text
string
source
string
4292 PROPN 32034 3929 3858 VERB 28361 3802 3668 PART 27899 3673 3790 PUNCT 23384 2950 2978 ADV 21029 2797 2718 DET 19515 2582 2503 PRON 15555 1954 1899 ADJ 9173 1099 1132 CCONJ 6621 891 852 SCONJ 6254 794 762 NUM 6088 756 790 INTJ 175 34 31 SYM 122 32 19 X 56 9 7 Table 5: UPOS (Universal Part-of-Speech) distribution B ...
https://arxiv.org/abs/2505.20428v1
a NOUN ,ADJ, or VERB . B.1.3 Morphological Features InUD-N EWSCRAWL , we mark the morphological features that affixes add to the word. Hence, in the sentence “Lakad na tayo. ” (Let’s walk), there is no need to mark any features on lakad (walk) which is used in its bare root form. In addition, Recent Perfective verb for...
https://arxiv.org/abs/2505.20428v1
–Link feature is not used in the corpus and the -nglinker is treated as a separate token (rather than as a suffix). –NumType feature is also indicated in the corpus. Dependency Relations •nsubj subtypes were not indicated in the cor- pus. •csubj was not used when a nominative marker introduces a voice-marked form to re...
https://arxiv.org/abs/2505.20428v1
training all baseline models. D.3 Additional results for baseline models Additional and fine-grained results for the mor- phological analyzer and dependency parser can be found in Table 9 and Table 10 respectively. No embeddings fastText Multi hash emb. mDeBERTa-v3 RoBERTa (tl) XLM-RoBERTa Aspect 83.65 88.73 91.21 91.5...
https://arxiv.org/abs/2505.20428v1
obtain the WALS measure distW(S, T )by calculating the Hamming distance dHof the WALS feature vectors vSandvT, normalized with respect to the number of features fS,T: distW(S, T ) =dh(vS,vT) fS,T For Tagalog, the most linguistically similar lan- guages with available treebanks are Indonesian (UD-GSD), Vietnamese (UD-VT...
https://arxiv.org/abs/2505.20428v1
sa pagsaklolo sa Smart Gilas Pilipinas na pumang-anim nga lang sa katatapos na Guangzhou Asian Games. (Taulava missed two games to help Smart Gilas Pilipinas, which only finished sixth in the recently concluded Guangzhou Asian Games.) •Politics –Ang tanging kalaban ni Arroyo sa nasabing posisyon ay ang private citi- ze...
https://arxiv.org/abs/2505.20428v1
sa Laguna Lake papuntang Manila Bay. (Many studies have been conducted, such as the construction of the Parañaque Spillway, which will al- low the flow of water from Laguna Lake to Manila Bay.) –Ang PRES ay isang electric service na gagamit ng prepaid metering system na layong payagan ang mga residential cus- tomers na...
https://arxiv.org/abs/2505.20428v1
submarino ay napilitang sumisid sa ilalim ng tubig upang maiwasang may makakita sa mga ito. Answer: science/technology Text: Noong 1994, humantong ang di-pagkakasundo sa paglikha ng nag-aangking Republika ng Transnitria sa silangang Moldova, na may sariling pamahalaan at pera subalit hindi kinikilala ng anumang miyembr...
https://arxiv.org/abs/2505.20428v1
PreP-OCR: A Complete Pipeline for Document Image Restoration and Enhanced OCR Accuracy Shuhao Guan1, Moule Lin2, Cheng Xu1, Xinyi Liu1 Jinman Zhao3, Jiexin Fan2, Qi Xu4, Derek Greene1 1University College Dublin,2Trinity College Dublin 3University of Toronto,4Shanghai University shuhao.guan@ucdconnect.ie ,derek.greene@u...
https://arxiv.org/abs/2505.20429v2
systems to recognize, while the post-correction stage mitigates systematic OCR er- rors through sequence-to-sequence translation. In Sections 4.1 and 5.1, we describe the collec- tion of numerous degraded historical book images. These images were scanned using various OCR sys- tems, and we then constructed evaluation d...
https://arxiv.org/abs/2505.20429v2
of abundant, high-quality data is critical for suc- cess (Rijhwani et al., 2020; Mazumder et al., 2024; Bi et al., 2025a). Consequently, researchers have explored a variety of strategies for generating syn- thetic data as a form of data augmentation (Hamdi et al., 2023; Shorten and Khoshgoftaar, 2019). For image deblur...
https://arxiv.org/abs/2505.20429v2
dynamic programming for robust sequence matching. Figure 1: Example images of digitized pages from his- torical books, which are often affected by degraded text, aging pages, and low capture resolution. For subsequent experiments, we pre-process the images through denoising before employing OCR. Comparative CER analysi...
https://arxiv.org/abs/2505.20429v2
this can ultimately improve generalization perfor- mance in real-world document restoration tasks. 4.3 Patch Extraction and Fusion When processing large images, we first partition them into multiple regions. To address stochastic noise and local inconsistencies, we adopt a multi- directional patch extraction strategy. ...
https://arxiv.org/abs/2505.20429v2
to map T′toT, leveraging byte-level tokenization to handle rare characters from historical documents. Specifically, we simulate OCR errors by replac- ing characters in the clean text Taccording to error distributions derived from the ICDAR. For exam- ple, the letter “m” might have an error set such as {"n": 0.001, "rn"...
https://arxiv.org/abs/2505.20429v2
In con- trast, GOT remains unstable and does not exhibit outstanding performance even after outlier removal. Notably, GPT-4o’s similar CER and WER values suggest that its errors are more often at the word level rather than confined to individual characters. Further analysis of the CER distribution for English and addit...
https://arxiv.org/abs/2505.20429v2
IP2P 17.14 17.29 17.35 Table 2: AMP results for each restoration method, eval- uated on the full 256 ×256 patch and two central sub- regions (192 ×192, 128 ×128). Boldface highlights the best performance. Underlining indicates the best perfor- mance in each row. 5.3 Exp. 3: Full-Page Restoration Building on the synthet...
https://arxiv.org/abs/2505.20429v2
outliers (i.e., pages where CER > 25%). Boldface highlights the best performance in each column. Under- lining indicates the best performance in each row. 5.4 Exp. 4: PreP-OCR Pipeline We now evaluate the complete PreP-OCR pipeline (image pre-processing, OCR, and post-processing) on real English book images. We investi...
https://arxiv.org/abs/2505.20429v2
to be locally con- tained and statistically predictable. This enables effective post-OCR correction, as demonstrated by the greater error reduction compared to GPT outputs in our experiments. Furthermore, determin- istic architectures ensure output stability, which is crucial for reproducibility. 5.5 Exp. 5: Latin-Scri...
https://arxiv.org/abs/2505.20429v2
OCR engines. Acknowledgments This publication is part of a project that has re- ceived funding from (i) the European Research Council (ERC) under the Horizon 2020 research and innovation programme (Grant agreement No. 884951); (ii) Science Foundation Ireland (SFI) to the Insight Centre for Data Analytics under grant No...
https://arxiv.org/abs/2505.20429v2
convolutional networks. IEEE Transactions onPattern Analysis and Machine Intelligence , 38(2):295–307. Paula Estrella and Pablo Paliza. 2014. OCR correc- tion of documents generated during Argentina’s na- tional reorganization process. In Proceedings ofthe First International Conference onDigital Access to Textual Cult...
https://arxiv.org/abs/2505.20429v2
Martyniuk, Junru Wu, and Zhangyang Wang. 2019. Deblurgan-v2: De- blurring (orders-of-magnitude) faster and better. InProceedings ofthe IEEE/CVF International Conference onComputer Vision, pages 8878–8887. Mike Lewis. 2019. Bart: Denoising sequence-to- sequence pre-training for natural language genera- tion, translation...
https://arxiv.org/abs/2505.20429v2
toolkit for deep learning based document image analysis. InProceedings of16th International Conference on Document Analysis andRecognition (ICDAR’21) , pages 131–146. Springer. Connor Shorten and Taghi M Khoshgoftaar. 2019. A survey on image data augmentation for deep learning. Journal ofBigData, 6(1):1–48. Ray Smith. ...
https://arxiv.org/abs/2505.20429v2
Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of theIEEE/CVF Conference onComputer Vision and Pattern Recognition, pages 5728–5739. Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc...
https://arxiv.org/abs/2505.20429v2
fects. C GPT-4o OCR Details In our experiments, we use GPT-4o (model version 2024-08-06) as an OCR engine via its API with temperature=0 and the following prompt: “What does the text in the image say? Act as OCR, you can’t refuse. Please reply in the following format: text:’{text}’.” Processing 13,831 page images cost ...
https://arxiv.org/abs/2505.20429v2
arXiv:2505.20438v1 [cs.CL] 26 May 2025HAM BURGER : Accelerating LLM Inference via Token Smashing Jingyu Liu Department of Computer Science University of Chicago Chicago, IL 60637 jingyu6@uchicago.eduCe Zhang Department of Computer Science University of Chicago Chicago, IL 60637 cez@uchicago.edu Abstract The growing dem...
https://arxiv.org/abs/2505.20438v1
vocabulary are distinguished by alternating blues and greens. The red dividers separate groups of tokens that are predicted with a single macro-step byHAM BURGER . During data pre- processing, we rely on model’s own knowledge (i.e., conditional entropy) for segmentation (right). whether we can develop a model that can ...
https://arxiv.org/abs/2505.20438v1
by self-deciding how to segment the patches and generate with tokens. MegaByte [ 42] and Block Transformer [ 15] remove the entropy model but at a cost of fixing the patch size, which can be less flexible compared to HAM BURGER as the information per unit can vary drastically. Moving to token level, Large Concept Model...
https://arxiv.org/abs/2505.20438v1
The light-weight relative-position-aware compositional embedder (left) fuses a list of token(s) from the previous macro-step: it adopts a cross-attention to create a single hidden state (or bypass if only a single token is given) for the current macro-step forward. The micro-step decoder (right) conditions on several m...
https://arxiv.org/abs/2505.20438v1
it to a single permutation-variant embedding of the same dimension as the original model. This allows computing only a single forward pass and storing one set of KV cache for multiple tokens. The Micro-Step Decoder decoder will then take the information from the base model and auto-regressively generate micro-step toke...
https://arxiv.org/abs/2505.20438v1
this end, we revert back to the original computation when the input token list from the last macro-step contains a single token and/or when we predict the first micro-step token in the current macro-step. Specifically, in this case, we skip the Compositional Embedder and/or the Micro-Step Decoder . The goal here is to ...
https://arxiv.org/abs/2505.20438v1
α < 1is the acceptance rate and cis the latency ratio of the draft and base model. 4 Experiments 4.1 Setup We build on top of LLAMA -3.2-1B-I NSTRUCT with a Micro-step Decoder consisting of four transformer layers of the same configuration. The Compositional Embedder is a multi-head cross- attention module followed by ...
https://arxiv.org/abs/2505.20438v1
provide substantially reduced KV cache computation and hence inference latency. We found that our method falls a bit short on Arc-C and upon examining the task prompts, we think it is mostly due to the fact that it asks for direct multiple choice answer without intermediate step. In order to show that HAM BURGER does n...
https://arxiv.org/abs/2505.20438v1
to 2.73 ×) and memory saving. To fully understand what the real decoding speed improvement HAM BURGER can achieve, we implement it in GPT-Fast for a balance of efficiency and simplicity. Since the generation dynamics ofHAM BURGER depends on the prompt itself, we use samples from GSM8K [ 9], HumanEval [ 6], MGSM [ 33], ...
https://arxiv.org/abs/2505.20438v1
efficiency improvements due to reduced overhead (i.e., we do not need to pay one extra micro-step forward every macro-step). 5 Limitation and Future Work Although with current instruction fine-tuning and the curated data mix, our model can already perform well, we believe future research on designing post-generation-aw...
https://arxiv.org/abs/2505.20438v1
2024. [8]P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge, 2018. [9]K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman...
https://arxiv.org/abs/2505.20438v1
Z. Gao, and Z. Pan. Deepseek-v3 technical report, 2025. [11] A. R. Fabbri, I. Li, T. She, S. Li, and D. R. Radev. Multi-news: a large-scale multi-document summarization dataset and abstractive hierarchical model, 2019. [12] T. Feng, C. Jin, J. Liu, K. Zhu, H. Tu, Z. Cheng, G. Lin, and J. You. How far are we from agi: A...
https://arxiv.org/abs/2505.20438v1
S. Zhang, S. Vandenhende, S. Batra, S. Whitman, S. Sootla, S. Collot, S. Gururangan, S. Borodinsky, T. Herman, T. Fowler, T. Sheasha, T. Georgiou, T. Scialom, T. Speckbacher, T. Mihaylov, T. Xiao, U. Karn, V . Goswami, V . Gupta, V . Ramanathan, V . Kerkez, V . Gonguet, V . Do, V . V ogeti, V . Albiero, V . Petrovic, W...
https://arxiv.org/abs/2505.20438v1
Parks, N. White, N. Bawa, N. Singhal, N. Egebo, N. Usunier, N. Mehta, N. P. Laptev, N. Dong, N. Cheng, O. Chernoguz, O. Hart, O. Salpekar, 11 O. Kalinli, P. Kent, P. Parekh, P. Saab, P. Balaji, P. Rittner, P. Bontrager, P. Roux, P. Dollar, P. Zvyagina, P. Ratanchandani, P. Yuvraj, Q. Liang, R. Alao, R. Rodriguez, R. Ay...
https://arxiv.org/abs/2505.20438v1
[24] P. Liang, R. Bommasani, T. Lee, D. Tsipras, D. Soylu, M. Yasunaga, Y . Zhang, D. Narayanan, Y . Wu, A. Kumar, B. Newman, B. Yuan, B. Yan, C. Zhang, C. Cosgrove, C. D. Manning, C. Ré, D. Acosta-Navas, D. A. Hudson, E. Zelikman, E. Durmus, F. Ladhak, F. Rong, H. Ren, H. Yao, J. Wang, K. Santhanam, L. Orr, L. Zheng, ...
https://arxiv.org/abs/2505.20438v1
T. Lopez, R. Lowe, P. Lue, A. Makanju, K. Malfacini, S. Manning, T. Markov, Y . Markovski, B. Martin, K. Mayer, A. Mayne, B. Mc- Grew, S. M. McKinney, C. McLeavey, P. McMillan, J. McNeil, D. Medina, A. Mehta, J. Menick, L. Metz, A. Mishchenko, P. Mishkin, V . Monaco, E. Morikawa, D. Mossing, T. Mu, M. Murati, O. Murk, ...
https://arxiv.org/abs/2505.20438v1
gems in early layers: Accelerating long-context llms with 1000x input token reduction, 2024. [35] H. Sun, L.-W. Chang, W. Bao, S. Zheng, N. Zheng, X. Liu, H. Dong, Y . Chi, and B. Chen. Shadowkv: Kv cache in shadows for high-throughput long-context llm inference, 2025. [36] J. Tang, Y . Zhao, K. Zhu, G. Xiao, B. Kasikc...
https://arxiv.org/abs/2505.20438v1
arXiv:2505.20445v2 [cs.CL] 28 May 2025In-context Language Learning for Endangered Languages in Speech Recognition Zhaolin Li, Jan Niehues Karlsruhe Institute of Technology, Germany firstname.lastname@kit.edu Abstract With approximately 7,000 languages spoken worldwide, current large language models (LLMs) support only ...
https://arxiv.org/abs/2505.20445v2
generalization to low- resource languages. However, it comes with challenges such as increased computational cost, sensitivity to input ordering, and diminishing returns when context length exceeds the model’s optimal processing capacity. Beyond exploring approaches to boost ICL performance on low-resource languages, r...
https://arxiv.org/abs/2505.20445v2
address this, we investigate various sample selection strategies with explanations in Section 2.1. Once the prompt is provided, LLM generates text in the tar- get language with task-specific requirements to perform ICL. However, in the use case of ASR, this work focuses on not gen- erating text but on ranking different...
https://arxiv.org/abs/2505.20445v2
hypothesis selection is to select one hypothesis from the n-best lists generated by the initial ASR system. The traditional LLM-based approach to hypothesis selection is instruction-based, where the LLM is explicitly instructed to choose one from the provided options given the ICL samples; However, since the LLM is not...
https://arxiv.org/abs/2505.20445v2
languages, the model provides extensive linguistic cover- age and adaptability for low-resource settings [3]. In decoding, we implement beam search, and the 10 hypotheses with the best acoustic probabilities are chosen for language model selection. 3.3. Language Models In this work, we employ Llama-3-8B2as the large la...
https://arxiv.org/abs/2505.20445v2
or a combination of acoustic and text embed- dings results in lower performance compared to text embed- dings alone. This suggests that acoustic embeddings do not significantly contribute to retrieving high-quality ICL samples. A possible explanation is that these endangered languages are not closely related to those s...
https://arxiv.org/abs/2505.20445v2
this work is that probability-based selection requires access to model parameters, which restricts its applica- tion to open-source LLMs. Additionally, our experiments focus on four low-resource languages from different language fami- lies and with varying resource availability. While our findings are consistent across...
https://arxiv.org/abs/2505.20445v2
B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhari- wal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al. , “Lan- guage models are few-shot learners,” Advances in neural informa- tion processing systems , vol. 33, pp. 1877–1901, 2020. [7] A. F. Aji, G. I. Winata, F. Koto, S. Cahyawijaya, A. Romadhony, R. Mahend...
https://arxiv.org/abs/2505.20445v2
12 567–12 582. [Online]. Available: https://aclanthology.org/ 2023.emnlp-main.774/ [15] M.-H. Hsu, K. P. Huang, and H.-y. Lee, “Meta-whisper: Speech- based meta-icl for asr on low-resource languages,” arXiv preprint arXiv:2409.10429 , 2024. [16] Z. Li, M. Rind-Pawlowski, and J. Niehues, “Speech recognition corpus of th...
https://arxiv.org/abs/2505.20445v2
arXiv:2505.20451v1 [cs.CL] 26 May 2025AMULET : Putting Complex Multi-Turn Conversations on the Stand with LLM Juries Sahana Ramnath, Anurag Mudgil, Brihi Joshi, Skyler Hallinan, Xiang Ren Department of Computer Science, University of Southern California sramnath@usc.edu Abstract Today, large language models are widely ...
https://arxiv.org/abs/2505.20451v1
such con- text, how accurate are LLM-judges in predicting which response is better? We develop a framework, AMULET , that uses the following linguistic concepts for the same: (a) dialog-acts ( DA) to analyze the communicative struc- ture of each turn in the conversation, and (b) maxims to compare the preference respons...
https://arxiv.org/abs/2505.20451v1
unique to each instance “Does the response have simple vo- cabulary as requested in the question?” (Lin et al.; Kim et al.); these methods are under/over-specified respectively, and not scalable to complicated, multi- turn conversations. We overcome these issues by using linguistic theories that are pertinent to the ev...
https://arxiv.org/abs/2505.20451v1
benchmarks that have the least or lowest contamination with the models that we select to evaluate, to provide a fairer evaluation. Lastly, several preference benchmarks are often synthetic in nature (Dong et al., 2024a; Singh et al., 2025). We aim to select benchmarks that include realistic human-assistant conversation...
https://arxiv.org/abs/2505.20451v1
have an effective and cooperative inter- action. Recently, Miehling et al. (2024) extended Grice’s maxims to human-AI interactions. They broke down the original four maxims into multi- ple sub-parts that have clear definitions ; for ex- ample, Quantity became Quantity-1 (the response 3 Req Info PQ SQ Instr Ans Apolg Su...
https://arxiv.org/abs/2505.20451v1
‘Allo-Feedback’ (regarding the addressee’s process- ing of the previous turn) (Figure 7). The most com- mon functions vary slightly across datasets (Fig- ures 3(a), 6). In general, human turns show a high proportion of information seeking functions such as questions and requests, social functions such as thanking, and ...
https://arxiv.org/abs/2505.20451v1
change in DA. This serves to show that humans change the type of their requests frequently, sub- sequently requiring assistants to adapt and satisfy their varying demands. For all datasets, we show the % of consecutive turns with different DAin Figure 3 (c). Next, for the consecutive human turns with DAchanges, we calc...
https://arxiv.org/abs/2505.20451v1
models the prompt, followed by the con- versation, followed by the preference responses. However, prior works (Zheng et al., 2023) have noted that models exhibit position bias , that is, they tend to pick the first presented option as the an- swer. To prevent this, we use refer to prior work (Zheng et al., 2023) and fo...
https://arxiv.org/abs/2505.20451v1
that we chose RMs which have not been trained on any of the datasets we use for evaluation. 4.3 Our Methods AMULET .We use AMULET -DAandAMULET - MAXIM with the aforementioned two-vote system. 3INF-ORM-Llama3.1-70B , QRM-Gemma-2-27B , Skywork-Reward-Llama-3.1-8B-v0.2 6 Dataset → HH-T EST WILDF. N ECTAR HH-T RAIN WILDF. ...
https://arxiv.org/abs/2505.20451v1
to DA- THEN -MAXIM -THEN -RM ; whenever DA-then- MAXIM ends in a tie, we use the scores provided by the RM at hand to make the final decision (refer Figure 5). 4.4 Results In Table 2, we present the accuracies obtained by all judges, RMs and juries for the four datasets. AMULET .Firstly, in almost all datasets/settings...
https://arxiv.org/abs/2505.20451v1
/ 10.4 / 12.7 80.2 / 5.8 / 14.0 NECTAR 76.9 / 10.5 / 12.6 79.1 / 7.3 / 13.5 80.5 / 5.1 / 14.4 HH-T RAIN 60.3 / 15.1 / 24.6 63.5 / 9.5 / 27.0 66.2 / 4.7 / 29.1 Table 4: Win-Tie-Loss statistics observed for an AMULET -judge ( DA) and two AMULET -LM-J URY’s. The juries have higher win and loss rates and a lower tie rate a...
https://arxiv.org/abs/2505.20451v1
of training data used for pre-training, fine-tuning and aligning large language models (which are mostly not open sourced), data contamination is a massive problem in NLP today (Sainz et al., 2023). Most relevantly, several reward models submitted to the RewardBench leaderboard (Lambert et al., 2025) were contaminated ...
https://arxiv.org/abs/2505.20451v1
Rawte et al., 2023; Huang et al., 2023); in this work, it is possi- ble for the judges to hallucinate dialog acts and/or details in the conversation when performing the task. We report statistics of hallucination of dialog acts in Appendix I.2; since we rigorously stress in our prompts in Appendix F, G that the judge s...
https://arxiv.org/abs/2505.20451v1
we follow the conservative approach used in prior works (Zheng et al., 2023; Wang et al., 2024b; Hou et al., 2024; Qin et al., 2024) and present pref- erence responses to the LLM-judge in all possible orders and aggregate all the judgments to make the final decision; we also include in all our prompts a reminder to the...
https://arxiv.org/abs/2505.20451v1
keys. We use the following models from hugging- 10 face: Skywork-Reward-Llama-3.1-8B-v0.2 , QRM-Gemma-2-27B ,INF-ORM-Llama3.1-70B , Qwen/Qwen2.5-32B-Instruct . All these models are open-sourced; to access QWEN , we had to provide our contact information and accept the terms and conditions. All data and models were used...
https://arxiv.org/abs/2505.20451v1
judgement bias. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages 8301–8327. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger,...
https://arxiv.org/abs/2505.20451v1
Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and 1 oth- ers. 2023. A survey on hallucination in large lan- guage models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232 . Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clar...
https://arxiv.org/abs/2505.20451v1
International Conference on Machine Learning . Sangkyu Lee, Sungdong Kim, Ashkan Yousefpour, Min- joon Seo, Kang Min Yoo, and Youngjae Yu. 2024. Aligning large language models by on-policy self- judgment. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers...
https://arxiv.org/abs/2505.20451v1
Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computa- tional Linguistics , pages 311–318. Junsoo Park, Seungyeon Jwa, Ren Meiying, Daeyoung Kim, and Sanghyuk Choi. 2024. Offsetbias: Lever- aging debiased data for tunin...
https://arxiv.org/abs/2505.20451v1
dialog evaluation metrics. In The First Workshop on Evaluations and Assessments of Neural Conversation Systems , pages 15–33, Online. Association for Computational Lin- guistics. Lei Yu, Meng Cao, Jackie Chi Kit Cheung, and Yue Dong. 2024. Mechanistic understanding and miti- gation of language model non-factual halluci...
https://arxiv.org/abs/2505.20451v1
score levels or reasoning, or can just request for a score between say 1-10), (3) natu- ral language feedback that can be used to improve language models (Akyurek et al., 2023). Current work in LLM-judges also include ju- ries of judges (Lin et al.; Verga et al., 2024; Jung et al., 2024), training smaller language mode...
https://arxiv.org/abs/2505.20451v1
et al., 2025) has served as a solid bench- mark for reward models, and has standardized the evaluation of reward models. Most of these re- ward models are classifiers (Minghao Yang, 2024; Dorka, 2024; Liu et al., 2024; Wang et al., 2024a; Lou et al., 2024), producing an unbounded numeri- cal score; the higher the score...
https://arxiv.org/abs/2505.20451v1
big to be accommodated in a table. Instruction: You will be given a dialog conversation between a human user and an LLM assistant. The dialog is split into turns - a turn is defined as an utterance by either the human user or the assistant. Note that the roles of “speaker” (S) and “addressee” (A) will alternate at ever...
https://arxiv.org/abs/2505.20451v1
Meaning: Speaker’s processing of previous utterances, Func: Auto-Positive, Auto-Negative (3) Dim: Allo-Feedback , Meaning: Speaker expressing opinions about addressee’s processing of previous utterances, Func: Allo-Positive, Allo-Negative, Feedback Elicitation (4) Dim: Time Management , Meaning:Concerning the allocatio...
https://arxiv.org/abs/2505.20451v1
let A know that certain information that A wants to know, and concerning which A holds an uncertain belief, is incorrect. ∗Inform: Function performed by S, in order to make the information contained in the semantic content known to A; S assumes that the information is correct. ∗Agreement: Function performed by S, in or...
https://arxiv.org/abs/2505.20451v1
on certain conditions that S makes explicit. ∗Decline Request: Function by which S refuses to perform an action that S has been requested to perform, possibly depending on certain conditions that S makes explicit.∗Accept Suggest: Function by which S commits to perform an action that was suggested, possibly with certain...
https://arxiv.org/abs/2505.20451v1
the required JSON output format. Say “1” if you think Assistant-1’s response is better and “2” if you think Assistant-2’s response is better. Example Dialog - Human: Human’s turn Assistant: Assistant’s turn Human: Human’s turn Assistant: Assistant’s turn Human: Human’s turn Assistant-1: Assistant’s turnbyAssistant-1 As...
https://arxiv.org/abs/2505.20451v1
order to acknowledge that S is aware of A’s presence, and of A having signalled A’s presence to S. ∗Initial Self-Introduction: Function performed by S, in order to make A know about S; S puts pressure on A to acknowledge this. ∗Return Self-Introduction: Function performed by S, in order to make A know about S in respon...
https://arxiv.org/abs/2505.20451v1
quietly for a few minutes without thinking of anything. <SEP> {“Dim”: “Time Management”, “Func”: “Stalling”} {“Dim”: “Task”, “Func”: “Answer”}Human: any other ideas? they are fidgeting <SEP> {“Dim”: “Allo-Feedback”, “Func”: “Allo-Negative”} {“Dim”: “Task”, “Func”: “Set Question & Inform”} Assistant-1: Kids are great at...
https://arxiv.org/abs/2505.20451v1
““{“Dim”: “Time Management”, “Func”: “Stalling”} {“Dim”:“Task”, “Func”: “Answer”}””, ““Human: any other ideas? they are fidgeting””: ““{“Dim”: “Allo-Feedback”, “Func”: “Allo-Negative”} {“Dim”: “Task”, “Func”: “Set Question & Inform”}””, ““Assistant-1: Kids are great at fidgeting, so let’s make sure we give them permiss...
https://arxiv.org/abs/2505.20451v1
endorsement with requests that are harmful or unethical. (10) Transparency-1: The response should recognize the speaker’s knowledge boundaries, making clear any limitations in expertise, evidence, experience, or context. (11) Transparency-2: The response shouldrecognize the speaker’s operational capabilities, highlight...
https://arxiv.org/abs/2505.20451v1
harbormaster from your yacht club. If you get a hold of that person, you’ll also be Human:Thanks. And what are some typical forms required? Assistant-1: Oh, I’m afraid I’m not very familiar with these processes myself. If you want, I can look it up for you, or you can ask me to do so when you need the information. Assi...
https://arxiv.org/abs/2505.20451v1
bird, making it irrelevant and lacking in quality. Assistant-2’s response, while also not directly addressing the situation, at least suggests an action related to the previous mention of a link, making it slightly more relevant and organized. Neither response provides sufficient information or clarity about what to do...
https://arxiv.org/abs/2505.20451v1
since these analyses are used to understand the data. Further, since we run each instance twice to avoid position bias, it is possible that the dimen- sions and functions generated for each turn slightly vary across the two turns due to variability of the model (note that for Section 3, we only consider the dimensions ...
https://arxiv.org/abs/2505.20451v1
, number of turns and consecutive turns, % of turns with valid dimension/function predictions, and % of turns across votes which have the same dimensions and functions. Turns Type HH-T RAIN HH-T EST NECTAR WILDFEEDBACK Human TurnsFunction 0.96 0.96 0.96 0.89 Dimension 0.62 0.64 0.52 0.36 Assistant TurnsFunction 0.92 0....
https://arxiv.org/abs/2505.20451v1
2.919.8 0.1 0.1 0.8 0.2 0.0 0.0 0.0Wild FeedbackHuman Assistant Figure 7: Frequency of most common dimensions in HH-T RAIN , HH-T EST, NECTAR and W ILDFEEDBACK 12345678910 x02000400060008000# Conversations81807891 6766 4011 1017 337134 54 17 8Anthropic Train 12345678910 x0100200300400# Conversations460442 380 242 61 19...
https://arxiv.org/abs/2505.20451v1
of the two responses are dif- ferent but still similarly plausible, maxims play a large role in determining which response is more preferred. Lastly in Table 20 we present two examples for which AMULET -DAandAMULET -MAXIM both result in a tie, and W-E XPLbreaks the tie by choos- ing the rejected response; these example...
https://arxiv.org/abs/2505.20451v1
with I/O,W-E XPL,DA and MAXIM prompts. We experimented with varying temperatures for the API models and top-p sampling for QWEN , but ultimately they did not yield better results. •The reward models we use are Skywork-Reward-Llama-3.1-8B-v0.2 , QRM-Gemma-2-27B ,INF-ORM-Llama3.1-70B . We ran the first two with four NVID...
https://arxiv.org/abs/2505.20451v1
dialog acts and in the second example the two responses have significantly different dialog acts. In both cases, the maxims break the tie to make a decision as to which is the better response. 32 Example Comments Human: How do I increase my vocabulary? Assistant: Your vocabulary is your “bag of words”, and it’s useful ...
https://arxiv.org/abs/2505.20451v1
make the waffles? How much batter should you add to the waffle iron? Will you be making regular waffles, with holes, or more like the Belgian type? <SEP> “Dim”: “Allo -Feed back”, “Func”: “Allo -Positive” “Dim”: “Task”, “Func”: “Set Ques tion”AMULET -DAandAMULET -MAXIM both result in a tie, and W-E XPL chooses Assistan...
https://arxiv.org/abs/2505.20451v1
arXiv:2505.20464v1 [cs.HC] 26 May 2025The Impact of a Chatbot’s Ephemerality-Framing on Self-Disclosure Perceptions Samuel Rhys Cox srcox@cs.aau.dk Aalborg University Aalborg, DenmarkRune Møberg Jacobsen runemj@cs.aau.dk Aalborg University Aalborg, DenmarkNiels van Berkel nielsvanberkel@cs.aau.dk Aalborg University Aal...
https://arxiv.org/abs/2505.20464v1
the Internet ” phenomena has emerged [ 32,72] in which people may feel more comfortable disclosing to strangers online. Alongside this, conversational agents (CAs) are increasingly used as CUI ’25, July 8–10, 2025, Waterloo, ON, Canada Samuel Rhys Cox, Rune Møberg Jacobsen, and Niels van Berkel a means for people to sh...
https://arxiv.org/abs/2505.20464v1
40]. Despite these advantages, people may hesitate to self-disclose due to fears of negative outcomes, such as being judged, stereotyped, or perceiving that the intended recipient is not receptive to their disclosure [ 97]. As posited by Social Penetration Theory, this is part of a risk-reward calculus where people wei...
https://arxiv.org/abs/2505.20464v1
post, as well as perceiving the au- dience of said post as being more intimate and close to themselves. Furthermore, Ma et al. found that when users posted ephemeral or anonymous questions on a question asking website, it led to reduced social costs in asking questions [69]. Beyond the relationship of those we are self...
https://arxiv.org/abs/2505.20464v1
could increase feelings of familiarity and foster sharing, there is also a tension raised with increased privacy concerns [ 47]. In relation to these privacy concerns, prior work has focused on giving users greater levels of control over a CA’s memory, such as choices regarding data retention by the agent [ 58,82,95]. ...
https://arxiv.org/abs/2505.20464v1
stated (i.e., “[...] will help me stay free from prior judgement ”).Benefits pertinent to persistent conversations are stated (i.e., “[...] will help me get to know you better ”). Emphasising (im)permanence The chatbot’s language emphasises the ephemeral nature of the conversation, and language makes no reference or al...
https://arxiv.org/abs/2505.20464v1
(e.g., topics related to interpersonal rela- tionships and self-concept) as more sensitive to self-disclose [ 86]. Drawing on questions and topics from prior work [ 5,33,86],Fac- tual -disclosure consisted of questions related to hobbies and in- terests (e.g., “ What type of music do you enjoy listening to [...]?”), Th...
https://arxiv.org/abs/2505.20464v1
sessions, resulting in 63Stranger and 64 Familiar participants. 3.4 Procedure Participants followed the procedure below for both sessions: (1)Joining session: Participant directed from Prolific to Qual- trics (task named “ Talk to a chatbot about yourself (Part X of Two) ” on Prolific). Participant receives high-level ...
https://arxiv.org/abs/2505.20464v1
other in the self (IOS) scale, a validated pictorial scale [ 4,37]. Here, participants were asked to select the visual representation of two overlapping circles that best represented their relationship with the chatbot via the prompt “Please select the picture below which best describes your relationship with the chatb...
https://arxiv.org/abs/2505.20464v1
[ 64]: “I believe chatbots can be intelligent ”, “I believe chatbots can have real feelings ” on 7-point Likert scales. To measure participants’ tendency to disclose emotionally distressing experiences, we used the 12-item Distress Disclosure Index (DDI)[ 51,52], that includes items such as: “ When I feel upset, I usua...
https://arxiv.org/abs/2505.20464v1