text string | source string |
|---|---|
of Multilingual Code- switched Soap Opera Speech. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018) , Miyazaki, Japan. European Language Resources Association (ELRA). Digital Umuganda. 2023. Afrispeech kin- yarwanda male and female tts datasets. https://huggingface... | https://arxiv.org/abs/2505.18436v1 |
for Computational Linguistics , 11:250–266. Tu Anh Nguyen, Benjamin Muller, Bokai Yu, Marta R Costa-Jussa, Maha Elbayad, Sravya Popuri, Christophe Ropers, Paul-Ambroise Duquenne, Robin Algayres, Ruslan Mavlyutov, et al. 2025. Spirit- lm: Interleaved spoken and written language model. Transactions of the Association for... | https://arxiv.org/abs/2505.18436v1 |
languages to promote acces- sibility and linguistic diversity. Jörgen Valk and Tanel Alumäe. 2021. V oxlingua107: A dataset for spoken language recognition. In Pro- ceedings of the IEEE Spoken Language Technology Workshop (SLT) .Charl Van Heerden, Neil Kleynhans, and Marelie H. Davel. 2016. Improving the lwazi asr base... | https://arxiv.org/abs/2505.18436v1 |
from various publicly available Bemba sources, including books, show transcripts, and YouTube transcripts. It contains 15,000 utterances totaling 24.5 hours of audio, making it a valuable resource for ASR and linguistic research for the Bemba lan- guage. Mozilla Common Voice (Mozilla Foundation, 2023) is a multilingual... | https://arxiv.org/abs/2505.18436v1 |
2020) is a high-quality crowdsourced dataset of Yoruba audio recordings designed for speech processing applications. It includes transcribed WA V files, with separate archives for female and male speakers and the corresponding transcription. It is manually quality-checked and provides valuable resources for developing ... | https://arxiv.org/abs/2505.18436v1 |
in 10 African languages, showcasing the linguistic diversity of the continent. The included languages are Edo, Tamazight, Yoruba, Swahili, Hausa, Tiv, Shona, Ibibio, Igbo, and Nigerian Pidgin. UDHR (Universal Declaration of Human Rights Audio, 2025): The website UDHR.audio hosts raw audio recordings of the Universal De... | https://arxiv.org/abs/2505.18436v1 |
evaluate two variants: whisper-large-v3 , which of- fers high accuracy for multilingual ASR tasks, and Language ISO-3 Hours Dataset Breakdown (Color-coded) Afrikaans afr 255.3 (, 4.28) ( , 138.73) ( , 3.31) ( , 0.15) ( , 108.39) ( , 0.43) Akuapim-twi aka 98.92 (, 60.57) ( , 38.35) Asante-twi aka 31.96 (, 1.53) ( , 30.4... | https://arxiv.org/abs/2505.18436v1 |
Yoruba yor 738.31 (, 614.98) ( , 0.12) ( , 94.05) ( , 4.03) ( , 25.13) Standard Moroccan Tamazight zgh 1.07 (, 1.07) Zulu zul, eng-zul 197.24 (, 4.28) ( , 187.5) ( , 0.01) ( , 5.45, CS - English) Multiple⋆142.42 (, 142.42) English - Accented eng 200 (, 200) Table B.1: Total duration of audio (in hours) available per la... | https://arxiv.org/abs/2505.18436v1 |
LD PT – – Taita dav – – – – – – Tem kdh – – STTSLD – – – Tigre tig – – LD – PT – Tigrinya tir – – STTSLD – – – Tiv tiv – – LD – – – Tonga (Zambia) toi – – LD PT – – Tsonga tso – – STTSLD PT – – Tswana tsn – – LD PT PT – Twi twi – – – – – – Venda ven – – LD PT – – Western Maninkakan mlq – – LD – – – Wolof wol – – STLD P... | https://arxiv.org/abs/2505.18436v1 |
learning rate of 5×10−5over 30 epochs.We use the HuggingFace Transformers (Wolf et al., 2020) for training and evaluation.7For TTS mod- els, we adopt the finetuning procedure outlined in the Vits repository and follow the default hyperpa- rameter configuration provided in the repository8. Evaluation Metrics. For ASR, w... | https://arxiv.org/abs/2505.18436v1 |
Lwazi 70.04/29.11 132.03/86.73 248.55/193.15 110.71/52.4 61.59/17.94 38.2/10.41 18.63/7.24 102.48/54.0 31.81/11.55 S. Sotho (sot) NCHTL 79.97/27.48 154.26/111.44 743.88/591.26 145.42/113.09 23.94/6.31 26.84/6.87 18.15/5.58 44.74/12.54 24.47/7.3 Serer (srr) Kallaama 105.41/69.85 255.33/233.38 1046.88/977.99 479.84/571.0... | https://arxiv.org/abs/2505.18436v1 |
22.64 15.59 Nigerian Pidgin (pcm) OlogoAfrica 73.68 74.32 Shona (sna) OlogoAfrica 90.91 92.34 Shona (sna) V oxLingua 86.61 88.23 Somali (som) V oxLingua 97.96 95.54 Swahili (swa, swh) OlogoAfrica 99.03 94.14 Swahili (swa, swh) UDHR 99.60 94.29 Swahili (swa, swh) V oxLingua 99.96 94.29 Tiv (tiv) OlogoAfrica 66.67 69.93 ... | https://arxiv.org/abs/2505.18436v1 |
arXiv:2505.18440v1 [cs.CL] 24 May 2025Efficient Long CoT Reasoning in Small Language Models Zhaoyang Wang1†Jinqi Jiang1†Tian Qiu1†Hui Liu2 Xianfeng Tang2Huaxiu Yao1 1University of North Carolina at Chapel Hill2Amazon {zhaoyang,huaxiu}@cs.unc.edu Abstract Recent large reasoning models such as DeepSeek-R1 exhibit strong ... | https://arxiv.org/abs/2505.18440v1 |
since SLMs have relatively poor capability and generalization. 1 To address this issue, existing works propose to use heuristic rules such as minimum reason- ing length with correct final answer (Chen et al., 2025), design length based rewards for reinforce- ment learning (Aggarwal and Welleck, 2025; Yi and Wang, 2025;... | https://arxiv.org/abs/2505.18440v1 |
Thought. Chain of thought (CoT) rea- soning has been widely adopted to enable LLMs to perform reasoning in a step-by-step manner (Wei et al., 2022a,b; Kojima et al., 2022; Wang et al., 2023a; Zhou et al., 2023). Recently, the emergence of large reasoning models such as OpenAI’s o1 series (OpenAI., 2024), DeepSeek-R1 (G... | https://arxiv.org/abs/2505.18440v1 |
et al., 2025). The goal of this paper is to enable open-source SLMs to take advantage of long CoT reasoning ability while maintaining efficiency. To achieve this, most of existing works introduce length based penalties into the reward function (Aggarwal and Welleck, 2025; Yang et al., 2025; Yi and Wang, 2025), then tra... | https://arxiv.org/abs/2505.18440v1 |
3 On-policy prompt : Give the final answer to the problem directly based on the provided thinking steps . Question: {Question} Thinking: {Thinking steps} <think> Step1, … Step N/2 </think> Binary Cutting Target SLMThe final answer is *** Answer check <think>Step1, … Step (k+2k)/2 </think> Backtracking Target SLMThe fin... | https://arxiv.org/abs/2505.18440v1 |
prefix is accepted only ifϕ= 1under M. By relying on the target model’s own outputs rather than an external judge model, we ensure that the distilled CoT segments align with the SLM’s native reasoning capacity. This on-policy mechanism overlooked by prior methods yields a more coherent long CoT segment, since each reta... | https://arxiv.org/abs/2505.18440v1 |
(Cobbe et al., 2021), MATH (Hendrycks et al., 2021), and AIME (MAA, 2024). GSM8k is a primary school level mathematical dataset re- quiring basic arithmetic and logic. MATH we used is a widely used subset of original dataset which contains 500 challenging high school competition- level math problems. AIME consists of e... | https://arxiv.org/abs/2505.18440v1 |
show obvious better per- formance than other methods which often need less than half of the generation. Figure 3 clearly shows that about 40% long CoT data can be stream- lined over 50% redundant reasoning steps by our method. This supports the perspective that long CoT contains redundant steps especially for the easy ... | https://arxiv.org/abs/2505.18440v1 |
↓)2615 (58.69% ↓) SFT+DPO FCS 78.24 (13.42% ↓)524 (48.17% ↓)49.40 (22.81% ↓)1914 (29.42% ↓)21.22 (25.57% ↓)4131 (34.74% ↓) SFT Ours 89.16 (1.34% ↓)382 (62.22% ↓)61.20 (4.38% ↓)901 (66.78% ↓)25.51 (10.52% ↓)3021 (52.27% ↓) SFT+DPO Ours 89.92 (0.50% ↓)278 (72.50% ↓)56.60 (11.56% ↓)489 (81.97% ↓)21.54 (24.45% ↓)1836 (71.0... | https://arxiv.org/abs/2505.18440v1 |
SFT Oursand SFT+DPO Oursrespectively. Model MethodGSM8K MATH AIME Acc (%) #Token Acc (%) #Token Acc (%) #Token LlamaSFT Ours 87.34 502 54.00 2322 18.01 5480 SFT Random 85.44 (2.18% ↓)995 (98.21% ↑)47.20 (12.59% ↓)2894 (24.63% ↑)8.68 (51.81% ↓)5771 (5.31% ↑) SFT Qwen data 87.19 (0.17% ↓)411 (18.13% ↓)52.8 (2.22% ↓)1896 ... | https://arxiv.org/abs/2505.18440v1 |
2)Our work focuses on the distillation scenario, which relies on a large reasoning model to provide high-quality reasoning traces. We do not consider reinforcement learning or self- training strategies for the SLM. While distilla- tion offers high efficiency, alternative training paradigms are still valuable and comple... | https://arxiv.org/abs/2505.18440v1 |
MAA. 2024. American invitational mathematics exami- nation — aime. American Invitational Mathematics Examination – AIME 2024, February 2024. Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, Eric Malmi, and Aliaksei Severyn. 2023. Teaching small language models to reason. In Proceedings ofthe61st Annual Meeti... | https://arxiv.org/abs/2505.18440v1 |
Advances inNeural Information Processing Systems. Tong Wu, Chong Xiang, Jiachen T Wang, and Pra- teek Mittal. 2025. Effectively controlling reasoning models through thinking intervention. arXiv preprint arXiv:2503.24370. Heming Xia, Yongqi Li, Chak Tou Leong, Wenjie Wang, and Wenjie Li. 2025. Tokenskip: Control- lable ... | https://arxiv.org/abs/2505.18440v1 |
5-point scoring system described below. Points are accumulated based on the satisfaction of each criterion: - Add 1 point if the CoT is relevant and initiates a reasoning process related to the user’s inquiry, even if it is incomplete or contains some irrelevant steps. - Add another point if the CoT addresses a substan... | https://arxiv.org/abs/2505.18440v1 |
think it involves the total sum of all the numbers in the square divided by the number of rows (or columns), since each row needs to add up to the magic sum. Let me check if that makes sense. The total sum of numbers from 1 to n2can be calculated using the formula for the sum of the first mintegers:m(m+1) 2. Here, misn... | https://arxiv.org/abs/2505.18440v1 |
me think how to compute that. Dividing by 0.31 is the same as multiplying by1 0.31. Let me calculate 1÷0.31. Well, 0.31×3 = 0 .93, and 0.31×3.2 would be 0.992. So 1÷0.31is approximately 3.2258. But I should probably do this more accurately. Let’s do the division: 1÷0.31. Alternatively, I can write 0.31 as a fraction, w... | https://arxiv.org/abs/2505.18440v1 |
BRIT: Bidirectional Retrieval over Unified Image-Text Graph Ainulla Khan*Yamada Moyuru*Srinidhi Akella Fujitsu Research India {ainulla.khan, yamada.moyuru, akella.srinidhi}@fujitsu.com Abstract Retrieval-Augmented Generation (RAG) has emerged as a promising technique to enhance the quality and relevance of responses ge... | https://arxiv.org/abs/2505.18450v1 |
when the retrieved information is provided as input. The existing methods (Chen et al., 2022; Yang et al., 2023; Sharifymoghaddam et al., 2024) often rely on training or fine-tuning techniques to ad- dress these limitations, but these approaches may not be effective in the enterprise settings, partic- ularly when they ... | https://arxiv.org/abs/2505.18450v1 |
questions that necessitate cross-modal, multi-hop retrieval to identify the key information. Our main contributions can be summarized as follows: •We propose BRIT, a novel multi-modal RAG framework integrates diverse text-image links within a unified graph. This enables effec- tive retrieval of relevant texts and image... | https://arxiv.org/abs/2505.18450v1 |
multi-modal graph retrieval have not been explored enough. 2.3 Multi-Modal Graph and LMMs Several recent works (Yang et al., 2023; Yoon et al., 2023) attempt to use multi-modal graph for summa- rization task and QA task with Large Multi-modal Model (LMM). They train their neural networks on the specific datasets and te... | https://arxiv.org/abs/2505.18450v1 |
graph: (1) Text-to-image retrieval, and (2) Image-to-text retrieval. Sub-graph retrieval . Given Gd, to retrieve query- specific sub-graphs GqfromGd, we first extract named entities from the query using a 1-shot LLMprompt. Then, the entities of query and the tex- tual nodes of Gdare embedded using an embed- ding model ... | https://arxiv.org/abs/2505.18450v1 |
Ar -Br (B) R -MgBr (C) Ar -R (D) NiL2 Katalysezyklus der Kumada - Kupplung , The Kumada coupling employs both a nucleophilic alkylation step subsequent to the oxidative addition of the aryl halide (L = Ligand, Ar = Aryl)Captions chemical structure of triethyloxonium tetrafluoroborate, Triethyloxonium ……(A) The fastest ... | https://arxiv.org/abs/2505.18450v1 |
least 3 images and 1 section from a validation split. In more detail, average number of images and sections per sample is 5.61 and 8.83. Then, an image is randomly picked for each question type (3 images in total in each sample). The picked image is fed into a LMM (Large Multi-Modal Model) along with the relevant texts... | https://arxiv.org/abs/2505.18450v1 |
corresponding question without NEA. 5 Experimental Settings 5.1 Benchmark We use MM-RAG test set described in Sec.4 to evaluate RAGs on multi-modal documents con- taining texts and images. MM-RAG test set is a new test set we built based on WikiWeb2M (Burns et al., 2023) and contains 500 questions for 100 Wikipedia pag... | https://arxiv.org/abs/2505.18450v1 |
✓ 0.86 0.68 0.77 316.0 1.86 ✓ ✓ ✓ 0.88 0.72 0.80 261.0 1.83 Table 2: Recall ratio in Retrieval. We evaluate various text-image linking and their combinations in terms of the recall rate and the number of retrieved words and images. CA: Caption-based, LP: Layout-based (page), LS: Layout-based (section), and SI: Similari... | https://arxiv.org/abs/2505.18450v1 |
understood.", "singing. Some prominent examples of movie …Lloyd Hamilton → leading lady in → numerous comedies Irene Dalton → married → Lloyd Hamilton Irene Dalton → born on → September 1, 1901 The image caption states that the image is a still from the American comedy short film *Rolling Stones* (1922) and shows Lloyd... | https://arxiv.org/abs/2505.18450v1 |
image-to-text and text-to-image links ex- tracted from a document bidirectionally. This pa- per has comprehensively evaluated the effective- ness of the various links and their combinations for RAG on multi-modal document using a new test set, MM-RAG test set we built, which contains complex cross-modal multi-hop quest... | https://arxiv.org/abs/2505.18450v1 |
, abs/2404.16130. Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. 2024. Colpali: Efficient document retrieval with vi- sion language models. Preprint , arXiv:2407.01449. Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michi- hiro Yasunaga, and Yu Su. 2024. Hipporag:... | https://arxiv.org/abs/2505.18450v1 |
6), 2) Image-Text questions (Fig. 7), and 3) Image- Image questions (Fig. 8). We use GPT-4o (gpt-4o- 2024-05-13) for question generation. 12 C a p t i o n : { t e x t u a l _ c o n t e x t } You a r e p r o v i d e d wi th an image c a p t i o n and a c o r r e s p o n d i n g image . The q u e s t i o n s h o u l d be... | https://arxiv.org/abs/2505.18450v1 |
u t i t s h o u l d a l l o w t h e c a p t i o n t o g u i d e t h e u s e r t o f i n d t h e r e l e v a n t image . − Avoid s p e c u l a t i v e or ambiguous q u e s t i o n s . Ensure t h e q u e s t i o n and answer a r e l o g i c a l l y c o n n e c t e d t o bot h t h e c a p t i o n and t h e image . R e f e... | https://arxiv.org/abs/2505.18450v1 |
v e t r i a n g l e wi th Nate Cooper and Duncan S t e w a r t ( a c t o r B e n e d i c t Wall p i c t u r e d ) . ' A f t e r you g e n e r a t e d t h e q u e s t i o n , p l e a s e a l s o d e s c r i b e t h e c a p t i o n you used as a f a c t t o g e n e r a t e t h e q u e s t i o n as f o l l o w s : " used ... | https://arxiv.org/abs/2505.18450v1 |
**must be based p u r e l y on t h e **v i s u a l a s p e c t s **of t h e image ( e . g . , o b j e c t s , a t t i r e , s e t t i n g , people , background , e t c . ) . − The **answer **must be d e r i v e d onl y from t h e **c a p t i o n ** , and n o t from any v i s i b l e t e x t or v i s u a l a s p e c t s... | https://arxiv.org/abs/2505.18450v1 |
e s ' : [ ' ( A) Robbo ' , ' ( B) B e n e d i c t Wall ' , ' ( C) Ja k e Ryan ' , ' (D) Ricky Sharpe ' ] , ' answer ' : ' ( B) ' # The q u e s t i o n i s v i s u a l l y based , and t h e answer ( B e n e d i c t Wall ) comes from t h e ** c a p t i o n * *. R e f e r r e d c a p t i o n : ' The Way Way Back A u s t r... | https://arxiv.org/abs/2505.18450v1 |
l e a s e a l s o d e s c r i b e t h e c a p t i o n you used as a f a c t t o g e t t h e answer as f o l l o w s : " used t e x t u a l f a c t s " : ' P o l i s h f o o t b a l l p l a y e r Tomasz Frankowski P o l s k i , Frankowski i n 2010 ' The o u t p u t f o r m a t s h o u l d be : " q u e s t i o n " : g e ... | https://arxiv.org/abs/2505.18450v1 |
n f o r m a t i o n ** t h a t i s n o t v i s i b l e i n t h e image t o form t h e q u e s t i o n or answer . − The image s h o u l d be i d e n t i f i a b l e from t h e q u e s t i o n . R e f e r t o t h e examples below f o r b e t t e r u n d e r s t a n d i n g : Example 1 : " q u e s t i o n " : " What c o ... | https://arxiv.org/abs/2505.18450v1 |
arXiv:2505.18451v1 [cs.LG] 24 May 2025µ-MoE: Test-Time Pruning as Micro-Grained Mixture-of-Experts Toshiaki Koike-Akino1Jing Liu1Ye Wang1 Abstract To tackle the huge computational demand of large foundation models, activation-aware compression techniques without retraining have been intro- duced. However, since these r... | https://arxiv.org/abs/2505.18451v1 |
LLMs. We consider such a mixture of micro- experts, referred to as µ-MoE. µ-MoE employs test-time adaptation to reduce the total test-time compute, i.e., online dynamic pruning to reduce the number of active weights for inference computation. Besides the computational efficiency, the online dynamic pruning may potentia... | https://arxiv.org/abs/2505.18451v1 |
ity, while Wanda is a viable candidate. The total complexity with online Wanda pruning is O[3dd′+dT+ρdd′T]. The complexity ratio (compared to the original O[dd′T]) is on the order of: 3dd′+dT+ρdd′T dd′T=ρ+3 T+1 d′≃ρ,(T, d′≫1). This suggests that instant Wanda pruning for every prompt has almost no additional complexity... | https://arxiv.org/abs/2505.18451v1 |
20.7 20.5 Wanda (PTB Calib) 10.9 15.2 14.2 13.4 13.4 16.8 17.4 15.9 20.6 20.5 24.6 21.9 Wanda (C4 Calib) 10.9 15.2 14.2 13.4 13.4 16.8 17.4 15.9 20.6 20.5 24.6 21.9 µ-MoE 10.6 15.0 12.3 12.7 11.5 16.4 12.9 13.6 14.3 20.2 14.6 16.4 3. Expriments Experiments Setup We conduct experiments for LLM benchmarks to evaluate the... | https://arxiv.org/abs/2505.18451v1 |
40% 42.81 27.90 40.36 43.60 37.08 38.40 40.16 37.05 39.05 Wanda 40% 32.99 28.23 34.73 30.16 31.68 35.47 33.22 31.05 32.45 µ-MoE 40% 45.16 35.21 37.64 44.62 37.43 40.07 43.28 37.24 41.12 Table 3. Accuracy in percent ( ↑) on TextVQA dataset of LLaV A- 7B model with different compression methods at 40–60% active weights. ... | https://arxiv.org/abs/2505.18451v1 |
Beyond efficiency: A systematic survey of resource-efficient large language models. arXiv preprint arXiv:2401.00625 , 2024a. Bai, G., Li, Y ., Ling, C., Kim, K., and Zhao, L. SparseLLM: Towards global pruning for pre-trained language models. arXiv preprint arXiv:2402.17946 , 2024b. Bansal, H., Gopalakrishnan, K., Dingl... | https://arxiv.org/abs/2505.18451v1 |
S., Ye, S., Wang, S., Yu, S., Zhou, S., Pan, S., Li, S. S., Zhou, S., Wu, S., Ye, S., Yun, T., Pei, T., Sun, T., Wang, T., Zeng, W., Zhao, W., Liu, W., Liang, W., Gao, W., Yu, W., Zhang, W., Xiao, W. L., An, W., Liu, X., Wang, X., Chen, X., Nie, X., Cheng, X., Liu, X., Xie, X., Liu, X., Yang, X., Li, X., Su, X., Lin, X... | https://arxiv.org/abs/2505.18451v1 |
with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301 , 2023. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., Chen, W., et al. LoRA: Low-rank adaptation of large language models. ICLR , 1(2):3, 2022. Hua, W., Zhou, Y ., De Sa, C. M., Zhang, Z., and Suh, G. E. Cha... | https://arxiv.org/abs/2505.18451v1 |
to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS) , 2022. Ma, X., Fang, G., and Wang, X. LLM-Pruner: On the structural pruning of large language models. Advances in neural information processing systems , 36:2170... | https://arxiv.org/abs/2505.18451v1 |
arXiv:2404.13628 , 2024. Xie, Z., Ma, Y ., Zheng, X., Chao, F., and Ji, R. Automated fine-grained mixture-of-experts quantization. Xu, C. and McAuley, J. A survey on model compression and acceleration for pretrained language models. In Proceed- ings of the AAAI Conference on Artificial Intelligence , volume 37, pp. 105... | https://arxiv.org/abs/2505.18451v1 |
demonstrated the existence of prompt-dependent and task-specific sparsity in LLMs. Dynamic Pruning Dynamic networks (Lin et al., 2017; Liu & Deng, 2018; Hua et al., 2019; Gao et al., 2018; Chen et al., 2019) selectively execute a subset of modules at inference time based on input samples. Typically, module selections a... | https://arxiv.org/abs/2505.18451v1 |
implementation on hardware. An empirical experiment is shown in Figure 3, where the average runtime for Wanda pruning over different embedding size dis measured on Apple M1 CPU and NVIDIA A100 GPU. It does not include the computation of linear affine transforms after weight pruning. We see that torch.topk and torch.kth... | https://arxiv.org/abs/2505.18451v1 |
subset “en”, containing 364,868,892 and 364,608 samples for train and validation splits, respectively, while we use the first shard for each split in https://huggingface.co/datasets/allenai/ c4. ScienceQA ScienceQA (Lu et al., 2022) is collected from elementary and high school science curricula (i.e., grades 1 through ... | https://arxiv.org/abs/2505.18451v1 |
arXiv:2505.18452v1 [cs.CL] 24 May 2025Preprint. Under review. MedScore : Factuality Evaluation of Free-Form Medical An- swers Heyuan Huang, Alexandra DeLucia, Vijay Murari Tiyyala & Mark Dredze∗ Center for Language and Speech Processing Johns Hopkins University Baltimore, MD 21218, USA {hhuan134, aadelucia, vtiyyal1, m... | https://arxiv.org/abs/2505.18452v1 |
require a more sophisticated approach. For example, recent work has demonstrated that LLMs can effectively answer patient medical questions (Ayers et al., 2023b). Allen et al. (2024) identified several factors for evaluation of patient-facing systems, including personalization, perceived empathy (Ayers et al., 2023a), ... | https://arxiv.org/abs/2505.18452v1 |
(2024) provide reasons for why a claim can be “unverifiable”, but the reasons are general and not formalized. 2 Preprint. Under review. Answer the question based on the given context. “Time to return to work and surgeons' recommendations after carpal tunnel release.” Time to return to work after carpal tunnel release i... | https://arxiv.org/abs/2505.18452v1 |
the specifics of y our sur gery. Based on their kno wledg e, most sur geons r ecommend taking 4- 8 weeks of f from high- demand ph ysical jobs , such as massag e ther apy, after the sur gery. Most surgeons recommend taking 4-8 weeks off from high-demand physical jobs after ganglion removal surgery. High-demand physical... | https://arxiv.org/abs/2505.18452v1 |
- Your concerns are about tetanus. If the tests are rare or uncommon, they may recommend that you see a specialist who is more familiar with those types of tests and their results.- The tests are rare. Hallucinated claims - The tests are uncommon. Context-dependent (vague reference) - They may recommend that you see a ... | https://arxiv.org/abs/2505.18452v1 |
use this decomposition method to transform sentences into atomic claims that can then be verified in a decompose- then-verify factuality evaluation system. We also consider three verification methods that differ by the source for claim-level verification: model internal knowledge, a customized reference corpus, and a m... | https://arxiv.org/abs/2505.18452v1 |
5 Preprint. Under review. Internal Knowledge. We prompt a large, general-purpose model, GPT-4o , and an open- sourced medical-specialized model, Llama3-OpenBioLLM-70B (Ankit Pal, 2024), to use their own knowledge to determine whether a claim is True or False. The internal knowledge test prompt is in Appendix Table 11. ... | https://arxiv.org/abs/2505.18452v1 |
all decomposition strategies are in Table 2.6For the AskDocsAI we assume every response has at least one valid claim since the chatbot responses are high-quality, rewritten doctor responses (Section 3.1). When a decomposition method returns no claims for a response, this indicates that there are no 4After qualitative e... | https://arxiv.org/abs/2505.18452v1 |
molecular” fact levels. The average number of tokens of claims extracted by FActScore, MedScore , and VeriScore is 10.31, 12.84 , and 12.77, respectively. These claim lengths show that FActScore decomposes sen- tences into atomic facts while MedScore and VeriScore generate molecular facts by including more information ... | https://arxiv.org/abs/2505.18452v1 |
claims than when verifying against the external MedCorp. The scores drop by 32% at most, to 62-65%. In the external corpus verification, we find that Mistral Small 3, the backbone verifier LLM, has strong capabilities to understand medical passages, associate the claim with the passages, and reason about the truthfulne... | https://arxiv.org/abs/2505.18452v1 |
the adjusted verifiable rates in Table 5. MedScore is more robust to colloquial language. Similar to the chatbot response decompo- sition results, FActScore has the most claims while VeriScoreQA has the fewest claims in Table 2. The 0-claim rate of VeriScoreQA becomes extremely high, 53.69%, because PUMA answers are sh... | https://arxiv.org/abs/2505.18452v1 |
MedScore claims are verifiable and valid, but there are no corresponding annotated verifiable spans in the PUMA dataset, which results in an underestimate of its verifiable rate, as shown in Appendix Table 14. 5 Conclusion We proposed MedScore , which decomposes medical answers into verifiable claims while preserving m... | https://arxiv.org/abs/2505.18452v1 |
Alan Ritter, and Lu Wang (eds.), Proceedings of the 2025 Conference of the Na- tions of the Americas Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers) , pp. 3563–3599, Albuquerque, New Mexico, April 2025. Association for Computational Linguistics. ISBN 979-8... | https://arxiv.org/abs/2505.18452v1 |
Zoe Papakipos, Aaditya Singh, Aaron Grattafiori, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alex Vaughan, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andre... | https://arxiv.org/abs/2505.18452v1 |
Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victori... | https://arxiv.org/abs/2505.18452v1 |
of large language models. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of the Association for Computational Linguistics: EMNLP 2023 , pp. 5831–5847, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.388. URL https://aclanthology. org/2023.findings... | https://arxiv.org/abs/2505.18452v1 |
Ian Sohl, Ibrahim Okuyucu, Ikai Lan, Ilya Kostrikov, Ilya Sutskever, Ingmar Kanitscheider, Ishaan Gulrajani, Jacob Coxon, Jacob Menick, Jakub Pachocki, James Aung, James Betker, James Crooks, James Lennon, Jamie Kiros, Jan Leike, Jane Park, Jason Kwon, Jason Phang, Jason Teplitz, Jason Wei, Jason Wolfe, Jay Chen, Jeff ... | https://arxiv.org/abs/2505.18452v1 |
Wojciech Zaremba, Yash Patil, Yilei Qian, Yongjik Kim, Youlong Cheng, Yu Zhang, Yuchen He, Yuchen Zhang, Yujia Jin, Yunxing Dai, and Yury Malkov. Gpt-4o system card, 2024a. URL https://arxiv.org/abs/2410.21276 . OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diog... | https://arxiv.org/abs/2505.18452v1 |
Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Fe- lipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, ... | https://arxiv.org/abs/2505.18452v1 |
2022. Associa- tion for Computational Linguistics. doi: 10.18653/v1/2022.emnlp-main.340. URL https://aclanthology.org/2022.emnlp-main.340/ . Miriam Wanner, Benjamin Van Durme, and Mark Dredze. DnDScore: Decontextualization and Decomposition for Factuality Verification in Long-Form Text Generation, December 2024a. URL h... | https://arxiv.org/abs/2505.18452v1 |
various specialties and practitioner types. We excluded posts with fewer than 40 words to ensure substantive content and removed posts containing URLs to minimize reference dependencies. For posts with multiple physician responses, we selected either the earliest response or the most prominently marked one (e.g., from ... | https://arxiv.org/abs/2505.18452v1 |
doctor emphasized the importance of keeping your wrist elevated, as you’ve been doing, to help reduce the swelling. ClaimsSubclaim : Your doctor emphasized something. Decontextualized : The doctor who is responsible for your treatment emphasized something. Subclaim : The emphasis was on the importance of keeping your w... | https://arxiv.org/abs/2505.18452v1 |
newly-released models (Llama 3.1-8B, Mistral-7B, and Mistral Small 3) and select the LLM with the highest factuality score as our verifier model. The results are in Table Table 9. The doctor responses are decomposed by MedScore using GPT-4o-mini to ensure the highest claim quality. When verifying doctor responses’ clai... | https://arxiv.org/abs/2505.18452v1 |
you are unable to verify the claim, then output "False". user prompt: Using your own knowledge, answer the question. Input:{Your doctor wanted to clarify a few things.} True or False? Output: Table 11: Claim verification prompt using LLM internal knowledge. {claim} is enclosed in braces. user prompt: Answer the questio... | https://arxiv.org/abs/2505.18452v1 |
sentence is. You break down a sentence into as many facts as possible. The facts should be objective and verifiable against reliable external information such as Wikipedia and PubMed. All subjective personal ex- periences ("I was or someone did") and personal narratives (stating a past event) are not verifiable and sho... | https://arxiv.org/abs/2505.18452v1 |
concern is that your partner’s underlying condition, Addison’s disease, may not significantly complicate things if well- treated, but it could become an issue when the anabolic cycle is stopped. They strongly advise that your partner consult with a medical professional, ideally their endocrinologist, to discuss the pot... | https://arxiv.org/abs/2505.18452v1 |
facts: They also wanted to remind you that rabies has a relatively long incubation period, typically ranging from 1-3 months, before symptoms start to show. Facts: - Rabies has a relatively long incubation period. - The incubation period for rabies typically ranges from 1-3 months. - Rabies symptoms start to show after... | https://arxiv.org/abs/2505.18452v1 |
chair, standing, and eventually walking, until you’re back to your normal self. Please breakdown the following sentence into independent facts: Once the soreness starts to subside, they recommend that you try to gradually increase your activities, starting with small steps like sitting in a chair, standing, and eventua... | https://arxiv.org/abs/2505.18452v1 |
about the quality of care you received from your initial surgeon, your doctor advises that medical malpractice is a complex issue that depends on many factors, including the specific circumstances of your case and the laws in your location. If you’re interested in exploring this further, they recommend consulting with ... | https://arxiv.org/abs/2505.18452v1 |
phrase that is subjective and not fully decontextual- ized (i.e., “your specialist” versus “a specialist”). Ex. Claim: If you stop doing these exercises, they won’t be effective. Original sentence: Your doctor likens these exercises to medicine, meaning that if you stop doing them, they won’t be effective in managing y... | https://arxiv.org/abs/2505.18452v1 |
Hybrid Latent Reasoning via Reinforcement Learning Zhenrui Yue1, Bowen Jin1, Huimin Zeng1, Honglei Zhuang2, Zhen Qin2, Jinsung Yoon2, Lanyu Shang3, Jiawei Han1, Dong Wang1 1University of Illinois Urbana-Champaign,2Google,3LMU {zhenrui3,bowenj4,huiminz3,lshang3,hanj,dwang24}@illinois.edu, {hlz,zhenqin,jinsungyoon}@googl... | https://arxiv.org/abs/2505.18454v1 |
latent reasoning incorporates hidden represen- tations from previous steps to enhance reasoning performance (between <think> and</think> ). and overlooking the inherent reasoning capabilities of LLMs [ 11,8,34]. For example, Coconut [ 11] requires multi-stage training on CoT steps, which not only increases training com... | https://arxiv.org/abs/2505.18454v1 |
space computation within trans- former models [ 2,47]. For example, Biran et al. [2]study multi-hop reasoning and show that ‘back-patch’ features from later layers can improve performance on challenging queries. Alternatively, latent representations can be used to construct informative features as in-context demonstrat... | https://arxiv.org/abs/2505.18454v1 |
improved reasoning performance. 3 Methodology 3.1 Hybrid Reasoning with Gating We first describe our notation and settings for hybrid latent reasoning. For input query x= [x1, x2, . . . , x t]and its corresponding token embeddings E= [e1, e2, . . . , e t], we describe the raw hidden states from the LLM output at step t... | https://arxiv.org/abs/2505.18454v1 |
range of atconverges to an optimum range and thus incorporates informative features from both hidden representations and sampled tokens. 4 Overall, our hybrid reasoning approach projects hidden states into the embedding space via weighted interpolation. Moreover, the sampling steps preserve stochasticity for effective ... | https://arxiv.org/abs/2505.18454v1 |
to dynamically integrate sampled tokens with latent representations, delivering stable and efficient on-policy hybrid reasoning training without a separate value function. 4 Experiments We evaluate HRPO on both knowledge- and reasoning-intensive tasks: (1) open-domain & multi-hop knowledge-intensive question answering ... | https://arxiv.org/abs/2505.18454v1 |
NQ. (3) Interestingly, GRPO underperforms PPO by 4.6% on the 1.5B backbone but outperforms it by 1.6% on the 3B model, likely a consequence of sparser rewards and limited sampled trajectories with a smaller model. (4) RL- based methods perform on par with the best-performing RAG baseline, with HRPO delivering the large... | https://arxiv.org/abs/2505.18454v1 |
leader on MATH by 2.0%. (3) At 1.5B, HRPO improves on the strongest alternative GRPO with notable boosts on MATH and MATH500 (1.6% and 1.2%), whereas the average gain narrows at 3B, implying that HRPO is more beneficial for smaller models. (4) HRPO registers the highest accuracies recorded for sub-7B models on MATH (0.... | https://arxiv.org/abs/2505.18454v1 |
0.999] 0.336 0.534 0.258 0.275 0.216 0.324 Init RangeSTEM GSM8k MATH MATH500 MMLU-ST ARC-C Average [0.95 - 0.999] 0.705 0.516 0.536 0.569 0.735 0.612 [0.98 - 0.999] 0.703 0.509 0.532 0.563 0.732 0.608 [0.99 - 0.999] 0.720 0.518 0.526 0.567 0.742 0.614 8 Figure 5: Sensitivity analysis for temperature τin Equation (3). W... | https://arxiv.org/abs/2505.18454v1 |
intention of the story as the little lizard did not abandon its partner even if it did not move. B. It tells us to give more help to our love. This is not completely in line, because the lizard in the story did not give more help, it just waited and took care of its partner. C. It tells us to sympathize with the trappe... | https://arxiv.org/abs/2505.18454v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.