text string | source string |
|---|---|
to 1/|y|, ensuring each response’s total weight sums to 1. SamPO employs a more subtle downsampling on the longer response to only consider the same amount of random tokens as the other response. LDDPO only down-weights the over-lengthy parts to reduce δ. These methods apply some heuristic kind of weighting scheme to m... | https://arxiv.org/abs/2505.18720v1 |
“France’s big city? That’s Paris, I guess. The birthplace of Victor Hugo, Émile Zola, Charles Baudelaire. ” Building on the token-pair level transformation of the DPO loss, we now focus on exploiting this finer granular ∆rto differentiate token pairs based on their semantic relevance. We aim to empha- size semantically... | https://arxiv.org/abs/2505.18720v1 |
under an optimal transport framework, whereas the first term corresponds to the Wasserstein distance be- tween the two responses’ distribution. After solving for the optimal transport plan Γ∗, the token-level weights ω∗ candω∗ rare obtained by summing along the respective dimensions as in Eq. 7 and normalizing to a pre... | https://arxiv.org/abs/2505.18720v1 |
to most of all preference optimization methods. Please re- fer to Appx. A for more detailed descriptions. Baselines. We primarily compare OTPO with DPO (Rafailov et al., 2024b) and other direct pref- erence optimization methods, excluding RLHF approaches that require training an additional reward model, following prior... | https://arxiv.org/abs/2505.18720v1 |
in 3 out of 4 settings, with up to 3.5% increase compared to the best baseline, demonstrating its robust effective- 0 50 100 150 200 250 Number of CasesDPOSimPOSamPOLDDPOOTPO 54.7%55.5%63.3%44.4%71.9%Comparison of Win/Tie/Lose Cases Across Methods Win Tie LoseFigure 4: TL;DR summarization win rates compared to the base... | https://arxiv.org/abs/2505.18720v1 |
WR Length Initial 22.92 22.57 1899 DPO 48.14 51.52 2168 OTPO 53.37 47.58 1791 (1) Ablation of Optimal Transport Uniform 52.60 46.36 1796 Similarity 53.28 46.09 1757 (2) Ablation of Weight Normalization None 26.38 26.07 1939 Mean 52.79 46.69 1791 Max 49.85 44.77 1808 Length 48.51 52.12 2167 Table 3: Ablation study of OT... | https://arxiv.org/abs/2505.18720v1 |
fluctuations in total weight |Γ|, thus leading to too aggressive gradient updates. Using mean or max for normalization leads to a decrease in both length-controlled win rate and win rate. This may relate to the large fluctuations in the shorter response’s weight upscale across samples, which can lead to suboptimal perf... | https://arxiv.org/abs/2505.18720v1 |
Time includes time for hyperparameter tuning based on DPO’s optimal settings. Empirical Efficiency. As shown in Fig. 6, OTPO exhibits training efficiency comparable to existing preference optimization methods. Despite the slight increase in per-run cost due to the OT computation, OTPO remains favorable in terms of tota... | https://arxiv.org/abs/2505.18720v1 |
direct Preference Optimization (OTPO), a context-aware token weighting scheme to reinforce semantically meaningful differences in reward estimation. OTPO leverages optimal transport to dynamically assign a fixed total weight budget to each token pair in the chosen and rejected response based on their semantic similarit... | https://arxiv.org/abs/2505.18720v1 |
building world-class universities (disci- plines) of Renmin University of China and Public Computing Cloud, Renmin University of China. References Martin Arjovsky, Soumith Chintala, and Léon Bottou. 2017. Wasserstein generative adversarial networks. InInternational conference on machine learning , pages 214–223. PMLR. ... | https://arxiv.org/abs/2505.18720v1 |
Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, An- ish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2024. A framework for few-shot language model evaluation. Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bi- lal Piot, Remi Munos, Mark Rowland, Michal... | https://arxiv.org/abs/2505.18720v1 |
a reference-free reward. In Advances in Neural In- formation Processing Systems , volume 37, pages 124198–124235. Curran Associates, Inc. Eric Mitchell. 2023. A note on dpo with noisy prefer- ences & relationship to ipo. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, ... | https://arxiv.org/abs/2505.18720v1 |
10592, Miami, Florida, USA. Association for Com- putational Linguistics. Jun Wang, Bohan Lei, Liya Ding, Xiaoyin Xu, Xian- feng Gu, and Min Zhang. 2024b. Autoencoder-based conditional optimal transport generative adversarial network for medical image generation. Visual Infor- matics , 8(1):15–25. Peiyi Wang, Lei Li, Li... | https://arxiv.org/abs/2505.18720v1 |
Advances in Neural Information Processing Systems , 36:46595–46623. A Implementation Details We use UltraFeedback (Cui et al., 2024), the preference-transformed HelpSteer2 (Wang et al.,2025), and TL;DR (Stiennon et al., 2020) as pref- erence training datasets. UltraFeedback contains 61,135 examples in its training spli... | https://arxiv.org/abs/2505.18720v1 |
Llama-3.2-3B UF HS UF HS count 59876 9084 60692 9087 mean -29 -24 -91 -75 std 233 168 447 257 min -4058 -1854 -4094 -2016 25% -67 -75 -99 -115 50% -11 -14 -16 -30 75% 31 37 30 28 max 2218 1193 4079 1266 Table 6: Response length difference summary statistic of each on-policy dataset. UF denotes UltraFeedBack, while HS d... | https://arxiv.org/abs/2505.18720v1 |
subsets. We select the model’s answer based on the probabilities of ‘A’, ‘B’, ‘C’, and ‘D’, as suggested in the original paper, and report the overall accuracy. •GSM8K (Cobbe et al., 2021): A mathemati- cal benchmark of grade school math problems for evaluating reasoning abilities. We evalu- ate the original test split... | https://arxiv.org/abs/2505.18720v1 |
loss, enhancing ro- bustness to preference label noise. •EXO (Ji et al., 2024) replaces the forward KL with reverse KL when deriving DPO loss. •IPO (Gheshlaghi Azar et al., 2024) minimizes the squared loss of margin between the esti- mated reward margin and a predefined margin. •NCA (Chen et al., 2024) optimizes the ab... | https://arxiv.org/abs/2505.18720v1 |
token weight difference. Differing changes of ϵ1, ϵ2on normalization value τ.A larger ϵ2decreases τby imposing stronger penalties on marginal differences, whereas increasing ϵ1implicitly downweights the impact of marginal differences overall, leading to a higher τ. This interplay reflects the contrasting roles of these... | https://arxiv.org/abs/2505.18720v1 |
a node, with tokens from the chosen response on the left and those from the rejected response on the right. Each node is labeled with its position in the pairwise data, its token text, and its aggregate weight ωi ∗. Token positions are indicated using response codes (C for chosen and R for rejected) and their indices, ... | https://arxiv.org/abs/2505.18720v1 |
sentence. This progres- sion is evident in the intermediate-sized bars and connections associated with these tokens. As the re- sponses diverge semantically, the weights of unique tokens progressively diminish ( <0.7). Tokens like “known for its art” or “the birthplace of Victor Hugo” cannot be well-mapped to correspon... | https://arxiv.org/abs/2505.18720v1 |
agree- ment with OT. This implies that the middle layers are critical in bridging the gap between token-level contributions and global model explanations. J Evaluation Configuration Details J.1 Summarization Win Rate Calculation In this section, we include the details for GPT-4o to generate win rates for summarization.... | https://arxiv.org/abs/2505.18720v1 |
Human Evaluation Guideline. K Mathematical Derivations K.1 Token-level DPO loss We can transform a response y’s probability π(y|x)given x as follows: π(y|x) =|y|Y i=1π(yi|y<i, x) = exp(log|y|Y i=1π(yi|y<i, x)) = exp(|y|X i=1logπ(yi|y<i, x)) (13) For the reward difference term in Eq. 1: ∆r= logπθ(yc|x) πref(yc|x)−logπθ(... | https://arxiv.org/abs/2505.18720v1 |
arXiv:2505.18722v1 [eess.AS] 24 May 2025 This paper is accepted (camera-ready) for Interspeech 2025Evaluating the Usefulness of Non-Diagnostic Speech Data for Developing Parkinson’s Disease Classifiers Terry Yi Zhong, Esther Janse, Cristian Tejedor-Garcia, Louis ten Bosch, Martha Larson Centre for Language Studies, Rad... | https://arxiv.org/abs/2505.18722v1 |
is recorded in controlled settings, such as laboratories or clinical environments, for research or treatment purposes, but not specifically for PD diagnosis. For instance, the Turn-Taking (TT) dataset that we introduce in this work, originally developed for studying conversational turn-taking phenomena [14] as a functi... | https://arxiv.org/abs/2505.18722v1 |
of our knowledge, no prior study has explored the diagnostic potential of mid-level diagnostic-purpose PD datasets, such as the TT dataset. However, understanding the value of these datasets is crucial, as they offer the potential to allow the community to leverage a broader range of data. 2.2. TT Dataset (Mid-Level Di... | https://arxiv.org/abs/2505.18722v1 |
speech corpus specifically designed for PD diagnosis, we selected PC-GITA [16]. Since its release, it has been extensively used as a dataset for developing and validating PD classification methods [7, 9, 21, 13, 16]. This dataset includes 100 participants: 50 PD patients and 50 HC. Each group has 25 male and 25 female ... | https://arxiv.org/abs/2505.18722v1 |
commonly used in PD diagnostic research, whereas TT includes specialized tasks designed for other research purposes. 3. Method 3.1. Classification Model We utilized a machine learning classification approach, as motivated by recent advances [6, 7, 23], to explore the potential of PD diagnosis using TT. Based on a liter... | https://arxiv.org/abs/2505.18722v1 |
of this range ( RQ1). However, in order to reach this level, it is necessary to concatenate the TT data and also make sure that it is appropriately balanced ( RQ2). We recommend that when using non-diagnostic speech data, researchers should not forget that these are key design decisions to make concerning how a dataset... | https://arxiv.org/abs/2505.18722v1 |
and tested on TT model on only the BoundaryTone task of the TT. Here, we also observed that performance was uneven across speakers. We further analyzed the performance of individual participants from the TT-concat-balanced dataset. The three participants with the lowest accuracy (<20%) are all female (2 PD, 1 HC), as s... | https://arxiv.org/abs/2505.18722v1 |
work was conducted in close collaboration with the project “Turn-taking in Dialogue in Populations with Communicative Impairment" (https://www.ru.nl/en/research/research-projects/turntaking- in-dialogue-in-populations-with-communicative-impairment). This work used the Dutch national e-infrastructure with the support of... | https://arxiv.org/abs/2505.18722v1 |
pp. 47–59, 2024.[12] K. Svensberg, B. G. Kalleberg, L. Mathiesen, Y . Andersson, S. E. Rognan, and S. K. Sporrong, “The observer effect in a hospital setting - experiences from the observed and the observers,” Research in Social and Administrative Pharmacy , vol. 17, no. 12, pp. 2136–2144, 2021. [13] A. Favaro, A. Buta... | https://arxiv.org/abs/2505.18722v1 |
arXiv:2505.18744v1 [cs.CL] 24 May 2025LogicCat: A Chain-of-Thought Text-to-SQL Benchmark for Multi-Domain Reasoning Challenges Tao Liu1, Hongying Zan1, Yifan Li1, Dixuan Zhang1, Lulu Kong1,Haixin Liu1,Jiaming Hou1,Aoze Zheng1, Rui Li1,Yiming Qiao1,Zewei Luo1,Qi Wang1, Zhiqiang Zhang1,Jiaxi Li1,Supeng Liu1,Kunli Zhang1,... | https://arxiv.org/abs/2505.18744v1 |
implicit data compu- tations. Spider 2.0 introduces multi-step enterprise workflows and complex nested queries, further challenging LLMs. However, its fixed database de- sign limits cross-database generalization, with per- formance slightly below Spider 1.0. These datasets collectively highlight a gap: while they chall... | https://arxiv.org/abs/2505.18744v1 |
2B, questions in this category are characterized by a high density of computational steps, distin- guishing them from physical knowledge reason- ing. Notably, every question-SQL pair in LogicCat incorporates arithmetic reasoning. Additionally, while the subsequent sections on commonsense and hypothetical reasoning also... | https://arxiv.org/abs/2505.18744v1 |
- strftime ( '%Y' , '2001 -01- 01') - (strftime ( '%m-%d', date_became_customer) < strftime ( '%m-%d', '2001 -01-01')) AS age FROM Customers WHERE first_name = 'Beijing ' Figure 2: Overview of Reasoning Categories in LogicCat. analyses are provided in Appendix B. 2.4 Ideal Hypothetical Reasoning Ideal hypothetical reas... | https://arxiv.org/abs/2505.18744v1 |
hypothetical reasoning, with both realistic and exaggerated scenarios (e.g., setting outdoor temperature to an extreme value as in Fig- ure 3); and (5) Questions were encouraged to re- quire advanced SQL syntax such as GROUP BY , ORDER BY , and JOIN . All questions were initially written in Chinese, specifying reasonin... | https://arxiv.org/abs/2505.18744v1 |
FormulaQuestion: Give me IDs of all customers ... Table Info Text LLM Step1 Step2 Step3 Step1: Calculate age for non -Beijing customers...4.SQL Annotation Ensuring Clarity Verifying Formulas Writing SQL Queries Validating SQL Q ueries SELECT customer_id, strftime ( '%M', date_became_customer) - strftime ( '%D', date_of... | https://arxiv.org/abs/2505.18744v1 |
BIRD and Spider 2.0 introduce a substan- tial amount of value prediction, where values are explicitly mentioned in the questions. In Logic- Cat, each SQL query references an average of 2.47 tables, indicating that a significant proportion of questions require information from multiple tables for SQL derivation. Moreove... | https://arxiv.org/abs/2505.18744v1 |
models were tested with a tempera- ture of 0.0, and inputs exceeding the model’s to- ken limit were truncated from the beginning. As shown in Figure ( ?), injecting chain-of-thought (CoT) reasoning into prompts led to substantial per- formance gains across all difficulty levels for both VES and EX metrics. Claude 3.7-S... | https://arxiv.org/abs/2505.18744v1 |
39.89 45.21 45.16 41.32 43.11 42.98 44.56 45.91 48.92 47.12 48.13 47.89 48.74 44.59 44.18 32.83 33.25 33.42 35.10 36.36 22.92 35.66 33.75 27.81 31.21 29.31 36.11 37.25 39.39 40.10 40.02 39.02 37.28 33.52 32.12 14.10 13.91 14.97 14.88 15.01 12.68 15.01 15.21 14.79 12.61 13.16 14.02 15.91 18.21 18.00 17.95 16.01 16.21 15... | https://arxiv.org/abs/2505.18744v1 |
67.78 62.88 68.83 61.27 60.78 66.31 61.21 67.32 60.91 66.86 67.69 68.59 69.59 69.67 69.69 66.23 67.77 66.38 22.75 17.54 21.40 11.29 19.65 3.61 12.38 27.16 5.21 14.26 7.22 21.35 22.32 26.52 27.43 26.92 24.89 23.54 21.98 11.99 70.35 68.58 70.31 66.59 69.76 65.24 66.59 70.09 68.26 69.36 66.16 70.79 71.87 74.17 75.01 74.92... | https://arxiv.org/abs/2505.18744v1 |
human reason- ing. 7 Conclusion We present LogicCat, a benchmark that inte- grates complex reasoning and chain-of-thought analysis for SQL inference. LogicCat features a di- verse set of tasks encompassing rich physical back- grounds, physics and mathematical formulas, and complex database schemas, covering a broad ran... | https://arxiv.org/abs/2505.18744v1 |
Chengda Lu. 2025b. Deepseek-v3 technical report. Preprint , arXiv:2412.19437. Longxu Dou, Yan Gao, Xuqi Liu, Mingyang Pan, Dingzirui Wang, Wanxiang Che, Dechen Zhan, Min-Yen Kan, and Jian-Guang Lou. 2023. To- wards knowledge-intensive text-to-sql semantic parsing with formulaic knowledge. Preprint , arXiv:2301.01067. D... | https://arxiv.org/abs/2505.18744v1 |
2024. Gpt-4 technical report. Preprint , arXiv:2303.08774.Mohammadreza Pourreza, Hailong Li, Ruoxi Sun, Yeounoh Chung, Shayan Talaei, Gaurav Tarlok Kakkar, Yu Gan, Amin Saberi, Fatma Ozcan, and Sercan O Arik. 2025a. CHASE-SQL: Multi-path reasoning and preference optimized candidate selec- tion in text-to-SQL. In The Th... | https://arxiv.org/abs/2505.18744v1 |
Zhang, Caiming Xiong, Richard Socher, Walter Lasecki, and Dragomir Radev. 2019a. CoSQL: A conversational text-to-SQL challenge towards cross- domain natural language interfaces to databases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Confere... | https://arxiv.org/abs/2505.18744v1 |
since these elements are not formally defined or derived from previous steps. Such errors highlight the model’s difficulty in preserving variable scope and logical consis- 11 tency across chained query components. These issues underscore the persistent challenge of gener- ating coherent, syntactically valid, and execut... | https://arxiv.org/abs/2505.18744v1 |
models, this time incorporating our annotated chain-of-thought (CoT) steps into the prompts. As shown in Figure 7, our ex- periments revealed that, without injecting chain- of-thought (CoT) content into the prompts, o4- mini achieved the best performance with an EX of 14.96% and VES of 75.01%, slightly outper- forming ... | https://arxiv.org/abs/2505.18744v1 |
day of the 21st Century.” Solving this problem involves a sequence of reasoning steps: (1) calculating the age for non-Beijing customers by subtracting the birth year from the registration year, with an adjust- ment for unpassed birthdays using a DATE_FORMAT comparison; (2) handling Beijing customers as a special case ... | https://arxiv.org/abs/2505.18744v1 |
and validate the necessary formulas. In parallel, ten graduate students specializing in software engi- neering—each with strong backend development experience—were invited to participate in the cre- ation of the chain-of-thought (CoT) reasoning steps. Leveraging their expertise in formula manipula- tion and SQL, these ... | https://arxiv.org/abs/2505.18744v1 |
retaining detailed data for each computer. step3: Deduplicate and sort the output. Apply DISTINCT to remove duplicates from intermediate results containing repeated calculations, and sort the final output by battery health percentage in ascending order. Table 4: Example of Physical Konwledge Reasoning category and deta... | https://arxiv.org/abs/2505.18744v1 |
Logic Reasoning category and detailed cases. 17 Reasoning Category Example Ideal Hypothetical Reasoning db_id: air_conditioner question: If an air conditioner has a cooling capacity of 10,000 BTU, assuming in extreme conditions where the outdoor temperature reaches 60 °C and the indoor temperature is set to -10 °C, cal... | https://arxiv.org/abs/2505.18744v1 |
-maintenanceDate -maintenanceType -…MaintenanceRecordsEnergyConsumption -consumptionId -deviceId -timestamp -energyConsumed -…EnergyManagement Database InfoFigure 9: Information on EnergeManagement databases. 20 Model Name Size Base Model Qwen2.5-Coder-32B-Instruct (Hui et al., 2024) 32B Qwen2.5-32B Qwen2.5-72B-Instruc... | https://arxiv.org/abs/2505.18744v1 |
Unifying Attention Heads and Task Vectors via Hidden State Geometry in In-Context Learning Haolin Yang1Hakaze Cho2Yiqiao Zhong3Naoya Inoue2,4 1University of Chicago2JAIST3University of Wisconsin - Madison4RIKEN haolinyang2001@uchicago.edu, yfzhao@jaist.ac.jp yiqiao.zhong@wisc.edu, naoya-i@jaist.ac.jp Abstract The unusu... | https://arxiv.org/abs/2505.18752v1 |
ers (PTH) LM Heads Positive★ ★Separ ability(A) (B) LM Lay ers (IH)(C) ★ ★ Unembedding Direction Optimal Decoding(D) ★★ ☆☆ Unembedding Direction Task V ector = negativ e samples = positiv e samplesAlignmentFigure 1: (A)An example for ICL input. (B)In early layers, LLMs promote separability among the last tokens’ hidden ... | https://arxiv.org/abs/2505.18752v1 |
position—in early layers. (2)In middle-to-late layers, Induction Heads (IHs) enhance alignment by amplifying hidden state components along the unembedding directions of the labels. Thus, our findings not only explain why attention heads like PTHs and IHs are essential for ICL performance, but also clarify why induction... | https://arxiv.org/abs/2505.18752v1 |
the classification accuracy depends on two geometric properties of the hidden state collection HHH= [hhhi]n i=1∈Rn×d:(1)its separability, and (2)its alignment with the unembedding vectors. As mentioned before, an LLM performs classification using the unembedding matrix EEE, which is a multiclass classifier on the hidde... | https://arxiv.org/abs/2505.18752v1 |
maximum separability requires evaluating separability along infinitely many direc- tions in Sd−1and is intractable. As a practical proxy, we train a logistic classifier on a subset of HHH and evaluate its accuracy on held-out data. We call the accuracy separability score as it reflects the empirical maximum separabilit... | https://arxiv.org/abs/2505.18752v1 |
yAandyB—ontoEEEyA−EEEyB ∥EEEyA−EEEyB∥2and divide it by the pooled projected variance. A higher value implies better separation margin alongEEEyA−EEEyB ∥EEEyA−EEEyB∥2and better directional alignment. •Composite alignment (B.5) . To unify the mean- and variance-based alignment, we define composite alignment as their prod... | https://arxiv.org/abs/2505.18752v1 |
24. Accuracy improvements of increasing demonstrations are reflected by consistently improving alignment measures. (B)Changing the demonstration selection method to kNN retrieval preserves phase transition and improves accuracy through enhancing alignment. (C)Using uninformative demonstration labels hurts accuracy due ... | https://arxiv.org/abs/2505.18752v1 |
@ and #. Figure 3 reports the results averaged across datasets. Phase transition is a robust hallmark of ICL Figure 3 shows that the phase transition pattern persists across all three settings. In Figure 3 (A), it is evident at all demonstration counts except 0. Figure 3 (B) confirms that principled demonstration selec... | https://arxiv.org/abs/2505.18752v1 |
This suggests that as alignment improves, label-relevant semantics are injected into hidden states via consec- utive layer updates, enabling correct label decoding even at intermediate layers. To support this, we decode the top right singular vectors of hidden states near the transition point. Figure 4 (B) reveals a sh... | https://arxiv.org/abs/2505.18752v1 |
hidden states in middle-to-late layers as task vectors [Hendel et al., 2023]. Due to their improved alignment properties, steering or replacing zero-shot hidden states with ICL hidden states can align zero-shot hidden states better with the label unembedding vectors. Similarly, modifying zero-shot hidden states with IC... | https://arxiv.org/abs/2505.18752v1 |
accuracy. 0-shot TV ICL020406080Accuracy (%) 0.2465.1180.57 Figure 7: Average effect across datasets of steering zero-shot hidden states using task vectors created from IH outputs.These results not only support the presence of a two-stage ICL pro- cess—initially driven by separability, then by alignment—but also clarif... | https://arxiv.org/abs/2505.18752v1 |
J. Wu, C. Winter, C. Hesse, et al. Language models are few-shot learners. In Proceedings of the 34th International Conference on Neural Information Processing Systems , page 1877–1901, Online and Vancouver, Canada, 2020. URL https://dl.acm.org/doi/abs/ 10.5555/3495724.3495883 . S. Chan, A. Santoro, A. Lampinen, J. Wang... | https://arxiv.org/abs/2505.18752v1 |
L. Chiruzzo, A. Ritter, and L. Wang, editors, Findings of the Association for Computational Linguistics: NAACL 2025 , pages 1855–1888, Albuquerque, New Mexico, Apr. 2025. Association for Computational Linguistics. ISBN 979-8- 89176-195-7. URL https://aclanthology.org/2025.findings-naacl.100/ . A. Lee, X. Bai, I. Pres, ... | https://arxiv.org/abs/2505.18752v1 |
Brucher, M. Per- rot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learn- ing Research , 12:2825–2830, 2011. URL https://www.jmlr.org/papers/volume12/ pedregosa11a/pedregosa11a.pdf . 12 A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever. Language models are unsupervised ... | https://arxiv.org/abs/2505.18752v1 |
of the cells, and the supremum (or the maximum) value S∗is attained on one of the cells. We then show the proof of Theorem 1: Proof First, consider the case where Vonly consists of yAandyB. Then EEE∈R2×dand the two rows areEEEyAandEEEyB. Then we have Acc=1 nn ∑ i=11(yi=argmax vjEEEvjhhhi ∑ v′ j∈{yA,yB}EEEv′ jhhhi) (3) ... | https://arxiv.org/abs/2505.18752v1 |
∑ i=11(yi=argmax vjEEEvjhhhi ∑ j′∈1,...,|V|EEEv′ jhhhi) ≤1 nn ∑ i=11(yi=argmax vjEEEvjhhhi ∑ v′ j∈{y1,...,yk}EEEv′ jhhhi) =1 n(m ∑ k=1∑ xi∈Nk1(EEEykhhhi≥EEEyjhhhi,∀j̸=k,j∈1,...,m)) =1 n(m ∑ k=1∑ xi∈Nkmin j̸=k,j∈1,...,m1(EEEykhhhi≥EEEyjhhhi)) indicator function is either 1 or 0 ≤1 n(m ∑ k=1∑ xi∈Nk1 m−1∑ j,j̸=k,j∈1,...,m... | https://arxiv.org/abs/2505.18752v1 |
average of the variances of the two label clusters alongEEEyA−EEEyB ∥EEEyA−EEEyB∥2, i.e. |NA|Var({hhhT iEEEyA−EEEyB ∥EEEyA−EEEyB∥2:i∈NA})+|NB|Var({hhhT iEEEyA−EEEyB ∥EEEyA−EEEyB∥2:i∈NB}) |NA|+|NB|. Finally, divide the projected difference between means by the square root of the weighted average of the projected varianc... | https://arxiv.org/abs/2505.18752v1 |
symbols. E Supplementary materials for subsection 5.1 E.1 Replication of Figure 2 for other models In Figure 8-13, we provide the visualizations of the experiments presented in Figure 2 for other models, from which similar conclusions to those in subsection 5.1 can be drawn. Alignment measures explain the accuracy diff... | https://arxiv.org/abs/2505.18752v1 |
semantically relevant task tokens, marking the onset of alignment between the hidden states and the unembedding vectors of the symbols. This delayed alignment corresponds with the pattern seen in subplot (C) of Figure 3, where the phase transition under demonstrations with symbolized labels is postponed to later layers... | https://arxiv.org/abs/2505.18752v1 |
details concerning the identification of PTHs and IHs For each dataset, we use the first 50 queries to identify the set of IHs and PTH s. Denote the queries asx1,...,x50each with token length s(x1),...,s(x50). Identification of PTHs For each xi, the LLM will generate an attention tensor AAAttttttnnni∈ RL×Nh×s(xi)×s(xi)... | https://arxiv.org/abs/2505.18752v1 |
a sentence in the format “{Famous Person Name} is a:", and the label is the profession of the famous person. Once combined, they form an input sequence like, e.g., “Taylor Swift is a: Singer." that can be used as either a query or a demonstration. For each query, we sample k=4 demonstrations, ensuring that none share t... | https://arxiv.org/abs/2505.18752v1 |
impacts Understanding the internal mechanisms of large language models is increasingly critical as these models are deployed in high-stakes applications ranging from education and healthcare to legal and governmental decision-making. This work contributes to that understanding by providing a geometric framework that ex... | https://arxiv.org/abs/2505.18752v1 |
Number0.000.02 Composite alignment 0 10 20 30 Layer Number012 Mean-based alignment 0 10 20 30 Layer Number0.000.01 Var.-based alignment 8-shot 0-shotLlama2-7B: 8-shot Acc.: 75.13%; 0-shot Acc.: 0.31% Figure 10: Comparison of trends in separability and alignment measures of Llama2-7B hidden states between ICL and zero-s... | https://arxiv.org/abs/2505.18752v1 |
0 50 Layer Number0.000.02 Composite alignment 0 50 Layer Number02 Mean-based alignment 0 50 Layer Number0.0000.005 Var.-based alignmentLlama3-70B: Random Selection Acc.: 81.09%; KNN Selection Acc.: 89.44% Random Selection KNN Selection (B) KNN Demonstration selection 0 50 Layer Number02040 Effective dimension 0 50 Laye... | https://arxiv.org/abs/2505.18752v1 |
datasets(A) Varying number of demonstrations 0 10 Layer Number020 Effective dimension 0 10 Layer Number0.60.8 Separability score 0 10 Layer Number0.10.20.3 Singular alignment 0 10 Layer Number0.0000.0250.050 Composite alignment 0 10 Layer Number012 Mean-based alignment 0 10 Layer Number0.000.010.02 Var.-based alignment... | https://arxiv.org/abs/2505.18752v1 |
10 20 30 Layer Number0.00.20.40.60.8Output Alignment 0.000.050.100.150.200.250.30 Singular Alignment (D) RTE 0 10 20 30 Layer Number0.00.20.40.60.8Output Alignment 0.000.010.020.03 Composite Alignment 0 10 20 30 Layer Number0.00.20.40.60.8Output Alignment 0.000.050.100.150.200.25 Singular Alignment (E) CB 0 10 20 30 La... | https://arxiv.org/abs/2505.18752v1 |
ICL hidden states with label unembedding vectors on various datasets. 32 0 25 50 75 Layer Number0.00.20.40.60.8Output Alignment 0.0000.0020.0040.0060.0080.010 Composite Alignment 0 20 40 60 80 Layer Number0.00.20.40.60.8Output Alignment 0.000.020.040.060.080.100.12 Singular Alignment(A) SUBJ 0 25 50 75 Layer Number0.00... | https://arxiv.org/abs/2505.18752v1 |
nehm Layer 44 false, combination, inois, multiply, scenario, group, action, group, prec, Solo Layer 45 False, #, xspace, @,#, neutral, @, large, sentiment, False Layer 46 False, @, contr, #,@, neutral, cheer, group, ppo, monot Layer 78 false, False, $., @@, @, false, $., false, False, # Layer 79 false, !, ., #, True, <... | https://arxiv.org/abs/2505.18752v1 |
probable ,false ,uncertain , solo, neutral , tender, unknown , gender, unlikely Layer 70 ., reverse, false , scenario, maybe , reverse, identity, gender, conce, Gram Stage 2 Layer 80 rola, Dic, Kra, Column, Common, motor, ..., ge, Salt, kern Filtered Tokens Stage Layer Tokens Stage 1 Layer 25 brief, uclide, Ath, bul, r... | https://arxiv.org/abs/2505.18752v1 |
0 20 40 60 80 Layer Number0.0000.005 Var.-based alignmentLlama3-70B, average across datasets Normal Acc.: 81.09% | w/o PTH Acc.: 26.89% | w/o IH Acc.: 0.92% | Random Heads Ablated Acc.: 77.89% Normal w/o PTH w/o IH Random Heads Ablated Figure 27: Effects of attention heads ablation on the layer-wise separability and al... | https://arxiv.org/abs/2505.18752v1 |
Layer Number0.0000.0020.004 Var.-based alignmentLlama2-70B Famous Unseen label Acc.: 57.78% | w/o PTH Acc.: 1.11% | w/o IH Acc.: 37.78% | Seen label Acc.: 78.89% Unseen label w/o PTH w/o IH Seen labelFigure 32: Effects of attention heads ablation on the layer-wise separability and alignment measures of Llama2-70B hidde... | https://arxiv.org/abs/2505.18752v1 |
PTH Acc.: 0.00% | w/o IH Acc.: 0.00% | Seen label Acc.: 73.33% Unseen label w/o PTH w/o IH Seen label Figure 37: Effects of attention heads ablation on the layer-wise separability and alignment measures of Gemma-2B hidden states on Famous People dataset. 41 0 10 20 Layer Number020 Effective dimension 0 10 20 Layer Numb... | https://arxiv.org/abs/2505.18752v1 |
20 25 30 layer0.05 0.000.05Acc. ImprovementRank 10 Denoising (F) CB Figure 42: Accuracy gains of rank-10 denoising over layers on all datasets with Llama2-7B 0 5 10 15 20 25 30 35 40 layer0.000.050.100.15Acc. ImprovementRank 10 Denoising (A) SST-2 0 5 10 15 20 25 30 35 40 layer0.0000.0250.0500.075Acc. ImprovementRank 1... | https://arxiv.org/abs/2505.18752v1 |
20 30 40 50 60 70 80 layer0.00.10.2Acc. ImprovementRank 2 Denoising (F) CB Figure 47: Accuracy gains of rank-2 denoising over layers on all datasets with Llama2-70B 0 10 20 30 40 50 60 70 80 layer0.2 0.00.2Acc. ImprovementRank 1 Denoising (A) SST-2 0 10 20 30 40 50 60 70 80 layer0.00.10.20.3Acc. ImprovementRank 1 Denoi... | https://arxiv.org/abs/2505.18752v1 |
Article Few-Shot Optimization for Sensor Data Using Large Language Models: A Case Study on Fatigue Detection Elsen Ronando1,2,†,‡, ∗ and Sozo Inoue1,‡ 1Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu Ward, Kitakyushu 808-0135, Japan 2Department of Informa... | https://arxiv.org/abs/2505.18754v1 |
utilization and reducing computational overhead while maintaining high performance [ 9]. Notably, prompts that integrate examples closely aligned with the target task have been observed to enhance model accuracy by as much as 30% compared to prompts with generic or mismatched examples [ 6]. These findings underscore th... | https://arxiv.org/abs/2505.18754v1 |
methods to select representative examples reliably. These visual overlaps highlight the practical need for a more context-aware selection strategy to address semantic and numeric ambiguity in few-shot prompting tasks. We adopt fatigue detection from accelerometer data as our case study to provide a rigorous benchmark f... | https://arxiv.org/abs/2505.18754v1 |
random and distance-based selection, HED-LM achieves relative macro F1-score improvements of 16.6% and 2.3%, respectively, demonstrating the advantage of combining numerical similarity and contextual relevance in selecting examples. This study introduces a novel strategy for few-shot prompting that leverages numerical ... | https://arxiv.org/abs/2505.18754v1 |
was proposed by Shin et al. [9], which identified informative samples based on data distribution properties, though it required extensive labeling and lacked adaptability across different prompting tasks. Dynamic and similarity-based approaches have also been explored. Margatina et al. [26] proposed an active learning-... | https://arxiv.org/abs/2505.18754v1 |
comparing list-based and graph-based representations. Their results indicated that zero-shot learning often outperformed few-shot prompting when random example selection was used, highlighting a significant limitation of current selection strategies. This underscores the need for a more structured and informed approach... | https://arxiv.org/abs/2505.18754v1 |
This dual-filtering approach improves the accuracy, generalizability, and robustness of few-shot prompting, particularly in high-dimensional and noisy sensor data scenarios. The technical details of each component, including data preprocessing, feature 7 of 43 Figure 2. Proposed HED-LM framework for physical fatigue de... | https://arxiv.org/abs/2505.18754v1 |
quality is adequate for identifying features essential for fatigue detection. The filtering process is implemented using a Butterworth filter [ 34], chosen for its smooth fre- quency response and minimal distortion in the passband. The core idea of the filter is to attenuate frequency components above a defined cutoff ... | https://arxiv.org/abs/2505.18754v1 |
distance is then calculated to measure the similarity between this new vector and the feature vectors of previously labeled subjects. The Euclidean distance d(xnew,y)between xnewand a labeled subject vector yis computed in Eq. 6 as: d(xnew,y) =s m ∑ i=1(xnew, i−yi)2 (6) where m=30represents the dimensionality of the fe... | https://arxiv.org/abs/2505.18754v1 |
assesses the relevance of each candidate-new subject pair by assigning a score ranging from 0 to 1, where higher scores reflect more substantial alignment between the candidate’s label and the new subject’s signal patterns. An explanation accompanies each score, adding transparency and interpretability to the process. ... | https://arxiv.org/abs/2505.18754v1 |
in Appendix A.1. Moreover, more details on how to generate domain knowledge, the form of domain knowledge examples, and their placement in the prompting design are explained in Appendix A.2. 3.5. Re-Ranking The Re-Ranking stage in the proposed method serves as a critical intermediary between the distance-based filterin... | https://arxiv.org/abs/2505.18754v1 |
these examples, the numeric data of the new subject (unlabeled) is appended to the prompt, along with an instruction guiding the LLM to make a final decision: “ Please compare the new data with [Example 1 and Example 2] and determine the final label: ‘fatigue’ or ‘non-fatigue.’ ” This structured design ensures that the... | https://arxiv.org/abs/2505.18754v1 |
of 43 • If the response is “non-fatigue,” the new subject is labeled “non-fatigue.” In cases where the LLM produces an ambiguous response (e.g., “I believe it is non-fatigue, but maybe fatigue?”), a fallback mechanism is employed. This involves analyzing the frequency of “fatigue” and “non-fatigue” in the LLM’s output.... | https://arxiv.org/abs/2505.18754v1 |
process can affect the relevance assess- ment of the original subject to the new subject so that the final label becomes more precise in performance. Datasets and Sensor Contexts . We utilized the publicly available dataset from Kathirgamanathan et al. [33], comprising accelerometer-based time series data collected fro... | https://arxiv.org/abs/2505.18754v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.