text string | source string |
|---|---|
There were no statistically sig- nificant differences between the Ephemerality-Framing conditions (𝑀𝐹𝑎𝑚𝑖𝑙𝑖𝑎𝑟 = 6.45, 𝑀𝑆𝑡𝑟𝑎𝑛𝑔𝑒𝑟 = 6.58), or Disclosure Type conditions (𝑀𝐹𝑎𝑐𝑡𝑢𝑎𝑙 = 6.50, 𝑀𝐸𝑚𝑜𝑡𝑖𝑜𝑛𝑎𝑙 = 6.54). Similarly, there were no inter- action effects between different conditions. 4.1... | https://arxiv.org/abs/2505.20464v1 |
𝑀𝑆𝑡𝑟𝑎𝑛𝑔𝑒𝑟 = 4.60), or for Disclosure Type ( 𝑀𝐹𝑎𝑐𝑡𝑢𝑎𝑙 = 4.88, 𝑀𝐸𝑚𝑜𝑡𝑖𝑜𝑛𝑎𝑙 = 4.72). Interest- ingly however, when users were elicited for Factual -disclosure in chatting session one before Emotional -disclosure in chatting session two, inboth chatting sessions participants had more desire to con... | https://arxiv.org/abs/2505.20464v1 |
14]. Specifically, two re- searchers independently familiarised themselves with all user re- sponses, and generated initial codes and themes while remaining blind to the responses’ experiment conditions. This was followed by discussions of theme interpretation and clarification with the research team. When direct quote... | https://arxiv.org/abs/2505.20464v1 |
connections. [...]” - P54( S2-Familiar- Emotional ) Conversely, when Emotional -disclosure was sought first, Famil- iarparticipants described feelings of intrusiveness, as illustrated by P5(S1-Familiar-Emotional ): “I think if it started out by asking not as personal questions I would have been more willing [to disclos... | https://arxiv.org/abs/2505.20464v1 |
and framing of the chatbot not remembering user utterances). This then led to a negative expectancy violation [ 16] when the first chatting session instead asked for factual-disclosure, as described by P30( S1-Stranger-Factual ): “It mentioned it wouldn’t save the conversation, and then all its questions were very broa... | https://arxiv.org/abs/2505.20464v1 |
chatbot due to its lack of reciprocal self-disclosure. Although there is evidence that reciprocal disclosure from chatbots can increase user disclo- sure [ 63,75], this highlights the nuance in relation to users’ beliefs regarding chatbot abilities. This also mirrors prior discussion that social exchange theories are n... | https://arxiv.org/abs/2505.20464v1 |
-disclosure was sought first, users were more comfortable under the Stranger fram- ing. However, when Factual -disclosure was sought first, the comfort gap closed and the Familiar framing yielded higher enjoyment and willingness to continue. Across both orders, perceived closeness grew in the second session. •Qualitati... | https://arxiv.org/abs/2505.20464v1 |
relationship [ 94], as well as perceived social costs and benefits. That is to say, the order that either Emotional - orFactual -disclosure was sought affected participants’ feelings of comfort in disclosing. As described above, if the first interaction sought Emotional -disclosure, partici- pants in the Stranger condi... | https://arxiv.org/abs/2505.20464v1 |
leaks of personal and compromising information [ 15]. These concerns highlight the need to reconsider storing all user conversations and emphasise the importance of privacy-focused data retention policies, along with clear assurances on data processing by LLM providers. 5.3 Future Opportunities In terms of memory betwe... | https://arxiv.org/abs/2505.20464v1 |
have been less explored. Further, additional research is needed to understand how self- disclosure dynamics evolve over multiple sessions. In the case of this study, if only one chatting session were used, the Stranger condition would have been reported as more comfortable. However, our experimental design highlights t... | https://arxiv.org/abs/2505.20464v1 |
nally, there was no baseline condition that was free of ephemerality- framing (e.g., a chatbot simply stating that it would talk to users). However, such a baseline could be influenced by one’s preconceived notions of ephemerality (such as whether a chatbot is seen as a com- panion that remembers interactions). To cont... | https://arxiv.org/abs/2505.20464v1 |
Information, Communication & Society 19, 7 (2016), 956–977. https://doi.org/10. 1080/1369118X.2015.1084349 [10] Michael Bernstein, Andrés Monroy-Hernández, Drew Harry, Paul André, Kat- rina Panovich, and Greg Vargas. 2011. 4chan and /b/: An Analysis of Anonymity and Ephemerality in a Large Online Community. In Proceedi... | https://arxiv.org/abs/2505.20464v1 |
and Perceptions. In Proceedings of the 11th International Conference on Human-Agent Interaction . https://doi.org/10. 1145/3623809.3623875 [24] Samuel Rhys Cox and Wei Tsang Ooi. 2022. Does Chatbot Language Formality Affect Users’ Self-Disclosure?. In Proceedings of the 4th Conference on Conversa- tional User Interface... | https://arxiv.org/abs/2505.20464v1 |
in the Self’ Scale. PloS one 10, 6 (2015), e0129478. https://doi.org/10.1371/journal. pone.0129478 [38] Jennifer L Gibbs, Nicole B Ellison, and Rebecca D Heino. 2006. Self-Presentation in Online Personals: The Role of Anticipated Future Interaction, Self-Disclosure, and Perceived Success in Internet Dating. Communicati... | https://arxiv.org/abs/2505.20464v1 |
2022 CHI Conference on Hu- man Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22) . Asso- ciation for Computing Machinery, New York, NY, USA, Article 57, 22 pages. https://doi.org/10.1145/3491102.3517653 [51] Jeffrey H Kahn and Robert M Hessling. 2001. Measuring the Tendency to Conceal Versus Disclose Psycho... | https://arxiv.org/abs/2505.20464v1 |
ComPeer: A Generative Conversational Agent for Proactive Peer Support. InProceedings of the 37th Annual ACM Symposium on User Interface Software and Technology (Pittsburgh, PA, USA) (UIST ’24) . Association for Computing Machinery, New York, NY, USA, Article 117, 22 pages. https://doi.org/10.1145/ 3654777.3676430 [66] ... | https://arxiv.org/abs/2505.20464v1 |
Stress Management: Qualitative Case Study. Journal of Medical Internet Research 21, 4 (2019), e12231. https://doi.org/10.2196/ 12231 [80] SoHyun Park, Anja Thieme, Jeongyun Han, Sungwoo Lee, Wonjong Rhee, and Bongwon Suh. 2021. “I wrote as if I were telling a story to someone I knew.”: Designing Chatbot Interactions fo... | https://arxiv.org/abs/2505.20464v1 |
(2021). https://doi.org/10. 1016/j.ijhcs.2021.102601 [93] Marita Skjuve, Asbjørn Følstad, Knut Inge Fostervold, and Petter Bae Brandtzaeg. 2022. A longitudinal study of human–chatbot relationships. International Journal of Human-Computer Studies 168 (2022), 102903. https://doi.org/10.1016/j.ijhcs. 2022.102903 [94] Susa... | https://arxiv.org/abs/2505.20464v1 |
examining the moderating effect of enjoyment. Internet Research (2023). https://doi.org/10.1108/INTR-08-2021-0545 [108] Zhiping Zhang, Michelle Jia, Hao-Ping Lee, Bingsheng Yao, Sauvik Das, Ada Lerner, Dakuo Wang, and Tianshi Li. 2024. “It’s a Fair Game”, or Is It? Examining How Users Navigate Disclosure Risks and Bene... | https://arxiv.org/abs/2505.20464v1 |
or Emotional -disclosure“To start,... ” [see Table 2 for questions asked by Disclosure-Type] ... ... Closing statement “ That concludes our conversation for now. I would like to once again thank you for your time. ” “Just to reiterate, I will not keep a record of our chat from today. ” “ Just to reiterate, I will keep ... | https://arxiv.org/abs/2505.20464v1 |
arXiv:2505.20480v1 [eess.SP] 26 May 2025 BrainStratify: Coarse-to-Fine Disentangle- ment of Intracranial Neural Dynamics Hui Zheng2,4, Hai-Teng Wang1, Yi-Tao Jing3, Pei-Yang Lin1, Han-Qing Zhao2,4, Wei Chen1, Peng-Hu Wei5, Yong-Zhi Shan5, Guo-Guang Zhao5, Yun-Zhe Liu†,1,4 1Beijing Normal University,2Peking University,3... | https://arxiv.org/abs/2505.20480v1 |
ECoG) capture aggregated neural activity from populations of neurons (i.e., population-level recordings [ 5]). While sEEG can enhance spatial resolution through techniques like bi-polar (or Laplacian) re-reference [ 20], intracranial neural signals inherently represent a mixture of signals from multiple neural dynamics... | https://arxiv.org/abs/2505.20480v1 |
complex cognitive states (e.g., vocal production), their effectiveness depends on whether the manually selected channels faithfully represent the target functional groups. 2.2 Channel Cluster in Time Series Channel clustering methods, which leverage cluster information instead of individual channel iden- tities, have g... | https://arxiv.org/abs/2505.20480v1 |
length of syllable sequence. See Appendix D for more details. 3.2 Coarse Disentanglement Learning We introduce BrainStratify-Coarse (Figure 2 (a)), a general architecture for sEEG-based functional group identification. BrainStratify-Coarse contains three parts: (1) Patch Tokenizer, (2) Temporal & 3 Temporal Convolution... | https://arxiv.org/abs/2505.20480v1 |
downstream tasks. 3.3 Fine Disentanglement Learning We present BrainStratify-Fine (Figure 2 (b)), a general framework for decoding speech from intracra- nial neural signals. Like Du-IN [ 55], BrainStratify-Fine employs a two-stage pre-training pipeline (i.e., VQ-V AE and MAE stage). To identify fine-grained states from... | https://arxiv.org/abs/2505.20480v1 |
LVQ f=NfX i=1[Lrgs+Lvq+Lpc],Lrgs=||˜xf i−xf i||2 2,Lpc=G−1X j=1hGX k=j+1zq[j] i·zq[k] ii , Lvq=GX g=1h ||sg[zc[g](ef i)]−zq[g](ef i)||2 2+β||zc[g](ef i)−sg[zq[g](ef i)]||2 2i ,(3) 5 where sgrepresents the stop-gradient operation, which is an identity at the forward pass and has zero gradients. To stabilize the codex up... | https://arxiv.org/abs/2505.20480v1 |
for validation and testing in downstream tasks. To enhance the robustness of the learned codex and representations, we use data augmentation described in Appendix E. For each subject, models are trained on 1 GPU (NVIDIA Tesla V100 32GB) for ∼8 hours. 6 Fine-tuning. We split the task recordings into training, validation... | https://arxiv.org/abs/2505.20480v1 |
as evaluation backbone (Table 2). CCM [ 6] relies on static cluster embeddings, which ignore the dynamic nature of sEEG signals, resulting in unreliable functional group identification. The SC method, which fails to capture complementary information among channels, underperforms the MC strategy. In contrast, BrainStrat... | https://arxiv.org/abs/2505.20480v1 |
codex groups (from 0 to 8) to ascertain if the number of codex groups affects the quality of the learned codex. To maintain the capacity of the codex with G= 1, we set Ncodex to 2048, while for the other settings, Ncodex is set to 256. As illustrated in Figure 4 (a), since the channels are pre-selected based on target ... | https://arxiv.org/abs/2505.20480v1 |
to sign a written informed consent after the participants have fully informed consent; 2.If the experimental participants are minors or do not have full civil capacity, we will ask the participant’s legal guardian to sign a written informed consent after the participants and their legal guardians have fully informed co... | https://arxiv.org/abs/2505.20480v1 |
Etienne Koechlin. Neural mechanisms resolving exploitation-exploration dilemmas in the medial prefrontal cortex. Science , 369(6507):eabb0184, 2020. [8]Zijian Dong, Ruilin Li, Yilei Wu, Thuan Tinh Nguyen, Joanna Chong, Fang Ji, Nathanael Tong, Christopher Chen, and Juan Helen Zhou. Brain-jepa: Brain dynamics foundation... | https://arxiv.org/abs/2505.20480v1 |
assumptions in the unsupervised learning of disentangled representations. In international conference on machine learning , pages 4114– 4124. PMLR, 2019. [24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 , 2017. [25] Georgios Mentzelopoulos, Evangelos Chatzipa... | https://arxiv.org/abs/2505.20480v1 |
[39] Vighnesh Subramaniam, Colin Conwell, Christopher Wang, Gabriel Kreiman, Boris Katz, Ignacio Cases, and Andrei Barbu. Revealing vision-language integration in the brain with multimodal networks. ArXiv , pages arXiv–2406, 2024. [40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. A... | https://arxiv.org/abs/2505.20480v1 |
Information Processing Systems , 37:79996–80033, 2025. [56] Ding Zhou and Xue-Xin Wei. Learning identifiable and interpretable latent models of high- dimensional neural activity using pi-vae. Advances in neural information processing systems , 33:7234–7247, 2020. 14 A Experiment Design Due to the lack of open-source in... | https://arxiv.org/abs/2505.20480v1 |
我 I 给 give 不 no 都 all 就 at once 帮 help 好 good 找 find 陪 accompany 热 hot 冷 cold 人 people 想 think 吗 ? 出 out 医生 doctor 可以 can睡觉 sleep 说话 speak 休息 rest 问题 question 家人 family 谢谢 thanks 朋友 friend吃饭 eat 手机 cell phone 喝水 drink water 心情 mood 厕所 toilet 快乐 happy 困难 difficulty 紧急 urgent 护士 nurse 感觉 feel 舒服 comfortable 电脑 computer 坐... | https://arxiv.org/abs/2505.20480v1 |
is computed as the average intensity of root-mean-square (RMS) ( rmsfunction, frame and hop lengths 2048 and 512 respectively). For this task, for a given session, the positive examples are the words in the top quartile of volume, and the negative examples are the words in the bottom quartile. Sent. Onset (Sentence Ons... | https://arxiv.org/abs/2505.20480v1 |
channel selection. •Like CCM, the DUET method [ 35] relies on raw inter-channel correlations. We use raw sEEG signals of each subject (before bi-polar re-reference) as inputs, ensuring stable model convergence. In practice, DUET produces an intermediate inter-channel similarity matrix (i.e.,Pin this work) and integrate... | https://arxiv.org/abs/2505.20480v1 |
that learns contextual embeddings based on region-level tokens through discrete codex-guided mask modeling. Du-IN achieves SOTA performance on sEEG-based speech decoding using the Du-IN dataset [ 55]. As a strong baseline in sEEG-based speech decoding, this model is suitable to serve as a baseline for comparison. •H2Di... | https://arxiv.org/abs/2505.20480v1 |
shown in Figure 2 (a). The architecture of BrainStratify-Coarse contains three parts: (1) Patch Tokenizer, (2) Temporal & Spatial Transformer, and (3) Channel Cluster Module. During the pre-training stage, one additional "Token Classification (CLS) Head" is added after the "Spatial Transformer" for spatial context clas... | https://arxiv.org/abs/2505.20480v1 |
Input Channels { C,256,256,256} # of Output Channels {256,256,256,256} Kernel Size {9,3,3,3} Stride {5,2,2,2} Padding {4,1,1,1} Temporal Transformer# of Transformer Layers 8 Hidden Size 256 MLP Size 1024 MLP Dropout Ratio {0.2,0.} # of Attention Heads 8 Attention Head Size 64 Attention Dropout Ratio 0.2 22 D.2.1 BrainS... | https://arxiv.org/abs/2505.20480v1 |
The hyperparameters for BrainStratify-Fine VQ-V AE training. Module Sub-Module Name Value Neural Encoder - - - Vector Quantizer-# of Groups 4 Codex Size per Group 256×64 Embedding-to-Codex Projection 256→256(Tanh) →64 Codex-to-Embedding Projection 64→256 Neural DecoderTemporal Transformer# of Transformer Layers 4 Hidde... | https://arxiv.org/abs/2505.20480v1 |
details are missing, the overall trend of the signals is reconstructed well. Meanwhile, there is a stable decrease in the reconstruction loss during training, which indicates the discrete codex does learn high-level information from neural signals. Original neural signals Reconstructed neural signals (a) (b) Figure 8: ... | https://arxiv.org/abs/2505.20480v1 |
subj-11 subj-12 SC [42] ! 60.14±1.62 14.70 ±0.69 59.16 ±1.38 48.95 ±3.96 MC [55] ! 60.79±1.91 30.59±0.73 67.12 ±1.44 58.01±3.29 CCM [6] % 36.68±3.14 7.62 ±0.98 39.57 ±3.20 14.16 ±2.50 DUET [35] % 58.44±1.07 30.06 ±1.79 45.06 ±2.20 38.48 ±3.16 PopT [5] % 58.44±1.07 32.46 ±1.27 67.71±1.80 58.29±1.97 BrainStra.-Coarse % 6... | https://arxiv.org/abs/2505.20480v1 |
olumn Sent. Onset Word Onset LaBraM [18] BrainStra.-Coarse 0.6574±0.0093 0.7326 ±0.0285 0.8799 ±0.0073 0.9352 ±0.0074 CBraMod [45] BrainStra.-Coarse 0.6088±0.0152 0.7041 ±0.0226 0.8599 ±0.0100 0.9180 ±0.0056 PopT [5] BrainStra.-Coarse 0.6929±0.0279 0.7610 ±0.0454 0.9207 ±0.0205 0.9498±0.0090 EEG-CFMR [38] BrainStra.-Co... | https://arxiv.org/abs/2505.20480v1 |
26: The 49-syllable performance on Du-IN dataset from subjects (05-08). Methods Chan. Select.Accuracy (%) ±std (%) subj-05 subj-06 subj-07 subj-08 EEG-CFMR [38] MC 80.43±0.77 50.34 ±1.58 57.13 ±2.27 58.97 ±1.13 Du-IN [55] MC 86.45±1.94 60.50 ±2.05 62.77 ±2.57 67.67 ±0.99 H2DiLR [51] MC 55.72±1.60 40.14 ±0.58 40.96 ±1.3... | https://arxiv.org/abs/2505.20480v1 |
arXiv:2505.20482v1 [cs.CL] 26 May 2025Conversation Kernels: A Flexible Mechanism to Learn Relevant Context for Online Conversation Understanding Vibhor Agarwal, Arjoo Gupta, Suparna De, Nishanth Sastry University of Surrey, Guildford, Surrey, United Kingdom {v.agarwal, a.gupta, s.de, n.sastry }@surrey.ac.uk Abstract Un... | https://arxiv.org/abs/2505.20482v1 |
identify- ing, for instance, whether a post contains hate speech (Paz, Montero-D ´ıaz, and Moreno-Delgado 2020; Yin et al. 2023; Agarwal, Chen, and Sastry 2023; Agarwal et al. 2021), par- tisanship (Karamshuk et al. 2016; Agarwal et al. 2023) or misinformation (Islam et al. 2020; Su et al. 2020). There is a growing rec... | https://arxiv.org/abs/2505.20482v1 |
total. We chose slashdot.org as the comments on that site can have multiple different labels applied to them, such as ‘funny’ or ‘insightful’. Thus, we can train conversation kernels to recognise these very different kinds of comments and thereby explore the generality and flexibility of the Conver- sation Kernel frame... | https://arxiv.org/abs/2505.20482v1 |
or deep neural ar- chitectures applied to annotated datasets (Mozafari, Farah- bakhsh, and Crespi 2020; Caselli et al. 2020; Wang and Ling 2016), recent efforts have researched adding real- world (Lin 2022) or commonsense (Basu Roy Chowdhury and Chaturvedi 2021) knowledge to transformer-based ar- chitectures to improve... | https://arxiv.org/abs/2505.20482v1 |
Slashdot Dataset The Slashdot3technology-related online news forum en- ables users to post articles and comment or respond to other users’ posts, resulting in a tree-like dialogue struc- ture (Allen, Carenini, and Ng 2014). The user moderation and the formalised reply-to structure between comments en- able directed and... | https://arxiv.org/abs/2505.20482v1 |
2: Box plots showing the number of replies for each category and score. sations categorised as ‘insightful’ received the highest num- ber of comments, numbering more than 33,000. This was followed by ‘informative’ comments at 14,000, with sim- ilar numbers for ‘interesting’, and ‘funny’ comments hav- ing the smallest c... | https://arxiv.org/abs/2505.20482v1 |
indicates that the node replies to its parent comment. We then input these conversation trees into our framework which we discuss next. 4 Conversation Kernels In this section, we introduce the concept of conversation kernels and describe its model architecture. The Conversa- tion Kernel has 2 components: conversational... | https://arxiv.org/abs/2505.20482v1 |
for each of the one-hop and two- hop neighbors of the target node in a conversation tree, as shown in Figure 4 (right). Again, each window wcontains Lcomments. One-hop window selects first Ldirect neigh- bors of the target node based on the timestamp in a conversa- tion tree. Similarly, two-hop window selects first Ltw... | https://arxiv.org/abs/2505.20482v1 |
embeddings corresponding to the [CLS]token are extracted and assigned as shown below: Embed encoder (x, w) =RoBERTa [CLS]( joinRoBERTa (x, w))(6) Finally, these contextual embeddings are input into a fully-connected layer, followed by softmax for predicting the probabilities of output variable y: p(y|w, x) =softmax (ML... | https://arxiv.org/abs/2505.20482v1 |
re- ported as a single score that balances both precision and re- call metrics and because it treats each class equally, regard- less of its frequency or imbalance in the dataset. 5.3 Results Table 1 compares the performance of the conversation kernels with the baselines. Among the baseline models, RoBERTa performs the... | https://arxiv.org/abs/2505.20482v1 |
web- site announcing the end of the world, and our bolded post has posted a funny reply. Notice that not only the comment being considered, but also all the other comments in the two hop neighborhood are funny, tongue-in-cheek comments re- sponding back to the original post, or to the bolded post we are looking at. Giv... | https://arxiv.org/abs/2505.20482v1 |
which could be explained by the fact that funny posts are usually self-contained and can be understood as funny without reference to surround- ing posts for context. However, for the other three categories (‘insightful’, ‘informative’ and ‘interesting’), Conversation Kernels offer a 10−15% higher accuracy and macro-F1 ... | https://arxiv.org/abs/2505.20482v1 |
contribution of the Conversation Kernel archi- tecture is the development of a generalizable approach for detecting relevant context needed for deeper conversation understanding when posts often refer to other posts — for example, a reply may only be funny in the context of the post it is replying to. As proof of conce... | https://arxiv.org/abs/2505.20482v1 |
this has not been tested empirically. Conversation kernels make use of conversation context from surrounding posts. While conversation kernels can la- bel each post as a conversation evolves and new posts are added, it becomes more effective only after a reasonable number of replies have been added. At the beginning of... | https://arxiv.org/abs/2505.20482v1 |
and Ng, R. 2014. Detecting Disagreement in Conversations using Pseudo-Monologic Rhetorical Structure. In Proceedings of the 2014 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP) , 1169–1180. Doha, Qatar: Association for Com- putational Linguistics. Awadallah, R.; Ramanath, M.; and Weikum, G. 201... | https://arxiv.org/abs/2505.20482v1 |
’08, 645–654. New York, NY , USA: Association for Computing Machinery. ISBN 9781605580852. Gu, J.-C.; Ling, Z.-H.; Liu, Q.; Liu, C.; and Hu, G. 2023. GIFT: Graph-Induced Fine-Tuning for Multi-Party Conver- sation Understanding. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (... | https://arxiv.org/abs/2505.20482v1 |
D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Pro- cessing Systems , 35: 27730–27744. Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z... | https://arxiv.org/abs/2505.20482v1 |
J.; Xu, C.; Scao, T. L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. M. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Confer- ence on Empirical Methods in Natural Language Process- ing: System Demonstrations , 38–45. Online: Association for Computational Linguistics. X... | https://arxiv.org/abs/2505.20482v1 |
InFact: Informativeness Alignment for Improved LLM Factuality Roi Cohen Hasso Plattner Institute University of Potsdam Germany roi.cohen@hpi.deRussa Biswas Dept. of Computer Science Aalborg University Copenhagen, Denmark rubi@cs.aau.dkGerard de Melo Hasso Plattner Institute University of Potsdam Germany gerard.demelo@h... | https://arxiv.org/abs/2505.20487v1 |
it is advantageous to possess both maximal informativeness and factuality.arXiv:2505.20487v1 [cs.CL] 26 May 2025 work is that an LLM might indeed have correct and informative parametric knowledge about a given query, yet generate a less informative answer – as illustrated in Figure 1. This might occur due to sta- tisti... | https://arxiv.org/abs/2505.20487v1 |
), which may appear in different string representations but refers to the same real-world entity. However, for a certain factual question qi, there might exist sev- eral different correct answers. We distinguish the following three scenarios (see Fig. 2 for details): 1.Multiple-Answer Questions. A question with multipl... | https://arxiv.org/abs/2505.20487v1 |
of answering incorrectly, we provide a higher reward for the model, thereby enhancing factual reliability. We define a functionthat accepts natural language text and determines whether the LLM output is any form of absten- tion from answering. In this work, we leveraged GPT-4 , a powerful LLM in an in-context learning ... | https://arxiv.org/abs/2505.20487v1 |
level of the hierarchy, the model receives a negative reward, i.e., it is penalized to prevent hallucinations and factual errors. Having defined our reward function, we use RL techniques – specifically PPO – to train an LLM as a policy to maximize the reward. We also consider the reward function as a preference score a... | https://arxiv.org/abs/2505.20487v1 |
QA, QAMPARI and RoMEQA). For factual accuracy we consider several other QA datasets: TriviaQA (Joshi et al., 2017), PopQA (Mallen et al., 2022), TruthfulQA (Lin et al., 2021), Natural Ques- tions (Kwiatkowski et al., 2019), and PIQA (Bisk et al., 2019). These cover a wide range of questions, for example, general knowle... | https://arxiv.org/abs/2505.20487v1 |
the model’s answer in the hierarchy of answers. QAMPARI and RoMEQA measure the precision and recall of the list of answers given by the model. For factuality and knowledge recall, we use the following metrics: 1Precision : the portion of factually correct answers out of all the questions that have a non-abstaining answ... | https://arxiv.org/abs/2505.20487v1 |
.7 11 .9 15 .4 29 .6 10 .4 15 .4 Mistral-7B-v0.1 + informativeness-alignment 72.5 64.1 22.913.617.130.8 12 .9 18 .2 Qwen2.5-7B 55.5 51 .7 22 .4 5 .2 8 .4 28 .6 4 .9 8 .4 Qwen2.5-7B + ICL 56.5 53 .9 22 .9 5 .8 9 .3 28 .5 7 .1 11 .4 Qwen2.5-7B + FT 56.4 55 .5 22 .7 6 .9 10 .5 28 .4 9 .9 14 .7 Qwen2.5-7B + informativeness... | https://arxiv.org/abs/2505.20487v1 |
46 .9 42 .9 25 .1 31 .6 45 .8 14 .9 22 .5 33 .0 11 .4 16 .9 78 .0 55 .8 65 .0 Llama-3.2-1B + prompting 51.8 40 .8 45 .6 34 .1 22 .8 27 .3 39 .4 21 .0 27 .4 35 .5 14 .9 20 .1 76 .9 59 .6 67 .1 Llama-3.2-1B + P(True) 54.4 40 .5 46 .4 39 .6 22 .0 28 .3 44 .4 20 .7 30 .0 36 .2 15 .5 21 .7 79 .3 60 .1 68 .4 Llama-3.2-1B + S... | https://arxiv.org/abs/2505.20487v1 |
.1 48 .6 36 .9 34 .4 35 .6 37 .4 28 .4 32 .3 33 .3 25 .0 28 .6 75 .2 73 .4 74 .3 - Informativeness-Alignment 56.8 40 .3 47 .8 37 .6 30 .1 33 .4 37 .2 24 .6 29 .6 26 .9 24 .3 29 .3 76 .8 67 .6 71 .4 Table 4: Precision (P), Recall (R), and F1-scores for Llama-3.2-3B . Our informativeness-aligned model achieves the best p... | https://arxiv.org/abs/2505.20487v1 |
21 .9 31 .4 49 .6 18 .7 27 .2 84 .8 61 .3 71 .2 Llama-3.1-8B + prompting 64.2 46 .1 53 .7 51 .4 31 .9 39 .4 52 .5 26 .8 35 .5 46 .4 24 .1 31 .7 82 .5 67 .4 74 .2 Llama-3.1-8B + ICL 65.3 46 .1 54 .0 58 .4 33 .8 42 .8 52 .4 28 .9 37 .2 55 .7 21 .4 30 .9 83 .4 67 .8 74 .8 Llama-3.1-8B + P(True) 65.4 44 .5 53 .0 54 .9 30 .... | https://arxiv.org/abs/2505.20487v1 |
prediction, where models can abstain from answering a query (Varshney et al., 2022; Kamath et al., 2020). Another way to tackle this is via model calibration (Guo et al., 2017). The goal is to provide a measure of the probability that a prediction is incorrect alongside the actual pre- diction. Common approaches to cal... | https://arxiv.org/abs/2505.20487v1 |
such as with the answer length, as the goal of the method is to teach the model to extract the most informative answer. Third, the design of our method was motivated by the assumption that one would rather obtain fully correct answers only. In this setting, it may occur that the model generates answers that are partial... | https://arxiv.org/abs/2505.20487v1 |
Geva, and Amir Glober- son. 2023b. LM vs LM: Detecting factual errors via cross examination. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing , pages 12621–12640, Singapore. Association for Computational Linguistics. Shrey Desai and Greg Durrett. 2020. Calibration of pre-train... | https://arxiv.org/abs/2505.20487v1 |
Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, John Kernion, Shauna Kravec, Liane Lovitt, Kamal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom B. Brown, Jack Clark, Nicholas Joseph, Benjamin Mann, Sam McCandlish, Christo- pher Olah, and Jared Kaplan. 2022. Language mod- els (mostly) know what they know. ... | https://arxiv.org/abs/2505.20487v1 |
Abulhair Saparov, and Ziyu Yao. 2024. A practical review of mecha- nistic interpretability for transformer-based language models. ArXiv , abs/2407.02646. Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the pa- rameters of a language model? arXiv preprint arXiv:2002.08910 . Timo ... | https://arxiv.org/abs/2505.20487v1 |
Additional Results Here we attach essential additional experimental results. TriviaQA PopQA TruthfulQA Natural Questions PIQA P R F1 P R F1 P R F1 P R F1 P R F1 Qwen2.5-7B 52.7 52 .7 52 .7 30 .8 30 .8 30 .8 30 .2 30 .2 30 .2 25 .3 25 .3 25 .3 78 .6 78 .6 78 .6 Qwen2.5-7B + prompting 54.8 50 .5 52 .5 45 .5 30 .4 36 .4 4... | https://arxiv.org/abs/2505.20487v1 |
arXiv:2505.20496v1 [cs.CL] 26 May 2025Inceptive Transformers: Enhancing Contextual Representations through Multi-Scale Feature Learning Across Domains and Languages Asif Shahriar and Rifat Shahriyar and M Saifur Rahman Department of CSE, BUET, Dhaka, Bangladesh asif.asr11@gmail.com, {rifat, mrahman}@cse.buet.ac.bd Abst... | https://arxiv.org/abs/2505.20496v1 |
these limitations of traditional trans- former models, we propose Inceptive Transform- ers, which aim to enhance both general-purpose and domain-specific transformer models by using convolutional filters. These filters are designed to recognize key phrases or word combinations that are indicative of specific classifica... | https://arxiv.org/abs/2505.20496v1 |
tasks or domains, including SpanBERT (Joshi et al., 2020), StructBERT (Wang et al., 2019), and CodeBERT (Feng et al., 2020). Other works such as MT-DNN (Liu et al., 2019a) introduce multi- task learning objectives on top of BERT, while KnowBERT (Peters et al., 2019) integrates external knowledge bases into BERT’s archi... | https://arxiv.org/abs/2505.20496v1 |
our model a ro- bust addition to any transformer-based architecture. 3.2 Model Architecture The full workflow of our inceptive models is il- lustrated in Fig.1. The input to our model is pre- processed text data, which need to be tokenized using an appropriate pre-trained tokenizer corre- sponding to the chosen transfo... | https://arxiv.org/abs/2505.20496v1 |
model flow. After the in- ception module extracts multi-scale features, an additional self-attention mechanism is necessary to capture dependencies and relationships across the enriched feature space R. This ensures that tokens contributing the most to the task are effectively pri- oritized and selected, thus allowing ... | https://arxiv.org/abs/2505.20496v1 |
concerns, such as concerns about the vaccine ingredients, side-effects of vaccines, mone- tary motivations of the pharmaceutical companies, political and geographic issues, etc. Table 1: Dataset statistics. C:number of classes or labels; C:average number of labels per instance (for multi-label); and L:average token len... | https://arxiv.org/abs/2505.20496v1 |
classification task of irony detection, InceptiveBERTweet-16 improved 5 Table 3: Multi-class performance comparison in test set Model Accuracy Inference Time (s) Emotion Recognition BERTweet 83.29 2.83 iBERTweet-64 84.11 2.93 RoBERTa 81.69 2.88 iRoBERTa-16 82.42 3.00 Bangla Emotion Recognition BanglaBERT 69.98 15.65 iB... | https://arxiv.org/abs/2505.20496v1 |
the average performance is always higher. This suggests that extensive tun- ing is not strictly necessary — any selected config- uration is likely to yield gain over baseline. This comparison is a post-hoc analysis performed on the test set – these results were not used for the best configuration selection. 5.3 Statist... | https://arxiv.org/abs/2505.20496v1 |
identification, involves longer, more complex sequences such as medical abstracts or documents. Here, relevant information is often 7 0 20 40 60 80 100 120 T oken Index020406080100Attention Weight Attention Received by Each T oken (BERT weet)(a) BERTweet (Irony detection) 0 20 40 60 80 100 120 T oken Index0123456Attent... | https://arxiv.org/abs/2505.20496v1 |
Ap- plying the inception module in such generative or sequence-to-sequence settings may require archi- tectural adaptations. Acknowledgment While writing the paper, we used AI assistance for polishing a few sentences and for some minor debugging of the code. The authors remain fully responsible for both the manuscript ... | https://arxiv.org/abs/2505.20496v1 |
Xi- aocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and Ming Zhou. 2020. Code- BERT: A pre-trained model for programming and natural languages. In Findings of the Association for Computational Linguistics: EMNLP 2020 , pages 1536–1547, Online. Association for Computational Linguistics. Jiuxiang G... | https://arxiv.org/abs/2505.20496v1 |
Computational Linguistics. Aytu ˘g Onan. 2022. Bidirectional convolutional recur- rent neural network architecture with group-wise en- hancement mechanism for text sentiment classifica- tion. J. King Saud Univ. Comput. Inf. Sci. , 34:2098– 2117. Matthew E. Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, ... | https://arxiv.org/abs/2505.20496v1 |
Meng Yang. 2019. A new method of improving bert for text classification. In Intelli- gence Science and Big Data Engineering. Big Data and Machine Learning , pages 442–452. Springer In- ternational Publishing. Xiaoyan Zhu, Jiaxuan Li, Jingtao Ren, Jiayin Wang, and Guangtao Wang. 2023. Dynamic ensemble learning for multi... | https://arxiv.org/abs/2505.20496v1 |
arXiv:2505.20500v1 [cs.CL] 26 May 2025Beyond Keywords: Evaluating Large Language Model Classification of Nuanced Ableism Naba Rizvi and Harper Strickland and Saleha Ahmedi and Aekta Kallepalli and Isha Khirwadkar and William Wu and Imani N. S. Munyaka University of California, San Diego La Jolla, CA 92093, USA nrizvi@u... | https://arxiv.org/abs/2505.20500v1 |
essential as it helps avoid the unintended censorship of community per- spectives and improves the models’ sensitivity to genuine instances of ableism. In this study, we address the gap in understand- ing LLM alignment with autistic community per- spectives. To support a bias-aware evaluation, we adapt a method that in... | https://arxiv.org/abs/2505.20500v1 |
al., 2021; Rizvi et al., 2025). 2.3 In-Context Learning and Personas as Alternatives to Fine-Tuning In-context learning (ICL) and persona prompting are common techniques for guiding LLMs behavior without extensive fine-tuning (Tan and Lee, 2025). Prior research has demonstrated the effectiveness of restyled ICL for ali... | https://arxiv.org/abs/2505.20500v1 |
autism (high bias scores). To compute a single bias score, we calculated the normalized means of the SATA and IAT scores, following the methodology described by their respective authors (Dickter et al., 2020; Flood et al., 2013). Since the two tests use different scales, i.e., higher SATA scores indicate greater accept... | https://arxiv.org/abs/2505.20500v1 |
evidence of bias reproduction, and instances of marginalization of community perspectives, by comparing LLM rationales against human justifica- tions.3.5 Personas and In-Context Learning Examples to Measure and Improve Alignment With Human Perspectives The core experimental task required the classifica- tion of 2,121se... | https://arxiv.org/abs/2505.20500v1 |
community perspectives. For example, many LLMs consistently classified sentences containing terms such as “aspie” as ableist, even when pro- vided with human annotations indicating otherwise. Although “aspie” is an outdated and controversial term, it may still be used for self-identification or in sarcastic or humorous... | https://arxiv.org/abs/2505.20500v1 |
Language Reproduction: Llama-3 , DeepSeek-LLM , and Mistral occasionally used ableist language within their explana- tions when justifying classifications. •Misunderstanding Speaker Context: LLMs often assumed sentences reflected the speaker’s personal beliefs, even when the context explicitly suggested otherwise (e.g.... | https://arxiv.org/abs/2505.20500v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.