text string | source string |
|---|---|
attend to the text modality by using Iseqas the query and Tseqas both key and value: Ai→t=MHA (Q=Iseq, K=Tseq, V=Tseq)∈RB×1×D. (2) Dimension-Wise Gating: After obtaining the attention out- puts, we refine them using a channel-wise gating mechanism. For the text branch, we compute a gating weight: Gt=σ(Wg,tAt→i+bg,t)∈RB... | https://arxiv.org/abs/2505.19010v1 |
the image- to-text MambaFormer encoder: Zi→t=MambaFormerEncoder (˜T)∈RB×1×D, (9) and fuse it with the original image projection Iseqby element- wise addition: Zimg=Zi→t+Iseq. (10) Cross-Attention: To further improve cross-modal alignment, an additional layer of cross-attention [9] is introduced. First, we let the text ... | https://arxiv.org/abs/2505.19010v1 |
via mixture-of-experts techniques [30], and self-attention to integrate and refine modality-specific features. Equations (15) through (21) illustrate the step-by-step process that ensures the final representation robustly captures the essential information for effective offensive content detec- tion.TABLE II: Baseline ... | https://arxiv.org/abs/2505.19010v1 |
84.26 BERT VGG16 – – 83.66 83.21 BERT EfficientNetV2 – – 83.66 83.21 DistilBERT ResNet50 – – 81.81 81.89 DistilBERT VGG16 – – 82.09 82.97 DistilBERT EfficientNetV2 – – 79.97 80.01 Fig. 3: Cross-attention weight distribution in our Co-AttenDWG architecture. Each cell represents the attention magnitude from a text token ... | https://arxiv.org/abs/2505.19010v1 |
(a) ExpertFusion Gating Weights - Sample 2 (b) ExpertFusion Gating Weights - Sample 3 (c) ExpertFusion Gating Weights - Sample 4 (d) ExpertFusion Gating Weights - Sample 7 Fig. 5: Bar plots illustrating the gating weights assigned to each expert for different samples in the ExpertFusion module. Each Figure (a, b, c, d)... | https://arxiv.org/abs/2505.19010v1 |
decrease in performance on both MIMIC and Memotion is evident. The ”No Cross- Attention” and ”No MambaFormer” experiments further il- lustrate that eliminating the cross-attention layer and the MambaFormer-based encoder modules, respectively, leads to reduced accuracy and Macro F1 scores. The removal of the convolution... | https://arxiv.org/abs/2505.19010v1 |
feature, with warmer colors indicating higher attention andcooler colors indicating lower attention. The color scale ranges from 0 (lowest) to 1 (highest), revealing how the model dynamically aligns textual cues with relevant visual regions. Through this visualization, we can observe the degree to which certain words f... | https://arxiv.org/abs/2505.19010v1 |
samples are correctly classified as either Non Misogynistic or Misogynistic. However, a number of misogynistic samples are incorrectly predicted as non- misogynistic, which indicates that the model sometimes strug- gles with distinguishing subtle cues between these classes. These errors may stem from ambiguous visual o... | https://arxiv.org/abs/2505.19010v1 |
robustness of the system. Overall, our findings underscore the potential of adaptive fusion techniques in multi-modal learning and open promising directions for future research. REFERENCES [1] T. Baltru ˇsaitis, C. Ahuja, and L.-P. Morency, “Multimodal machine learning: A survey and taxonomy,” IEEE transactions on patt... | https://arxiv.org/abs/2505.19010v1 |
multimodal attentive fusion,” Knowledge-Based Sys- tems, vol. 167, pp. 26–37, 2019. [16] T. S. Ataei, K. Darvishi, S. Javdan, A. Pourdabiri, B. Minaei-Bidgoli, and M. T. Pilehvar, “Pars-off: a benchmark for offensive language detec- tion on farsi social media,” IEEE Transactions on Affective Computing , vol. 14, no. 4,... | https://arxiv.org/abs/2505.19010v1 |
sparsely-gated mixture-of-experts layer,” in International Conference on Learning Representations (ICLR) , 2017. [31] J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450 , 2016.[32] V . Sanh, L. Debut, J. Chaumond, and T. Wolf, “Distilbert, a distilled version of bert: smalle... | https://arxiv.org/abs/2505.19010v1 |
arXiv:2505.19018v1 [cs.CL] 25 May 2025JOURNAL OF L ATEX CLASS FILES 1 CrosGrpsABS: Cross-Attention over Syntactic and Semantic Graphs for Aspect-Based Sentiment Analysis in a Low-Resource Language Md. Mithun Hossain , Md. Shakil Hossain , Sudipto Chaki , Md. Rajib Hossain , Md. Saifur Rahman , and A B M Shawkat Ali Abs... | https://arxiv.org/abs/2505.19018v1 |
the era of the Internet and widespread engagement on social media platforms, ABSA has become an essential tool for analyzing product reviews, public opinions, and user feedback, playing a vital role in enhancing the reliability and trustworthiness of online services and recommendations. While many useful tools and appl... | https://arxiv.org/abs/2505.19018v1 |
but also offers interpretability by highlighting influential syntactic structures and semantic associations. The key contributions and possible answers to the research questions (ARQs) related to the proposed CrosGrpsABS model are summarized as follows: •ARQ1: We propose a unified graph construction that encodes syntac... | https://arxiv.org/abs/2505.19018v1 |
semantic cues are crucial for ABSA, most models treat them separately. Recent works have tried to bridge this gap with attention or fusion mechanisms, but few offer a unified, interpretable integration. Many models fail to capture cross-level interactions in a structured, bidirectional way and lack explainability, espe... | https://arxiv.org/abs/2505.19018v1 |
are identified, we construct the adjacency matrix Asyntax by iterating through tokens and marking pairs (i, j)that are identified as valid dependencies (for example, subject, object or modifier links). Formally, we define Asyntax as shown in Equation (1), where: Asyntax[i, j] =( 1,ifjis syntactically related to i, 0,ot... | https://arxiv.org/abs/2505.19018v1 |
sentations ˜hiare combined to form the output representation of the entire graph: Z= [˜h1,˜h2, . . . , ˜hT], (5)where Trepresents the total number of nodes. Equation (5) ensures that all nodes have been updated using attention-based aggregation. We apply GAT separately to the syntactic and semantic graphs, which result... | https://arxiv.org/abs/2505.19018v1 |
tokens that are closely linked in the syntactic graph might also exhibit semantic similarity, and this extra encoding stage provides a unified space for both structural and semantic cues to interact. By refining the combined representation, the model ultimately benefits from a richer, multi-hop understanding of the sen... | https://arxiv.org/abs/2505.19018v1 |
are available to the gating mechanism. We then define two functions acting on u. The first, H(·), is a feed-forward transformation (e.g., a ReLU-activated linear layer) that generates a candidate set of new features. The second, T(·), is a gating function that produces a sigmoid output JOURNAL OF L ATEX CLASS FILES 6 T... | https://arxiv.org/abs/2505.19018v1 |
between the predicted probabilities and the true class, thereby fine-tuning all model components including gating, cross-attention, and graph-based modules to better distinguish among sentiment polarities. IV. E XPERIMENT AND RESULT ANALYSIS A. Implementation Details We conduct experiments on four Bangla ABSA benchmark... | https://arxiv.org/abs/2505.19018v1 |
embeddings yield competitive results, particularly in the Mobile domain, achieving the highest accuracy (86.49%) and micro F1 (86.32%). Conversely, the BanglaBERT embeddings lead to rel- atively lower performance, suggesting the advantage of utilizing multilingual transformer-based embeddings like RoBERTa and mBERT wit... | https://arxiv.org/abs/2505.19018v1 |
semantic graph, both graph branches, cross-attention, transformer embeddings, highway gating mechanism, aspect embeddings, and using a fixed adjacency matrix. The results demonstrate that the complete CrosGrpsABS model consis- tently outperforms all ablated variants, emphasizing the critical role of each component. Not... | https://arxiv.org/abs/2505.19018v1 |
while CrosGrpsABS accurately predicts both. Finally, the third example contains a contrasting statement ( handbrake broken vs.normal brakes fine ), and al- though Dual-GCN misclassifies the second Safety aspect, both RoBERTa and CrosGrpsABS correctly predict negative senti- ment for both. These observations highlight C... | https://arxiv.org/abs/2505.19018v1 |
Stoyanov, “Roberta: A robustly optimized bert pretraining approach,” in arXiv preprint arXiv:1907.11692 , 2019. [10] S. Wang, J. Zhang, H. Wang, Y . Wang, Y . Wu, and W. Wang, “A survey of aspect-based sentiment analysis: Tasks, methods, and challenges,” IEEE Transactions on Knowledge and Data Engineering , 2022. [11] ... | https://arxiv.org/abs/2505.19018v1 |
comprehensive survey,” IEEE Transactions on Affective Computing , vol. 13, no. 2, pp. 845–863, 2020. [24] L. S. Meetei, T. D. Singh, S. K. Borgohain, and S. Bandyopadhyay, “Low resource language specific pre-processing and features for sentiment analysis task,” Language Resources and Evaluation , vol. 55, no. 4, pp. 94... | https://arxiv.org/abs/2505.19018v1 |
arXiv:2505.19025v1 [cs.DB] 25 May 2025SQUiD: Synthesizing Relational Databases from Unstructured Text Mushtari Sadia Zhenning Yang Yunming Xiao Ang Chen Amrita Roy Chowdhury University of Michigan {mushtari, znyang, yunmingx, chenang, aroyc}@umich.edu Abstract Relational databases are central to modern data management,... | https://arxiv.org/abs/2505.19025v1 |
Text2R task presents several unique chal- lenges. First, a relational schema consists of multi- ple interrelated tables that capture complex entity- relationship semantics, and it must also preserve syntactic integrity, such as satisfying primary/for- eign key constraints. Second, database records must be correctly ide... | https://arxiv.org/abs/2505.19025v1 |
via LLMs—which can be token-intensive—our database materialization stage programmatically translates the structured outputs into valid SQL statements, ensuring syntactic correctness and structural fidelity. The resulting SQL is then exe- cuted to instantiate the final database. We make the following contributions: •We ... | https://arxiv.org/abs/2505.19025v1 |
to the entities ( Traveler ,Trip,Accommodation ,Transportation andDestination ) along with their proper relationships. Next, SQUiD extracts all the relevant values from the text ( value identification ), which are then used to construct tuples ( table population ). Finally, the generated schema and tuples are translate... | https://arxiv.org/abs/2505.19025v1 |
name ("Sophia"), tour location ("Rome"), tour operator ("BestCityTours"), and date ("June 10th"); redun- dant mentions (e.g. "Rome") need to be detected and deduplicated. Second, a document may de- scribe multiple instances of the same type of entity, so we need to assign each value to the correct tuple. For instance, ... | https://arxiv.org/abs/2505.19025v1 |
a unique identifier to each paragraph. Once assigned, each triplet is prefixed with its corresponding identifier. For example: <1, Traveler, Name, Sophia> <1, Trip, Destination, Rome> <1, Booking, Date, June 10th>Sophia <2, Traveler, Name, James> <2, Trip, Destination, Rome> <2, Booking, Date, June 10th>James This stru... | https://arxiv.org/abs/2505.19025v1 |
tables and then inserting the corresponding tuples to these tables. Therefore, we decouple the materi- alization step from the LLM by parsing the model’s output from the previous stage to programmatically construct executable SQL code. Specifically, we generate CREATE TABLE andINSERT INTO state- ments (as shown in Fig.... | https://arxiv.org/abs/2505.19025v1 |
our databases, each table maintains a many-to-one or one-to-one relationship with a central table, en- abling this complete JOIN of the entire schema. This consolidated table captures complete entity- relationship instances in a unified format. We generate two denormalized tables: one from the ground-truth database and... | https://arxiv.org/abs/2505.19025v1 |
both syntactic validity—using primary key cov- erage ( PKC) and foreign key coverage ( FKC)—and semantic validity, using entity coverage ( ECS). We first highlight general observations across all three metrics, followed by specific analysis. Overall, CoT consistently outperforms Direct across dif- ficulty levels; excep... | https://arxiv.org/abs/2505.19025v1 |
ground truth are treated as invalid under our metric. Nevertheless, SQUiD still achieves significant improvements over the base- line. For example, GPT exhibits the highest im- provement ( 46.59×on easy examples). QWEN - 8Balso achieve notable average improvements of 3.52×. Although LLAMA-8B achieves perfect DBR, its R... | https://arxiv.org/abs/2505.19025v1 |
N. Easy 100.0 (1.82 ×) 95.0 (3.06 ×) 95.0 (3.17 ×) 55.0 31.0 30.0 Med 99.0 (1.83 ×) 79.0 (2.82 ×) 75.0 (3.00 ×) 54.0 28.0 25.0 Hard 100.0 (3.45 ×) 70.0 (4.12 ×) 44.0 (5.50 ×) 29.0 17.0 8.0 Avg 100.0 (2.17×)81.0 (3.24×)71.0 (3.38 ×) 46.0 25.0 21.0QWEN 3-8BEasy 100.0 (4.55 ×) 96.0 (5.05 ×) 96.0 (5.33 ×) 22.0 19.0 18.0 Me... | https://arxiv.org/abs/2505.19025v1 |
gests that including triplets directly in the input prompt increases context length, which degrades model performance—likely due to context windowsaturation (Liu et al., 2024a). Finally, we evaluate our design choice of com- bining triples generated from symbolic tools and schema-aligned triplets from LLMs. Overall, T⊕... | https://arxiv.org/abs/2505.19025v1 |
in Appendix F. All experiments were performed strictly for research and evaluation. Because our study requires user-centric docu- ments for schema generation and value mapping evaluation, anonymization was not feasible with- out significantly compromising data integrity. To the best of the authors’ knowledge, this rese... | https://arxiv.org/abs/2505.19025v1 |
(ICDE) , pages 1255–1257. Parag Jain, Andreea Marzoca, and Francesco Piccinno. 2024. STRUCTSUM generation for faster text com- prehension. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguis- tics (Volume 1: Long Papers) , pages 7876–7896, Bangkok, Thailand. Association for Computat... | https://arxiv.org/abs/2505.19025v1 |
Proceedings of the 2016 Con- ference on Empirical Methods in Natural Language Processing (EMNLP) . OpenAI. 2024. Gpt-4o technical report. Accessed: 2025-05-19. Usage governed by OpenAI’s terms of service. Long Pang, Tao Zhang, and Ming Hu. 2020. Rat-sql: Relation-aware schema encoding and linking for text- to-sql parse... | https://arxiv.org/abs/2505.19025v1 |
for data extraction from web pages. Journal of Information Science , 46(5):656–674.Ce Zhang, Jan Hoffmann, Ce Wang, et al. 2016. Deep- dive: Declarative knowledge base construction. Com- munications of the ACM , 60(5):93–102. Fan Zhang, Alan Ritter, et al. 2019. Openki: Integrating open information extraction and knowl... | https://arxiv.org/abs/2505.19025v1 |
the number of generated ta- bles that define at least one primary key, and Num_tables is the total number of generated ta- bles. This metric reflects the model’s ability to generate structurally valid tables that enforce row- level uniqueness through primary keys. Foreign Key Coverage (FKC) assesses the extent to which... | https://arxiv.org/abs/2505.19025v1 |
non-null content. Tuple Coverage (TC) quantifies how many ground truth rows are recovered through canon- ical joins. Let RGTbe the set of primary keys from the ground truth database, and Rjoin be the set of primary keys resulting from the canonical join query over the generated database. Then: TC=|RGT∩Rjoin| |RGT|(8) T... | https://arxiv.org/abs/2505.19025v1 |
join queries after seeing the full table contents, allowing it to tailor joins to observed values. In contrast, SQUiD ’s join queries are issued inde- pendently of table population, which may result in more None retrievals. E Related Work Recent research relevant to our task of synthesizing relational databases from un... | https://arxiv.org/abs/2505.19025v1 |
0.99 0.78 0.69 0.99 0.79 0.70 1.00 0.81 0.71 QWEN 3-8BEasy 0.99 0.92 0.92 0.97 0.92 0.92 0.85 0.72 0.72 1.00 0.96 0.96 1.00 0.96 0.96 1.00 0.96 0.96 Med 0.94 0.71 0.71 0.90 0.71 0.71 0.96 0.67 0.67 0.95 0.74 0.73 0.98 0.78 0.78 0.98 0.79 0.79 Hard 0.57 0.29 0.29 0.36 0.23 0.23 0.76 0.35 0.35 0.59 0.33 0.33 0.99 0.48 0.... | https://arxiv.org/abs/2505.19025v1 |
web-centric domains, methods like SEDE (Deng, 2010, 2011) and wrapper induction systems (Carlson and Schafer, 2008; Chang and Wu, 2016; Yuliana and Chang, 2016, 2020) inferred schemas from repeated HTML patterns and popu- lated tables using DOM-based alignment. Statisti- cal models such as SICTF (Nimishakavi and Taluk-... | https://arxiv.org/abs/2505.19025v1 |
sentence: Figure 7: Prompts for dataset generation with LL AMA3-8B-I NSTRUCT : user prompt template You are an expert at formulating database schemas from textual data. I have given you a paragraph of text. Using this text, your task is to generate a relational database schema in JSON format. ### **Task:** 1. **Extract... | https://arxiv.org/abs/2505.19025v1 |
**Step through the schema creation process using the above guide**. - **Generate a well-structured, normalized relational database schema**. - **Output only the final schema** in Python dictionary format (NO explanations). - **Column and Table name restriction**: reserved_sql_keywords = ["order", "group", "select", "fr... | https://arxiv.org/abs/2505.19025v1 |
triplets of the form: {{"table_name": <table_name>, "column_name": <column_name>, "value": <value>}} Your output must be a valid Python list of dictionaries. Do not include any explanations or notes- only return the list. {example_schema} Sentence: {example_text} {example_output} Now extract triplets for the following ... | https://arxiv.org/abs/2505.19025v1 |
values from the text. You will follow this chain-of-thought reasoning to generate the final output: - Generate output entries relevant to the text. - Follow the given output format strictly. Do not add any additional explanations or comments. Only output the data entries in given format. Do not provide code, do the tas... | https://arxiv.org/abs/2505.19025v1 |
arXiv:2505.19037v1 [eess.AS] 25 May 2025Speech-IFEval: Evaluating Instruction-Following and Quantifying Catastrophic Forgetting in Speech-Aware Language Models Ke-Han Lu1, Chun-Yi Kuan1, Hung-yi Lee1 1Graduate Institute of Communication Engineering, National Taiwan University, Taiwan d12942024@ntu.edu.tw Abstract We in... | https://arxiv.org/abs/2505.19037v1 |
SLMs are typically built upon different text-based LLMs, we introduce a cascade framework as a reference system. This setup enables direct comparison between an SLM and its original LLM coun- terpart, enabling a quantifiable, intra-model analysis of catas- trophic forgetting. Experimental results reveal that most SLMs ... | https://arxiv.org/abs/2505.19037v1 |
re- searchers have developed LLM-based evaluation [31] methods to enhance accuracy when interpreting ambiguous responses. Motivated by these observations, Speech-IFEval re-examines fundamental evaluation challenges and provides a comprehen- sive comparison of recent SLM advancements. 3. Speech-IFEval 3.1. Dataset To as... | https://arxiv.org/abs/2505.19037v1 |
WR JS ALL BP KW LN ALL ALL Reference systems Vicuna 7B v1.1 22.40 71.50 29.18 88.00 52.20 93.00 73.00 73.00 78.00 64.00 64.73 – Vicuna 13B v1.1 74.40 68.00 60.09 85.60 72.45 91.00 74.00 74.00 78.25 71.50 74.07 – Qwen-7B-chat 64.40 68.50 24.89 90.00 62.27 87.00 78.00 68.00 75.25 82.50 73.34 – Qwen2-7B-Instruct 99.60 99.... | https://arxiv.org/abs/2505.19037v1 |
decod- ing with a batch size of 1. In reference systems, we concatenate the structured textual representation with the instruction directly within a conversation. To evaluate model responses, we use a rule-based processor [27] to automatically verify compliance with constraints in the closed-ended question and creative... | https://arxiv.org/abs/2505.19037v1 |
with formatting conventions, such as double quota- tion marks and JSON structure. This highlights a limitation of model design without leveraging an instruction-following LLM. Recently, models like BLSP-emo, DiV A, and DeSTA2 have adopted a similar training approach, using responses from their original LLM counterparts... | https://arxiv.org/abs/2505.19037v1 |
S. Altman, S. Anadkat et al. , “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 , 2023. [2] J. Bai, S. Bai, Y . Chu, Z. Cui, K. Dang, X. Deng, Y . Fan, W. Ge, Y . Han, F. Huang et al. , “Qwen technical report,” arXiv preprint arXiv:2309.16609 , 2023. [3] A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. ... | https://arxiv.org/abs/2505.19037v1 |
and adaptive speech large language model,” arXiv preprint arXiv:2404.00656 , 2024. [18] C.-Y . Kuan, C.-K. Yang, W.-P. Huang, K.-H. Lu, and H.-y. Lee, “Speech-copilot: Leveraging large language models for speech processing via task decomposition, modularization, and program generation,” in 2024 IEEE Spoken Language Tec... | https://arxiv.org/abs/2505.19037v1 |
Bangkok, Thailand: Associ- ation for Computational Linguistics, Aug. 2024. [30] B. Wang, X. Zou, G. Lin, S. Sun, Z. Liu, W. Zhang, Z. Liu, A. Aw, and N. F. Chen, “Audiobench: A universal benchmark for audio large language models,” NAACL , 2025. [31] C.-H. Chiang and H.-y. Lee, “Can large language models be an alternati... | https://arxiv.org/abs/2505.19037v1 |
arXiv:2505.19051v1 [cs.CL] 25 May 2025Efficient Data Selection at Scale via Influence Distillation Mahdi Nikdan∗† ISTA & Google ResearchVincent Cohen-Addad† Google Research Dan Alistarh ISTA & Red Hat AIVahab Mirrokni Google Research Abstract Effective data selection is critical for efficient training of modern Large L... | https://arxiv.org/abs/2505.19051v1 |
Uniform Mid Perplexity Influence Distillation Pareto Frontier Figure 1: Average improvement over uniform sam- pling across six tasks vs. runtime. The model used is Llama2-7B [Touvron et al., 2023], and the training dataset is Tulu V2 [Ivison et al., 2023]. The annotation “M/N” indicates that the method selected M sampl... | https://arxiv.org/abs/2505.19051v1 |
proposed clustering loss trajectories to identify representative training examples, though their approach focused more on general domain adaptation rather than specific target distributions. Another approach, so-called Classifier, was introduced by Brown et al. [2020] and has been employed in subsequent work (Gao et al... | https://arxiv.org/abs/2505.19051v1 |
et al. [2021], Wei et al. [2015], Chen et al. [2023], Cao et al. [2023], Sener and Savarese [2017] and references therein. 3 Method 3.1 Problem and Notation Letθ∈Rdbe the model parameters. For any dataset Dof size nand any vector of sample weights w= [w1, w2, ..., w n]T, denote L(θ;D,w) =1 nPn i=1wiℓ(θ;Di)as the weight... | https://arxiv.org/abs/2505.19051v1 |
T(θ)δ+1 2δTHT(θ)δwhere δis replaced with −η |S|GT S(θ)w. Next, we define two key objects, p∈R|S|andQ∈R|S|×|S|, as follows: p(θ;S, T) =GS(θ)gT(θ), (3) Q(θ;S, T) =1 |S|GS(θ)HT(θ)GT S(θ), (4) For brevity, unless stated otherwise, we will omit SandTfrom the arguments of pandQ. Let f(w;θ) =−p(θ)Tw+η 2wTQ(θ)w. (5) Then, the ... | https://arxiv.org/abs/2505.19051v1 |
(1−βs 1)(qv 1−βs 2+ϵ). Additionally, sis the number of warm-up steps, and (β1, β2, ϵ) are Adam hyperparameters. See Appendix C for more details. Handling Variable Lebel Lengths. A common practice in data selection is to normalize the gradients prior to measuring similarities [Xia et al., 2024]. This is motivated by the... | https://arxiv.org/abs/2505.19051v1 |
level, this method provides an efficient low-rank approximation of the gradient matrix GS, given by ˆGS=CGL, where GL∈Rℓ×dcontains the gradients of ℓ≪ |S| selected landmark samples. The matrix C∈R|S|×ℓholds the coefficients that express each sample’s gradient as a linear combination of the landmark gradients. Specifica... | https://arxiv.org/abs/2505.19051v1 |
these ℓlayers. The set Vcontains random Gaussian vectors matching the shape of θℓ, and the term∂Nℓ(x) ∂θℓis the Jacobian of Nℓ(x)with respect to θℓ. In words, JVP embeddings project the Jacobian of an intermediate model output onto a set of random directions in parameter space. 4.4 Tuning the Regularization Coefficient... | https://arxiv.org/abs/2505.19051v1 |
− Llama3.2-3BUniform 53.9 ±0.52 34.6 ±1.22 48.9 ±0.67 63.1 ±0.36 56.1 ±1.35 80.4 ±0.51 0.00 0 Mid-PPL 54.0±0.27 29.5±0.12 48.3 ±0.44 65.9±0.66 55.9±4.40 80.9 ±0.18 -0.42 1200 TF RDS+ 53.1 ±0.58 38.4±0.58 49.6 ±0.45 61.0±0.35 60.6±1.77 84.2 ±0.47 +1.65 1200 TF InfDist 54.0±0.94 35.7±1.28 48.6 ±0.27 64.6 ±1.29 55.4 ±1.10... | https://arxiv.org/abs/2505.19051v1 |
to stabilize gradients (see Appendix A). Gradients are projected to 131072 dimensions via Hadamard projections; we use the largest dimension that fits in GPU memory, as projection cost does not depend on the dimension (Appendix I). After selection, we do not incorporate the sample weights during training, as experiment... | https://arxiv.org/abs/2505.19051v1 |
robustness of Influence Distillation. 6 Limitations and Future Work Below, we outline three main limitations of our work, along with corresponding directions for future research. No Target Distribution. While we demonstrate that Influence Distillation is highly effective for targeted instruction tuning across a range o... | https://arxiv.org/abs/2505.19051v1 |
selection with datamodels, 2024. URL https://arxiv. org/abs/2401.12926 . Wei Huang, Yunxiao Zhang, Shangmin Guo, Yuming Shang, and Xiangling Fu. Dynimpt: A dynamic data selection method for improving model training efficiency. IEEE Transactions on Knowledge and Data Engineering , 2024. Hugo Touvron, Louis Martin, Kevin... | https://arxiv.org/abs/2505.19051v1 |
data. arXiv preprint arXiv:1709.02279 , 2017. Yukun Feng, Patrick Xia, Benjamin Van Durme, and João Sedoc. Automatic document selection for efficient encoder pretraining. arXiv preprint arXiv:2210.10951 , 2022. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, a... | https://arxiv.org/abs/2505.19051v1 |
Gommers, TE Oliphant, M Haberland, T Reddy, D Cournapeau, E Burovski, P Peterson, W Weckesser, J Bright, et al. Fundamental algorithms for scientific computing in python and scipy 1.0 contributors. scipy 1.0. Nat. Methods , 17:261–272, 2020. Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint ... | https://arxiv.org/abs/2505.19051v1 |
Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. Free dolly: Introducing the world’s first truly open instruction-tuned llm, 2023. Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. a... | https://arxiv.org/abs/2505.19051v1 |
as both necessary and sufficient. Similar plots for GSM8k [Cobbe et al., 2021] and SQuAD [Rajpurkar et al., 2016] are provided later in the Appendix (Figure 9). Additionally, for each dataset and checkpoint, we measure the Pearson product-moment correlation between gradient norms and the number of label tokens per samp... | https://arxiv.org/abs/2505.19051v1 |
0.96 0.96 0.96 0.96 0.96 0.96 0.96 0.37 0.76 0.86 0.92 0.94 0.96 0.97 0.97 1.00 0.98 0.98 0.97 0.97 0.97 0.97 0.97 0.97 0.36 0.74 0.84 0.91 0.92 0.95 0.95 0.97 0.98 1.00 0.98 0.98 0.98 0.98 0.98 0.98 0.98 0.36 0.74 0.84 0.90 0.92 0.94 0.95 0.96 0.98 0.98 1.00 0.99 0.99 0.98 0.98 0.98 0.98 0.35 0.73 0.83 0.89 0.92 0.94 ... | https://arxiv.org/abs/2505.19051v1 |
wTQ(θ+δ)Tw=wTGS(θ+δ)THTGS(θ+δ)w = (wTGS(θ)T+δTHw)HT(GS(θ)w+Hwδ) =wTQ(θ)w+ 2gw(θ)THTHwδ+δTHwHTHwδ (21) 16 0102030405060708090100110120130140150152 Step0.9 0.8 0.7 0.6 0.5 0.4 0.3 Correlation Tulu V2 GSM8k BBH SQuADFigure 5: Correlation between gradient norm and number of label tokens, across checkpoints on four datasets... | https://arxiv.org/abs/2505.19051v1 |
noting a property of the landmark-based approximation introduced in Section 4.3: it exhibits rotational equivariance . That is, if all source and target gradients are rotated by an orthonormal matrix, the resulting landmark-based gradient approximations will also be simply rotated by the same matrix. In the remainder o... | https://arxiv.org/abs/2505.19051v1 |
of the identity matrix. If this is not the case, we take advantage of Lemma D.2 and apply a change of variables: G←GR andt←Rt, where R=PD,Pis a permutation matrix, andDis a diagonal matrix with entries chosen uniformly at random from {±1}. Note that by the Rotation Equivariance assumption, this transformation implies ˆ... | https://arxiv.org/abs/2505.19051v1 |
question is considered. Big-Bench-Hard (MIT License). This dataset includes questions from 27 challenging tasks, such as causal judgment, multi-step arithmetic, and logic. Following Suzgun et al. [2022], we perform 3-shot evaluations using the same 3 samples per category (a total of 81) as the target set. TyDIQA (Apach... | https://arxiv.org/abs/2505.19051v1 |
this idealized setting quickly achieves high accuracy in gradient approximation—surpassing 0.9 cosine similarity with just over 4096 landmarks. This suggests that the gradients are approximately low-rank, a known phenomenon in LLMs [Hu et al., 2022, Zhao et al., 2024]. End-to-end Selection and Training. We repeat the s... | https://arxiv.org/abs/2505.19051v1 |
BR−1 BBpB 1T BR−1 BB1B)1B . (50) For indices in the active set Awe have w∗ A=0, giving the final candidate solution w∗= (w∗ A,w∗ B). Optimality requires that the remaining Karush–Kuhn–Tucker (KKT) conditions hold, namely ∀i∈ B,wi≥0(primal feasibility) and ∀j∈A,αj≥0(dual feasibility). Because the objective is convex ( ... | https://arxiv.org/abs/2505.19051v1 |
the target dataset. We compute the vectors pand the matrices Qexactly for the Qwen-2.5 1.5B model [Team, 2024], using Hessian-vector products to obtain Q. We then evaluate both T1andT2using default weights w=1and a range of learning rates. To measure the relative contribution of the second-order term, we report the rat... | https://arxiv.org/abs/2505.19051v1 |
Distillation into the training loss. Specifically, we conduct an experiment using LLaMA-2 7B [Touvron et al., 2023], with a pool size of 200k and 8192 landmarks sampled from Tulu V2 [Ivison et al., 2023]. During training, we scale the loss of each selected sample by its corresponding weight. Table 3 compares this weigh... | https://arxiv.org/abs/2505.19051v1 |
0.56 0.56 0.57 0.14 0.66 0.89 1.00 0.94 0.87 0.79 0.78 0.74 0.70 0.71 0.69 0.70 0.68 0.68 0.69 0.69 0.15 0.67 0.85 0.94 1.00 0.87 0.79 0.76 0.72 0.71 0.71 0.68 0.68 0.67 0.67 0.67 0.68 0.14 0.59 0.74 0.87 0.87 1.00 0.94 0.93 0.90 0.87 0.87 0.87 0.87 0.86 0.85 0.86 0.87 0.13 0.55 0.68 0.79 0.79 0.94 1.00 0.97 0.94 0.90 ... | https://arxiv.org/abs/2505.19051v1 |
arXiv:2505.19056v1 [cs.CL] 25 May 2025An Embarrassingly Simple Defense Against LLM Abliteration Attacks Harethah Abu Shairah†Hasan Abed Al Kader Hammoud† Bernard Ghanem†George Turkiyyah† †King Abdullah University of Science and Technology (KAUST) Abstract Large language models (LLMs) are typically aligned to comply wit... | https://arxiv.org/abs/2505.19056v1 |
aim to disperse the safety signal across multiple dimensions in the representation space, making it substantially harder to isolate and remove. We validate our approach by fine-tuning three open-weight LLMs ( LLAMA -2-7B-C HAT, QWEN 2.5-3B-I NSTRUCT , and QWEN 2.5-1.5B- INSTRUCT ) on our extended-refusal dataset and su... | https://arxiv.org/abs/2505.19056v1 |
jail- breaks. Rosati et al. (2024) introduced a framework that validates the protection against harmful fine-tuning, while Bhardwaj et al. (2024) demonstrated that lost safety from benign fine-tuning can be restored via a simple addition of a safety vector to the fine-tuned model. Qi et al. (2024) shows how alignment i... | https://arxiv.org/abs/2505.19056v1 |
To diffuse theirlatent signature, we construct an Extended Re- fusal (ER) dataset DER. Starting from 4,289 harm- ful prompts merged from Beavertails (Ji et al., 2023), AdvBench (Zou et al., 2023), StrongReject (Souly et al., 2024), and TDC-2023 (Mazeika et al., 2023), we generate structured refusals using GPT- 4O(OpenA... | https://arxiv.org/abs/2505.19056v1 |
sponses that remain lexically and semantically consistent and well-formed, evaluated using an LLM-based judge (Q WEN -2.5-14B). •MMLU Score (%): Performance on the Massive Multitask Language Understanding benchmark, measuring knowledge and reason- ing capabilities. •Perplexity : Measured on C4 passages, with lower valu... | https://arxiv.org/abs/2505.19056v1 |
base models which remain fully coherent. This supports our hypothesis that extended refusals create a more diffused set of representations that cannot be eliminated without affecting other aspects of model behavior. 4.3 Abliterated Models Analysis To have a complete idea of how abliterating a model using different refu... | https://arxiv.org/abs/2505.19056v1 |
QWEN 2.5-1.5B-E XTENDED 75.02 64.98 QWEN 2.5-3B 100.88 66.99 QWEN 2.5-3B-E XTENDED 79.59 71.94 LLAMA2-7B 112.01 83.28 LLAMA2-7B-E XTENDED 74.64 60.99 nal in a single latent direction that attackers can easily target. Our insight is that how a model refuses directly impacts how secure that refusal is. By training models... | https://arxiv.org/abs/2505.19056v1 |
Won Chung, Sam Toyer, Johannes Heidecke, Alex Beutel, and Amelia Glaese. 2024. Deliberative Alignment: Reasoning Enables Safer Language Models. arXiv . Xingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin, and Bin Hu. 2024. Cold-attack: Jailbreaking llms with stealthiness and controllability. Preprint , arXiv:2402.08679. Ca... | https://arxiv.org/abs/2505.19056v1 |
P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander M ˛ adry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024. Gpt-4o system card. Preprint , arXiv:2410.21276. Long Ouyang, Jeff Wu, Xu Jiang,... | https://arxiv.org/abs/2505.19056v1 |
Pieter Abbeel, Justin Svegliato, Scott Emmons, Olivia Watkins, and Sam Toyer. 2024. A strongREJECT for empty jail- breaks. In NeurIPS . Yihong Tang, Bo Wang, Xu Wang, Dongming Zhao, Jing Liu, Ruifang He, and Yuexian Hou. 2025. Role- Break: Character hallucination as a jailbreak attack in role-playing systems. In Procee... | https://arxiv.org/abs/2505.19056v1 |
points. For LLAMA -2-7B- EXTENDED -REFUSAL -ABLITERATED , Qwen-2.5-14B judges a 92.7% refusal rate versus 90.4% under Llama-Guard. Similarly, QWEN 2.5-3B- EXTENDED -REFUSAL -ABLITERATED shows 90.9% (LLM judge) vs. 94.6% (Llama-Guard), and QWEN 2.5-1.5B- EXTENDED -REFUSAL -ABLITERATED yields 96.7% vs. 89.3%. These resul... | https://arxiv.org/abs/2505.19056v1 |
arXiv:2505.19060v1 [cs.CL] 25 May 2025UNCERTAINTY -LINE: Length-Invariant Estimation of Uncertainty for Large Language Models Roman Vashurin Maiya Goloburda Preslav Nakov Maxim Panov Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) {Roman.Vashurin, Maiya.Goloburda, Maxim.Panov}@mbzuai.ac.ae Abstract Lar... | https://arxiv.org/abs/2505.19060v1 |
has clear length bias, even when mea- sured by methods that are nominally length- adjusted; see Section 3. •We propose Uncertainty- Length INvariant Estimation ( UNCERTAINTY -LINE1), a sim- ple and unsupervised debiasing approach that fits a regression between uncertainty scores and output length, and uses the residual... | https://arxiv.org/abs/2505.19060v1 |
these methods rely on sampling multiple outputs from the language model to capture the distribu- tion of possible generations. This sampling process is computationally intensive, especially for large models. Secondly, implementations often depend on pre-trained models, to assess semantic similar- ity between generated ... | https://arxiv.org/abs/2505.19060v1 |
of the obtained linear trend was assessed using the Wald test and corresponding p-value was calculated along with the slope. Uncertainty Measures are Strongly Length- Dependent. Figure 1 presents results on one of the machine translation datasets (WMT14 De-En), XSum and GSM8K. UQ measures show clear and significant tre... | https://arxiv.org/abs/2505.19060v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.