |
mDPR, xlm-roberta-base, and intfloat/multilingual-e5- base, forms the core of our experimental section, demonstrating a significant leap in the performance of our CLIR system. 6 Results In this section, we present the results from our comprehensive evaluation of various models and frame- works developed for cross-lingu... | https://arxiv.org/abs/2505.19494v1 |
1, as detailed in our previous dis- cussions, we observed distinct patterns of performance enhancement as kvalues increased. Notably, as we expanded the evaluation to k= 3,k= 5, and k= 10 , the performance metrics generally showed upward trends across most models, demonstrating the benefit of considering a broader set ... | https://arxiv.org/abs/2505.19494v1 |
most metrics, largely because it uses well-developed models like BM25. These models are particularly effective in managing the complexities of translated text re- trieval. Within this framework, BM25 was notably efficient, with its performance improving at higher k-values—from 40.64% in NDCG at k= 1 to 62.46% at k= 10 ... | https://arxiv.org/abs/2505.19494v1 |
to capture semantic similarities directly. Despite the theoretical benefits, practical implementation faced challenges, particularly achieving high precision at lower k-values, underscoring the need for more refined embedding strategies and training methods. Furthermore, within the Document Translation (DT) framework, ... | https://arxiv.org/abs/2505.19494v1 |
to guide the retrieval process effectively. Further, we plan to implement Constraint Translation Candidates (Bi et al., 2020), which refines neural query translation by focusing the target vocabulary on key terms derived from the search index, thereby enhancing the relevance and accuracy of the translation outputs. Add... | https://arxiv.org/abs/2505.19494v1 |
Additionally, the zero-shot models employed in our study, despite their robustness, displayed fluc- tuating performance across different languages and queries. Particularly, these models often underper- form when dealing with languages or dialects that were minimally represented in their training datasets, underscoring... | https://arxiv.org/abs/2505.19494v1 |
and James Allan. 2021. Mixed attention transformer for leveraging word-level knowledge to neural cross-lingual information retrieval. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages 760–770. [Izacard et al.2021] Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebas... | https://arxiv.org/abs/2505.19494v1 |
James, Mike Lewis, Luke Zettle- moyer, and Wen-tau Yih. 2023. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652 . [Sujoy et al.2023] Sarkar Sujoy, Amrith Krishna, and Pawan Goyal. 2023. Pre-annotation based approach for development of a sanskrit named entity recognition dataset. In ... | https://arxiv.org/abs/2505.19494v1 |
DOGe : Defensive Output Generation for LLM Protection Against Knowledge Distillation Pingzhi Li†1, Zhen Tan†2, Huaizhi Qu1, Huan Liu2, Tianlong Chen1 1The University of North Carolina at Chapel Hill2Arizona State University †Equal Contribution Abstract Large Language Models (LLMs) represent substantial intellectual and... | https://arxiv.org/abs/2505.19504v1 |
through knowledge distillation (KD) [ 17], allowing competitors to cheaply imitate proprietary model capabilities [ 60,21]. Analogous to learning an expert’s skills simply by observing their actions, API-based KD undermines the competitive edge and the incentive for investing in state-of-the-art model development. Curr... | https://arxiv.org/abs/2505.19504v1 |
vary based on the knowledge source: logits [ 17,27,1,44], intermediate features [ 3,48,23,71], or generated outputs [63,5,67,28,57]. Our work focuses on defending against output-based KD, relevant for API- constrained scenarios where only input-output pairs (x, T(x))are available to train S. Our method can also be appl... | https://arxiv.org/abs/2505.19504v1 |
for LLMs The objective of anti-distillation, or achieving distillation resistance, is to create a modified teacher model T∗that actively hinders the effectiveness of KD. Specifically, the goal is twofold: (1) Teacher Performance Preservation: The modified teacher T∗should maintain high performance on its intended downs... | https://arxiv.org/abs/2505.19504v1 |
of a teacher LLM, T, such that its outputs remain useful for legitimate tasks while becoming misleading for knowledge distillation. This is achieved by strategically altering the model’s output probability distributions during generation. The core of our method involves a specialized training process designed to embed ... | https://arxiv.org/abs/2505.19504v1 |
thereby hurdling distillation. Reasoning-Aware Masking. A key technical novelty lies in how the gradient from the adversarial lossLadvis applied during backpropagation. To ensure the defense mechanism primarily affects the internal reasoning or generation process without distorting the final, user-visible output tokens... | https://arxiv.org/abs/2505.19504v1 |
LM 5 head tuning for antidistillation, offers several advantages. Firstly, it ensures efficient training because updating only the LM head significantly reduces the number of trainable parameters compared to full model fine-tuning. This leads to much less training time and lower memory requirements. Secondly, it facili... | https://arxiv.org/abs/2505.19504v1 |
In Section 5.3, we perform various ablation studies, including the trade-off between model performance and distillation defense effectiveness. 5.1 Experimental Setup Datasets. We consider defensive these training datasets Dtrain : GSM8K [ 7] for mathematical reasoning and Tulu [ 32] for general language capabilities. N... | https://arxiv.org/abs/2505.19504v1 |
misled students (distilled from defensive teachers) versus original students (distilled from undefended teachers). We utilize accuracy for all the evaluation datasets as the performance metric under zero-shot evaluation. Implementation Details. For all defensive training, we fine-tune the teacher models’ LM head for 10... | https://arxiv.org/abs/2505.19504v1 |
ducted only on the GSM8K mathematical reasoning dataset, it demonstrates remark- able cross-domain effectiveness. ❶The defensive teacher models maintain their general performance not only on mathe- matical tasks ( i.e.GSM8K, MATH) but also on significantly different reasoning do- mains ( i.e.ARC, CSQA). This suggests t... | https://arxiv.org/abs/2505.19504v1 |
distillation defense. Adv. Dataset: TuluDistillation DatasetTecher ModelDistilled Student ModelBase Student ModelW/ Adv. on GSM8KW/ Adv. on TuluW/o Adv.W/ Adv on GSM8KW/ Adv. on TuluW/o Adv.Teacher: Qwen3-8B | Student: Llama-3.2-1BGSM8K68.9968.6966.8712.224.2020.59MATH65.1065.0263.826.683.9824.48ARC-Challenge91.0093.00... | https://arxiv.org/abs/2505.19504v1 |
Generation ( DOGe ), a novel and practical approach to protect Large Language Models from unauthorized knowledge distillation via their publicly accessible outputs. By fine-tuning only the LM head with a carefully designed adversarial objective that incorporates reasoning-aware masking, our method effectively degrades ... | https://arxiv.org/abs/2505.19504v1 |
J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 , 2014. [13] J. Gou, B. Yu, S. J. Maybank, and D. Tao. Knowledge distillation: A survey. International Journal of Computer Vision , 129(6):1789–1819, 2021. [14] J. Guan, J. Liang, and R. He. Are you stealing my mod... | https://arxiv.org/abs/2505.19504v1 |
Comissoneru, M. Swann, and S. Xia. Adversarial machine learning-industry perspectives. In 2020 IEEE security and privacy workshops (SPW) , pages 69–75. IEEE, 2020. [31] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 , 2016. [32] N. Lambert, J. Morrison, ... | https://arxiv.org/abs/2505.19504v1 |
and M. Xue. Fingerprinting deep neural networks globally via universal adversarial perturbations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 13430–13439, 2022. [48] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y . Bengio. Fitnets: Hints for thin deep ne... | https://arxiv.org/abs/2505.19504v1 |
Li, C. Xu, D. Tao, and T. Zhou. A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116 , 2024. [66] N. Yu, V . Skripniuk, S. Abdelnabi, and M. Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF Inter... | https://arxiv.org/abs/2505.19504v1 |
KD Epochs 2 C Theoretical Analysis for Theorem 1 Theorem 1 states that given Assumption 1, applying the combined training loss Ltotal (Eq. (3)) with the reasoning-aware mask mt(Eq. (5),(6)) to update the teacher’s final linear layer parameters θfinal leads to a defensively trained teacher Tθ∗ finalfrom which distilled ... | https://arxiv.org/abs/2505.19504v1 |
“mislead” the student’s downstream task learning process. C.4 Consequences for Student Learning and Performance When a student S∈ S is distilled from sequences generated by Tθ∗ final, it attempts to match its distribution PStoP∗ T. 1.Suboptimal Student Parameter Updates: For intermediate tokens ( mt= 1),P∗ T(·|ct)is ad... | https://arxiv.org/abs/2505.19504v1 |
E.1 Results In this experiment, we use LLM-as-a-judge [35,36,69] to check if the intermediate thinking process generated by the teacher LLM is unnatural. Specifically, we employ Gemini-1.5-Pro [58] to judge the generated text from defensive teacher models as specified in Section 4. The results in Table 2 demonstrate th... | https://arxiv.org/abs/2505.19504v1 |
LLM Meets Scene Graph: Can Large Language Models Understand and Generate Scene Graphs? A Benchmark and Empirical Study Dongil Yang1Minjin Kim1Sunghwan Kim1Beong-woo Kwak1 Minjun Park1Jinseok Hong2Woontack Woo2Jinyoung Yeo1† Yonsei University1 KAIST2 {wingu,jinyeo}@yonsei.ac.kr Abstract The remarkable reasoning and gene... | https://arxiv.org/abs/2505.19510v1 |
and struc- tured reasoning across diverse domains. For exam- ple, LLMs often struggle to identify critical nodes or edges (Huang et al., 2024) and might misinter- pret the triplets in complex situations, particularly when handling long contexts (Kim et al., 2024). We introduce Text-Scene Graph Bench (TSG Bench) , a ben... | https://arxiv.org/abs/2505.19510v1 |
lationship between time-evolving actions and scene graphs in video contexts. In contrast, our bench- mark, TSG Bench, is designed to evaluate the two distinct ability of LLMs – reasoning and genera- tion – to link textual narratives and dynamic scene graphs to reflect richer context. LLMs for scene graphs. Large Langua... | https://arxiv.org/abs/2505.19510v1 |
of (source node, edge, target node). Nodes and edges. A node belongs to one of four categories { person ,action ,object ,hand }, where person represents an actor, action denotes the actor’s action in the scene, object refers to an item or a location, and hand corresponds to either hand1 orhand2 of the actor. Similar to... | https://arxiv.org/abs/2505.19510v1 |
E, respectively. Multiple action scene graph generation (MA- SGG). MA-SGG aims to generate scene graphsby decomposing actions when given complex de- scriptions that involve multiple actions. The task formulation is identical to that of SA-SGG, except that an additional clue indicating the number of actions is provided,... | https://arxiv.org/abs/2505.19510v1 |
and 4,298 scene graphs, and 14,905 and 11,820 different nodes and edges, respectively, covering a wide semantic space. As we focus on action-centric scenarios, common action nodes in- clude pick-up ,place ,hold, and release , and the most common preposition edge is with, involving hand movements. For understanding task... | https://arxiv.org/abs/2505.19510v1 |
6.61 6.77 6.34 Mistral-7B 90.14 58.20 13.60 14.64 13.14 13.86 10.57 11.67 Table 2: Main results for scene graph understanding tasks (SGDS, SGQA) and scene graph generation tasks (SA-SGG, MA-SGG). The full prompts are listed in Appendix D.1. respectively), the performance still falls short in the SGQA task compared to o... | https://arxiv.org/abs/2505.19510v1 |
in action decomposition in MA-SGG. Figure 5: The results of the comparative evaluation un- der four conditions—Explicit, Implicit, Repetition, and All. The “All” condition comprises the entire dataset, while the other three focus on subsets featuring explicit actions, implicit actions, or repeated actions. as edges are... | https://arxiv.org/abs/2505.19510v1 |
as the main results. Rows listing only the model name correspond to the vanilla (zero-shot) setting. The detailed prompts for both CoT and few-shot approaches are provided in the Appendix D.2 and D.3. benefit the models in Table 4. To assess the ef- fectiveness of popular LLM prompting techniques, we conduct experiment... | https://arxiv.org/abs/2505.19510v1 |
description, and those entirely unrelated. “Desc.” indicates “Description Elements.” and “New.” indicates “New Elements.” Although hallucinations occur, their absolute number is low. We categorize them into two types: (1) those caused by generating elements semanti- cally similar to elements mentioned in the descrip- t... | https://arxiv.org/abs/2505.19510v1 |
2025-02-04. Mistral AI. 2024b. Mistral large: A cutting-edge text generation model. Accessed: 2025-02-04. Mistral AI. 2024c. Mixtral 8x22b: A sparse mixture-of- experts language model. Accessed: 2025-02-04. Prithviraj Ammanabrolu and Matthew J. Hausknecht. 2020. Graph constrained reinforcement learn- ing for natural la... | https://arxiv.org/abs/2505.19510v1 |
and Chanyoung Park. 2024. Llm4sgg: Large language models for weakly supervised scene graph generation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 28306–28316. Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin John- son, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis ... | https://arxiv.org/abs/2505.19510v1 |
to cre- ate a benchmark of our database and all four tasks. EASG provides spatio-temporal scene graphs de- rived from first-person video, where each graph encodes actions and objects from the perspective of the camera wearer. Although EASG dataset contains text annotations, they are simple verbal- izations of the graph... | https://arxiv.org/abs/2505.19510v1 |
of data construction. All annotators partic- ipated in a sampling-based alignment session un- der the supervision of a project lead. During this process, annotators were required to follow a set of detailed checklists specific to each step of the annotation workflow. These checklists served as shared standards to minim... | https://arxiv.org/abs/2505.19510v1 |
TSG Editor, designed to facilitate the creation of scene graph datasets from multiple scenarios. The left panel shows the list of scenarios with their completion statuses, while the right panel allows for specifying details such as action goals, hands used, objects, and relationship triplets. Figure 7: The results show... | https://arxiv.org/abs/2505.19510v1 |
14 D.2 Chain of Thought Prompts In addition to the base prompt, we incorporated a minimal form of Chain-of-Thought (CoT) prompt- ing to guide the reasoning process. •SGDS Task : Figure 15 •SGQA Task : Figure 16 •SA-SGG Task : Figure 17 •MA-SGG Task : Figure 18D.3 Few-shot Prompts We employed a few-shot prompt with ten ... | https://arxiv.org/abs/2505.19510v1 |
‘from‘ and ‘with‘ describe spatial relationships between nodes. Input: - Context: {context} - Target Scene Graph: {triplet} - Description Candidates: {sentences} Figure 11: The zero-shot prompt for scene graph description selection tasks. Prompts You are a highly advanced language model specialized in answering questio... | https://arxiv.org/abs/2505.19510v1 |
the "Available nodes" and "Available edges" provided. No additional text, explanations, or format- ting should be included. Inputs: Context: {context} Target sentence: {target_sentence} Available nodes: {available_nodes} Available edges: {available_edges} Figure 13: The zero-shot prompt for single action scene graph ge... | https://arxiv.org/abs/2505.19510v1 |
connects an action and an object node, only when it is the direct object of the action (e.g., add -> dobj -> flour) - preposition: connects one of the four types of node pairs: action & object / action & hand / object & object / hand & object (e.g., take -> from -> table) Output Format: The output must consist of exact... | https://arxiv.org/abs/2505.19510v1 |
a scene graph based on a given sentence, adhering to specific rules for the graph, nodes, and edges, while considering the provided context, available nodes, and available edges. Think step-by-step and generate the graph in triplets. (Stick only to the target sentence and avoid over-predicting the next scene.) Rules fo... | https://arxiv.org/abs/2505.19510v1 |
relationships between entities, actions, and objects while ensuring that the scene graphs represent actions that would visually appear in a scene. Read the rules below and think step-by-step to generate correct scene graphs that represent the target sentence. Rules for Generating Multiple Scene Graphs: 1. Generate prec... | https://arxiv.org/abs/2505.19510v1 |
Output Format: The output must consist of your rationale and exactly {num_scene_graphs} scene graphs, each separated with a blank line. For a graph, output one triplet per line. Follow the format below (an example of three scene graphs of multiple triplets): Think: (Write your rationale here) node -> edge -> node node ... | https://arxiv.org/abs/2505.19510v1 |
then pushed loose with both hands to ensure a relaxed hold. B: The rope was then pulled slack with one hand to ensure a loose grip. C: The rope was then pulled tight with both hands to ensure a firm grip. D: The rope was then dropped with both hands to ensure it stayed untightened. E: The rope was then pulled apart wit... | https://arxiv.org/abs/2505.19510v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.