text
string
source
string
W. Bouaziz, and M.-P. Gill, “pyannote.audio: neural building blocks for speaker diarization,” inICASSP 2020, IEEE International Conference on Acoustics, Speech, and Signal Processing , Barcelona, Spain, May 2020. [17] J. P. Gala, P. A. Chitale, R. AK, V . Gumma, S. Doddapaneni, A. K. M., J. A. Nawale, A. Sujatha, R. Pu...
https://arxiv.org/abs/2505.18609v2
arXiv:2505.18610v1 [cs.CL] 24 May 2025PM-KVQ: Progressive Mixed-precision KV Cache Quantization for Long-CoT LLMs Tengxuan Liu∗1,2, Shiyao Li∗1,2, Jiayi Yang∗3, Tianchen Zhao1, Feng Zhou4, Xiaohui Song4, Guohao Dai5,2, Shengen Yan2, Huazhong Yang1, Yu Wang†1 1Tsinghua University,2Infinigence-AI,3Columbia University, 4O...
https://arxiv.org/abs/2505.18610v1
the quantization error. KIVI [ 16], SKVQ [ 6], and IntactKV [ 15] gain insights from the data distribution of the attention map and preserve the first or the most recent tokens in a higher bit-width within the KV Cache, thereby maintaining the performance. However, directly applying the above short-context-optimized me...
https://arxiv.org/abs/2505.18610v1
to enhance multi-step reasoning ability for complex tasks like mathematical proofs, scientific reasoning, and multi-hop QA. Models such as OpenAI- o1 [17], QwQ [ 21], and DeepSeek-R1 [ 8] employ advanced techniques to extend CoT reasoning depth. DeepSeek, specifically, integrates iterative self-refinement and tool-augm...
https://arxiv.org/abs/2505.18610v1
Key Cache and per-token quantization for Value Cache in a group-wise manner to reduce the quantization error. RotateKV [ 20] combines the channel-wise equalization and the rotation-based equalization with Hadamard matrices to reduce the quantization error. In this paper, we adopt effective strategies from prior work, c...
https://arxiv.org/abs/2505.18610v1
context length of the target long-CoT LLM for each transformer block. As shown in Figure 1(a) right, the Fbit in this example is 2-bit and the maximum context length is 32K. During generation, we initially store the KV Cache in 16-bit to alleviate the large cumulative quantization error. Once the memory budget is fully...
https://arxiv.org/abs/2505.18610v1
impact on the loss function, subject to a given memory budget. To this end, we formulate the block-wise bit-width allocation as the following Integer Programming problem: arg min xi,bNX iX bxi,b·si,b, (6) X bxi,b= 1,NX iX bxi,b·(Mem (Qb(Ki)) +Mem (Qb(Vi)))≤ M , (7) xi,b∈ {0,1}, b∈B, (8) where Nis the number of transfor...
https://arxiv.org/abs/2505.18610v1
in Figure 1(c) bottom, by applying positional interpolation, we can increase the largest positional index by s×without additional computation and memory overhead. 3.4 Method Pipeline In this paper, the proposed PM-KVQ combines the above three techniques to achieve better long- CoT performance with low bit-width KV Cach...
https://arxiv.org/abs/2505.18610v1
for DeepSeek-LLaMA-8B and 2-bit for other LLMs. Notably, the bit-width for the proposed PM-KVQ stands for the Fbit, as discussed in Section 3.1. In addition, for the optional bit-width set Bin Section 3.2, we use B={4,8}for DeepSeek-LLaMA-8B, and B={2,4}for other long-CoT LLMs. We use asymmetric group-wise quantization...
https://arxiv.org/abs/2505.18610v1
2-bit quantization, as 7 Table 2: Main results of long-CoT Language Models on reasoning-related benchmarks with SOTA KV Cache quantization methods. Models Quantization Bit-width AIME-2024 AIME-2025 CMIMC-2024 LiveCode (Target GPU) Methods (K-V) pass@1 V oting pass@1 V oting pass@1 V oting pass@1 - - 16-16 41.04 ±6.74 6...
https://arxiv.org/abs/2505.18610v1
Qwen-14B model, KIVI causes a performance degradation of 21.87% on CMIMC-2024, whereas PM-KVQ has a significantly lower degradation of only 1.87% and 2.91% under batch sizes of 16 and 12, respectively. For the 70B-level long-CoT LLM, we evaluate the 2-bit DeepSeek-R1-Distill-LLaMA-70B model on the AIME-2024 benchmark. ...
https://arxiv.org/abs/2505.18610v1
the other two strategies, increasing the pass@1 by 26.25% and 9.58%, and maintaining a lossless voting accuracy. Therefore, we adopt the Equivalent Right Shift strategy in PM-KVQ. 4.3.2 The Effect of Positional Interpolation We evaluate the long-CoT performance across varying lengths of calibration data and position sc...
https://arxiv.org/abs/2505.18610v1
Ar- mando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974 , 2024. [11] Shiyao Li, Xuefei Ning, Ke Hong, Tengxuan Liu, Luning Wang, Xiuhong Li, Kai Zhong, Guohao Dai, Huazhong Yang, and Yu Wang. Llm-mq...
https://arxiv.org/abs/2505.18610v1
Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems , 36:34661–34710, 2023. [27] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyu...
https://arxiv.org/abs/2505.18610v1
the attention mechanisms, such as the multi-head latent attention (MLA), which is quite different from the widely used Group-Query Attention (GQA). 13 Besides, we do not combine the proposed PM-KVQ with other system-level optimization techniques and inference engines, which yields for future work. D Broader Impacts Thi...
https://arxiv.org/abs/2505.18610v1
MA VL: A Multilingual Audio-Video Lyrics Dataset for Animated Song Translation Woohyun Cho Youngmin Kim Sunghyun Lee Youngjae Yu* Department of Artificial Intelligence Yonsei University k106419@yonsei.ac.kr Abstract Lyrics translation requires both accurate seman- tic transfer and preservation of musical rhythm, syllab...
https://arxiv.org/abs/2505.18614v1
strained Audio- Video LLM with Chain ofThought (SylAVL-CoT ).SylAVL-CoT enhances standard rea- soning by incorporating audio and video cues, en- abling better integration of contextual information across modalities for more accurate and musically coherent lyric translations.arXiv:2505.18614v1 [cs.CL] 24 May 2025 MA VL ...
https://arxiv.org/abs/2505.18614v1
and textual features, and development of special- ized evaluation metrics. For instance, (Guo et al., 2022) incorporated tonal constraints for translat- ing lyrics into Mandarin, while (Ou et al., 2023) experimented with integrating melody-length and phonetic constraints into translation systems. Also, (Kim et al., 202...
https://arxiv.org/abs/2505.18614v1
produced by Disney Corporation. For more details, refer to Section 3.1 reasoning across diverse languages and specialized knowledge domains (Hu et al., 2024b). Beyond these advancements, CoT prompting has been utilized in Multimodal machine translation tasks (Barrault et al., 2018) to improve translation quality (Rajpo...
https://arxiv.org/abs/2505.18614v1
denotes the individual lines within those sections. The number of videos is equal to the number of audio. corresponding audio and video. Lyrics Human Alignment. The human alignment process was crucial not only for segmenting lyrics but also for rigorously verifying their authenticity as official, singable dubbed versio...
https://arxiv.org/abs/2505.18614v1
"," 지 "," 는 ", " 마 "] 날 잊지마 , 슬퍼하지는 마Remeber me, D on ’ t let it make you cry[" 울 ", " 지 ", " 는 ", " 마 "] Revie w ing translatio n by segmented partsSyllable based re - tranlsation Figure 3: SylAVL-CoT pipeline for lyrics translation. This three-step process segments syllables utilizing au- dio, translates using video ...
https://arxiv.org/abs/2505.18614v1
simply comparing syllable counts without considering phonological and rhythmic dif- ferences across languages risks compromising the naturalness of translated lyrics. Therefore, it is es- sential to incorporate dubbed lyrics that adapt syl- lable counts. Syllable error (SE) measures how well the syl- lable count aligns...
https://arxiv.org/abs/2505.18614v1
0.947 0.984 Gemini 8.718 4.660 10.819 16.919 0.456 0.276 0.557 0.798 0.948 0.829 0.944 0.977w/ SCQwen2.5-72B 6.512 2.729 6.502 10.061 0.354 0.183 0.361 0.517 0.927 0.754 0.911 0.962 GPT-4o 3.164 1.442 3.084 7.221 0.185 0.138 0.182 0.378 0.795 0.676 0.935 0.935 Gemini 3.585 1.384 3.039 6.257 0.207 0.104 0.190 0.389 0.87...
https://arxiv.org/abs/2505.18614v1
target language. 4.2 Comparative Analysis In this section, we evaluate how well the translation models incorporate the three principles described in the Section 3.3. Additional experiments results can be found in Appendix C. Singability. As shown in Table 3, SylAVL-CoT achieves the lowest Syllable Error ,Syllable Dis- ...
https://arxiv.org/abs/2505.18614v1
SylAVL-CoT appears to retain a degree of naturalness similar to the original and dubbed renditions. 4.3 Ablation Studies Table 6 presents an ablation study examining the individual contributions of text, audio, and video inputs in our Multimodal CoT approach. When comparing semantic similarity with the originalModelLev...
https://arxiv.org/abs/2505.18614v1
A similar trend of improve- ment is observed when comparing the translated lyrics against the dubbed lyrics, with the combina- tion of both stages generally yielding the lowest Syllable Error . The prompt used for CoT without “Syllable List” or “Refine” can be found in Ap- pendix F. 4.4 User Study We conducted a user s...
https://arxiv.org/abs/2505.18614v1
focuses on ani- mated musicals and on five target languages (Span- ish, French, Japanese, and Korean, in addition to the original English). Although the dataset provides a rich testing ground for multimodal translation, their thematic variety may not represent the full range of musical genres, languages, and styles en-...
https://arxiv.org/abs/2505.18614v1
ment techniques, and more holistic metrics to cap- ture the creative and performative aspects of lyric translation. We hope that releasing MA VL and proposing SylAVL-CoT spark further innovations and encourage the research community to build on or address these limitations in future work. 7 Ethical Consideration In con...
https://arxiv.org/abs/2505.18614v1
Wong, and Rui Wang. 2024b. Large lan- guage model for multi-domain translation: Bench- marking and domain cot fine-tuning. Preprint , arXiv:2410.02631. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system c...
https://arxiv.org/abs/2505.18614v1
the Americas. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie- Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding. Preprint , arXiv:2004.09297. Moh Supardi and Dea Putri. 2018. Audio-visual trans- lation techniques: Subtitling and dubbing of movie soundtrack in frozen: Let it go. Bulet...
https://arxiv.org/abs/2505.18614v1
translated lyrics maintain a singable syllable count. Syllable Count Distance (SCD) Proposed by (Kim et al., 2024), SCD measures the congruity between the translated lyrics and the original struc- ture. It achieves this by calculating a normalized relative distance based on the absolute differences in syllable counts, ...
https://arxiv.org/abs/2505.18614v1
even when translations are relatively literal, potentially misrepresenting the true degree of translational fidelity. Therefore, the dubbed lyrics for each language (ES, FR, JP, KR) were first translated into English using Google Translate. Subsequently, the similarity between these English-translated dubbed lyrics and...
https://arxiv.org/abs/2505.18614v1
0.8776 0.8226 0.8103 0.7615 0.7656 0.6812 0.6764 GPT-4o 0.2042 0.2047 0.0703 0.0837 0.8295 0.8417 0.7932 0.7818 0.7421 0.7542 0.6759 0.6711 Gemini 0.1857 0.1785 0.0673 0.0826 0.7813 0.7904 0.7544 0.7327 0.7169 0.7234 0.6686 0.6604 SylA VL-CoT (Ours) 0.1900 0.1849 0.0679 0.0867 0.7652 0.7859 0.7289 0.6953 0.7142 0.7278 ...
https://arxiv.org/abs/2505.18614v1
d eeper 당신은 조금 더 깊이 파헤쳐 야합니다T ienes que cavar un poco m á s profundoT u dois creuser un peu plus profondémentも う少 し 深く掘り下げなければなり ま せんY ou need to dig a little deeperY ou need to dig a little deeperY ou have to dig a little deeperW e need to dig a little deeper .조금만 더 노력을 해 봐Es tu deber , trabaja dur oIl faut creuser en...
https://arxiv.org/abs/2505.18614v1
Average, 4: Good, 5: Very good)". This entire guide was translated and presented to participants in their respective evaluation languages. Finally, participants would click a "Start Evaluation" button to proceed. Example image of user study website can be seen in Figure 7. Each evaluation session lasted approximately 3...
https://arxiv.org/abs/2505.18614v1
2.732 9.213 2.996 3.096 3.001 4.017 Gemini (w/ SC) 1.030 1.412 1.854 4.718 0.997 2.519 2.280 5.339 1.397 1.896 1.534 4.364 2.107 2.033 2.137 1.631 SylA VL-CoT (Ours) 1.780 1.746 1.966 3.173 1.062 1.658 1.444 2.848 1.035 1.504 1.501 2.405 1.527 1.810 1.905 1.514 Table 11: Comparison of Syllable Error (SE) in Cross-Lingu...
https://arxiv.org/abs/2505.18614v1
and the utility of the MA VL dataset for multifaceted translation eval- uation. Notably, when Spanish is the source lan- guage, a relatively higher Syllable Error (SE) with Dubbed Lyrics can be observed. This aligns with the observation from Table 3, where Human Ex- pert translations into Spanish also showed a high SE....
https://arxiv.org/abs/2505.18614v1
employing Gemini 2.0 Flash, we bypass the need to develop and train a new model on a highly specialized, low-resource task. Instead, we explore whether prompt-based techniques alone can effectively solve the lyrics translation problem, even for content that requires multimodal understanding. Our approach centers on ada...
https://arxiv.org/abs/2505.18614v1
linguistic characteristics of each language, we utilized var- ious libraries, as shown in Table 20. After this process, to account for syllables in numbers, we used num2words7to convert numerical values into words, followed by IPA transcription (Mortensen et al., 2018) and syllable counting. I Experiment Details Comput...
https://arxiv.org/abs/2505.18614v1
Example: - Original syllable list: ["Three", "months", "of", "win", "ter", "cool", "ness"] - Translation syllable list: [" ᄉ ᅦ", "ᄃ ᅡ ᆯ", "ᄋ ᅴ", "ᄀ ᅧ", "울", "ᄎ ᅮ", "ᄋ ᅱ"] The target syllable list has 7 syllables, maintaining the count. 3. Iterate and Refine the Translation - After generating the initial translation, ...
https://arxiv.org/abs/2505.18614v1
following steps: 1. Identify the Core Lyric and Perform Syllable Segmentation - Read carefully to the provided lyrics to capture the natural rhythm, pronunciation, and any important phonetic, musical, or syllabic characteristics of the identified lyric line. - You will be also given the real syllable count of the origi...
https://arxiv.org/abs/2505.18614v1
following format: “‘json "translation": "final translation text" “‘ - Example: “‘json "translation": " ᄉ ᅦᄃ ᅡ ᆯᄋ ᅴᄀ ᅧ울ᄎ ᅮᄋ ᅱ" “‘ Table 16: Detailed Prompt for Text-Only Chain-of-Thought Lyrics Translation 1. Identify the Core Lyric and Perform Syllable Segmentation - The core lyric is: "And there’s a butterfly". - Au...
https://arxiv.org/abs/2505.18614v1
Iterate and Refine the Translation Upon reviewing the segmentation [" ᄂ ᅡ ᆯ", "ᄀ ᅵ", "ᄋ ᅥ ᆨ", "ᄒ ᅢ", "ᄌ ᅯ", "울", "ᄌ ᅵ", "는", "ᄆ ᅡ"], several issues become apparent: Upon reviewing the initial segmentation [" ᄂ ᅡ ᆯ", "ᄀ ᅵ", "ᄋ ᅥ ᆨ", "ᄒ ᅢ", "ᄌ ᅯ", "울", "ᄌ ᅵ", "는", "ᄆ ᅡ"], the primary issue is the syllable count. ...
https://arxiv.org/abs/2505.18614v1
ᅥᄂ ᅡᄀ ᅡᄌ ᅵ만" effectively, while maintaining a natural rhythm. 4. Generate the Final Translation The translation meets all the required conditions: natural flow, correct syllable count, and cultural appropriateness. The final English translation is: "Remember me though I have to leave you now" “‘json { "translation": ...
https://arxiv.org/abs/2505.18614v1
arXiv:2505.18630v1 [cs.CL] 24 May 2025DDO: Dual-Decision Optimization via Multi-Agent Collaboration for LLM-Based Medical Consultation Zhihao Jia1, Mingyi Jia1, Junwen Duan1*, Jianxin Wang1, 1Hunan Provincial Key Lab on Bioinformatics, School of Computer Science and Engineering, Central South University {zhihaojia, jia...
https://arxiv.org/abs/2505.18630v1
two decision-making sub-tasks presents significant challenges for unified optimization. To overcome these limitations, we propose DDO , 1 an LLM-based MC approach that leverages multi- agent collaboration to decouple and independently optimize the two core decision-making compo- nents. For symptom inquiry, DDO integrat...
https://arxiv.org/abs/2505.18630v1
tackle more complex tasks in the medical domain, recent studies (Kim et al., 2024; Bani-Harouni et al., 2024) have explored the use of multiple LLM agents, offering promising di- rections for enabling collaborative decision-making in challenging clinical scenarios. 3 Problem Definition A real-world Medical Consultation...
https://arxiv.org/abs/2505.18630v1
symptoms. All agents operate over a shared memory , comprising a static component that encodes prior disease-symptom knowledge derived from training data and a dy- namic component that is continuously updated with observed symptoms and diagnostic confidence dur- ing the consultation process. Each consultation round pro...
https://arxiv.org/abs/2505.18630v1
2022) to improve the Diagnosis Agent ’s ability to distinguish among similar diseases. For each pa- tient sub-trajectory Pc, the ground-truth diagnosis dlis treated as the positive instance, while all other candidate diseases serve as negatives. We construct a target distribution dist target= [ϵ, . . . , 1−ϵ, . . . , ϵ...
https://arxiv.org/abs/2505.18630v1
ranking, negative for worsened ranking, and zero if unchanged. The long-term reward Rlong(dp)assesses the fi- nal diagnostic prediction dp, yielding a positive reward if dp=dland a negative reward otherwise. 4.2.3 Masked Sampling for Candidate Actions The RL policy model’s sampling nature inherently prevents the guaran...
https://arxiv.org/abs/2505.18630v1
(Chen et al., 2024), and Direct Prompting (DP) . UoT plans future inquiries by computing information gain, while CoD employs instruction tuning to teach LLMs transparent di- agnostic reasoning. In contrast, DP relies solely on the inherent capabilities of the LLM, without prompt engineering or instruction tuning. 5 5.2...
https://arxiv.org/abs/2505.18630v1
signif- icantly improves diagnostic effectiveness. After symptom inquiry, it boosts diagnostic accuracy by 24.6%, 11.3%, and 3.2% on the DXY , GMD, and CMD datasets, respectively, over initial diag- noses based only on self-reported symptoms. DDO consistently achieves the highest accuracy, notably 94.2% on DXY . The DP...
https://arxiv.org/abs/2505.18630v1
and RL model for symptom inquiry, respectively. w/omasking disables action space mask- ing.w/oretry omits candidate actions regeneration. Instruct as the LLM backbone. Impact of Diagnostic Adapter. Removing the diagnostic adapter ( w/oadapter)—using only the vanilla BTP method to estimate diagnostic confi- dence—result...
https://arxiv.org/abs/2505.18630v1
inquiry during the early stages of MC. 5.8 Effect of Max Turns L As shown in Figure 3, we evaluate the diagnostic performance of the DDO framework on the DXYand GMD datasets for different maximum interac- tion turns ( L= 0/5/10/15). The line charts show an upward trend, indicating that increasing the max- imum number o...
https://arxiv.org/abs/2505.18630v1
scheme may hinder the model’s ability to softly distinguish between clinically similar diseases. Ethical Consideration Due to the hallucination problem inherent in large language models, they may generate content that is not factually accurate. Therefore, the DDO frame- work proposed in this paper is intended solely fo...
https://arxiv.org/abs/2505.18630v1
Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, Hae Park, et al. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems , 37:79410– 79452. 9 Abhishek Kumar, Robert Morabito, Sanzhar Umbet, Jad Kabbara, and Ali Emami. 2024. Confidence under the...
https://arxiv.org/abs/2505.18630v1
Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages 22546–22570. Donald E Stanley and Daniel G Campos. 2013. The logic of medical diagnosis. Perspectives in Biology and Medicine , 56(2):300–315. Zhoujian Sun, Cheng Luo, and Zhengxing Huang. 2024. Conversational disease diagnos...
https://arxiv.org/abs/2505.18630v1
Haiformer: Human-ai collaboration framework for disease diagnosis via doctor-enhanced transformer. In ECAI 2024 , pages 1495–1502. IOS Press. Zi’ou Zheng, Christopher Malon, Martin Renqiang Min, and Xiaodan Zhu. 2024. Exploring the role of rea- soning structures for constructing proofs in multi- step natural language r...
https://arxiv.org/abs/2505.18630v1
during RL policy training. To reduce train- ing costs, we used Qwen2.5-7B-Instruct with the BTP adapter as the diagnostic model in RL training. For Qwen2.5-14B-Instruct, we directly applied the learned RL policy for inference, demonstrating the policy’s transferability across different LLMs. Hyperparameters DXY GMD CMD...
https://arxiv.org/abs/2505.18630v1
appropriate action based on reasoning. The Patient Agent re- sponds with the presence or absence of symptoms based on the Medical Consultation Record (MCR). For symptoms not recorded in the MCR, the Pa- tient Agent infers their likely presence or absence based on the clinical presentation of the disease. TheDiagnosis A...
https://arxiv.org/abs/2505.18630v1
the most relevant, as it is strongly associated with pneumonia. Therefore, we choose short of breath as the next symptom to inquiry about. Symptom Status: True Response Reasoning: Short of breath is recorded in MCR, which indicates that the patient has this symptom. Diagnostic Confidence of Top Diseases: {"pneumonia": ...
https://arxiv.org/abs/2505.18630v1
Selected Symptom for Inquiry: expectoration Selection Reasoning: Pneumonia currently has the highest diagnostic confidence (0.68), much higher than upper respiratory infection and hand-foot-mouth disease (both 0.16). So, we focus on symptoms typical of pneumonia. Among the candidate symptoms, only expectoration (coughi...
https://arxiv.org/abs/2505.18630v1
use Strategy 2 and focus on symptoms related to the patient’s condition. The patient’s symptoms include cough, irritability, runny nose, fever, difficulty breathing, vomiting, phlegm, and anorexia. Looking at the frequency of candidate symptoms: Sneezing is more common in upper respiratory tract infection (0.233), whil...
https://arxiv.org/abs/2505.18630v1
unlikely to have the symptom "bubble-like spitting." The judgment is False. Diagnostic Confidence of Top Diseases: {"pneumonia": 0.68, "URTI": 0.16, "diarrhea": 0.16} Table 8: A medical consultation case by DDO. The symptom status of True and False represent the presence and absence, respectively. "Top Diseases" refers...
https://arxiv.org/abs/2505.18630v1
that the patient is likely to have the symptom. False indicates that the patient is unlikely to have the symptom. The criteria for judgment are as follows: If the symptom {inquiried_symptom} is relatively typical for the disease {disease_label} (i.e., it ranks high in the symptom occurrence frequency), then it is consi...
https://arxiv.org/abs/2505.18630v1
arXiv:2505.18638v1 [cs.CL] 24 May 2025Preprint. Under review. Multilingual Question Answering in Low-Resource Settings: A Dzongkha-English Benchmark for Foundation Models Md. Tanzib Hosain & Rajan Das Gupta Department of Computer Science American International University-Bangladesh Dhaka, Bangladesh {20-42737-1,18-3630...
https://arxiv.org/abs/2505.18638v1
et al. (2019), HellaSwag Zellers et al. (2019), CosmosQA Huang et al. (2019), and COPA Roemmele et al. (2011). X-COPA Ponti et al. (2020) is extensively used for non-English languages, and more recent initiatives such as Doddapaneni et al. (2023) have translated it into a number of Indic languages, while Dzongkha is st...
https://arxiv.org/abs/2505.18638v1
English, guaranteeing a high-quality parallel corpus. In educational research, this dual-language availability enables strong cross-linguistic analysis and application. 1https://bcsea.bt/ 2 Preprint. Under review. Table 1: Statistics of DZEN dataset grouped by various fields of subjects and levels. Group #Fields #Insta...
https://arxiv.org/abs/2505.18638v1
and Dzongkha are included in our suggested DZEN corpus. Some of the questions in this collection are from the 12th grade (55.36% (2,857)), the 10th grade 2https://translate.google.com/ 3 Preprint. Under review. (35.98% (1,857)), and the 8th grade (8.66% (447)). The disciplines covered in the 12th grade section include ...
https://arxiv.org/abs/2505.18638v1
open-source model trials because of this. In this work, GPT-3.5 Turbo is used for the majority of the other experiments (such as ablations, prompting method investigation, and others). This decision was taken in order to minimize the considerable expenses that come with more sophisticated proprietary models, especially...
https://arxiv.org/abs/2505.18638v1
dramatically in every topic. 5.2 Performance on CoT Prompting Prompting using CoT is often used to improve LLM performance on reasoning problems. We experimented across subjects and question categories to assess its efficacy on our dataset. 5.2.1 Subject-specific Performance Figure 1 shows how CoT reasoning enhances GP...
https://arxiv.org/abs/2505.18638v1
reasoning on the GPT-3.5 for DZEN English throughout k-shot. Note that w/ denotes with and w/o denotes without CoT. CoT00.050.10.150.20.250.30.350.40.45 1 2 3 4 5Accuracy Shot (k)w/ w/o Figure 6: Impact of CoT reasoning on GPT-3.5 across k-shot. Note that w/ denotes with and w/o denotes without CoT. For English, Figure...
https://arxiv.org/abs/2505.18638v1
is excluded from the advantages of AI, especially LLMs. A recently published open multilingual LLM, Aya ¨Ust¨un et al. (2024), represents a promising step in this direction. We also showed that it is possible to use query translation to have LLMs answer in the target language while utilizing the benefits of high-resour...
https://arxiv.org/abs/2505.18638v1
at SSRN , 2023. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457 , 2018. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lu...
https://arxiv.org/abs/2505.18638v1
chatgpt and gpt-4. arXiv preprint arXiv:2304.03439 , 2023. OpenAI. Gpt-4 technical report, 2023. Aleksandar Petrov, Emanuele La Malfa, Philip HS Torr, and Adel Bibi. Language model tokenizers introduce unfairness between languages. arXiv preprint arXiv:2305.15425 , 2023. Edoardo Maria Ponti, Goran Glava ˇs, Olga Majews...
https://arxiv.org/abs/2505.18638v1
our dataset during curation, thus it mostly consists of text-based questions. Given that visual issues frequently call for more sophisticated thinking, this constraint could limit the breadth of our findings. Furthermore, because the questions are multiple-choice, there’s a chance that models may skip some of the answe...
https://arxiv.org/abs/2505.18638v1
on the findings and task choices can be found in Appendix E. C Grammar Errors’ Impact on DZEN 00.10.20.30.40.50.60.7 10th Physics 10th Chemistry10th Math I 10th Math II 10th BiologyAccuracy Subjectw/ w/o Figure 11: Grammatical errors’ effects on GPT 3.5 performance. Note that w/ denotes with and w/o denotes without inc...
https://arxiv.org/abs/2505.18638v1
performance was marginally harmed in three instances, while it was beneficial in seven of the fourteen. Performance was essentially unchanged in the other four circumstances. F Experimental Prompts This section contains all the original prompts used for all the experiments. F.1 Prepared DZEN Questions with Correct Gram...
https://arxiv.org/abs/2505.18638v1
deduction to find the solution. Here is the question for you to evaluate: Figure 13: DZEN dataset categorized questions prompt. You are provided with a multiple-choice question and its options in English or Dzongkha. Your task is to correctly answer the question. Follow these steps: 1. Think and reason step by step, ex...
https://arxiv.org/abs/2505.18638v1
Sci Factual Knowledge 194 3.76% Procedural & Application 26 0.50% Reasoning 13 0.25% Total All 5161 100.00% 19 Preprint. Under review. Table 7: DZEN zero-shot benchmark. Language Subject GPT-4 GPT-3.5 Claude 2.1 LLaMA2 (13b) LLaMA2 (7b) Mistral 7b English12th Bio I 84.67 69.33 59.52 38.24 30.61 31.59 12th Bio II 82.15 ...
https://arxiv.org/abs/2505.18638v1
English Only Dzongkha Only Translation Append (Pre- mium)Translation Append (GPT) Causal Judgement 57.2 43.2 53.0 51.4 Date Understanding 56.0 20.2 36.8 28.1 Disambiguation QA 45.2 32.0 47.4 45.9 Formal Fallacies 55.2 42.9 42.8 41.4 Logical Deductions Five 35.2 17.3 15.1 16.6 Logical Deductions Seven 30.8 13.1 18.2 13....
https://arxiv.org/abs/2505.18638v1
Skip-Thinking: Chunk-wise Chain-of-Thought Distillation Enable Smaller Language Models to Reason Better and Faster Xiaoshu Chen, Sihang Zhou, Ke Liang, Xiaoyu Sun, Xinwang Liu National University of Defense Technology Abstract Chain-of-thought (CoT) distillation allows a large language model (LLM) to guide a small lang...
https://arxiv.org/abs/2505.18642v1
sharp minimum. More critically, in CoT distillation, the core reasoning tokens (such as the yellow and green ball in the rationale of Fig- ure 1) constitute a small proportion of rationales, while the prevalence of similar non-reasoning to- kens (e.g., those used for transition and summariza- tion) across different rat...
https://arxiv.org/abs/2505.18642v1
response speed of the answer. Among them, internalizing the explicit reasoning process (Deng et al., 2023, 2024) into the latent space has emerged as a promising direction. However, these internalization-based methods may compromise answer accuracy due to the lack of an explicit reasoning process. Similar to these appr...
https://arxiv.org/abs/2505.18642v1
mechanism, and latent space thought. Multi-task learning (Hsieh et al., 2023; Chen et al., 2024b; Liu et al., 2024) utilizes distinct pre- fixes to differentiate between tasks. For instance, when the input task prefix is [label] , the SLM di- rectly outputs the answer, whereas when the input task prefix is [rationale] ...
https://arxiv.org/abs/2505.18642v1
is thinking in mind. gation, the gradient of ϑcan be expressed as: ∂L ∂ϑ=∂ ∂ϑ 1 NNX i=1ℓi! =1 NNX i=1∂ℓi ∂ϑ(2) where N=B×(K−s)represents the token- level batch size and ℓiis the cross-entropy loss forithtraining token. Assume that we divide the training tokens into two sets S1andS2, where S1 is the training token set i...
https://arxiv.org/abs/2505.18642v1
]⊕qi⊕c1 i⊕c2 i...⊕cM i⊕ai(6) The reason for adding the prefix [m]and[answer ] is that it can tell the model what stage the current reasoning is at, thereby reducing the difficulty of reasoning. And the sin the Equation 1 is the start index of cm iandaiin these data at this time. 4.1.2 Search-based chunking Since the av...
https://arxiv.org/abs/2505.18642v1
accuracy of the SLM no longer increases. 4.3 Testing After training the SLM with the CWT and STT, when prompting SLM with the input [skip]⊕qtest, the SLM can adaptively skip the unimportant rea- soning chunk and only externalize the key reason- 5 ing chunks, thereby accelerating reasoning while ensuring reasoning accur...
https://arxiv.org/abs/2505.18642v1
We attribute the significant performance decline of the variant on the LLC dataset to the fact that the LLC dataset requires parallel reasoning rather than sequential reasoning, where each reasoning step is indepen- dent with no context dependence between them. Therefore, when using the variant, the SLM needs to reason...
https://arxiv.org/abs/2505.18642v1
that Scott requires the complete logits of each output token for implementation; thus, the rationales used in Scott are collected from the open-source model LLama3.1-70b-instruction (Dubey et al., 2024). Method Token type AD(%) TSO(%) Basecore reasoning tokens 87.37 89.25 other tokens 89.64 95.18 Base w. SBCcore reason...
https://arxiv.org/abs/2505.18642v1
the cases (Appendix G.1) where the correct answer is inferred after using CWT com- pared to base because the core reasoning token is predicted correctly. This also proves that CWT helps SLMs comprehend the core reasoning logic. Finally, the reasoning speed of the SLM trained with CWT based on SBC is faster than that of...
https://arxiv.org/abs/2505.18642v1
reduction techniques to mitigate toxicity in LLM reasoning. References Xiaoshu Chen, Sihang Zhou, Ke Liang, and Xinwang Liu. 2024a. Distilling reasoning ability from large language models with adaptive thinking. Preprint , arXiv:2404.09170. Xin Chen, Hanxian Huang, Yanjun Gao, Yi Wang, Jishen Zhao, and Ke Ding. 2024b. ...
https://arxiv.org/abs/2505.18642v1
Lin- guistics. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations . Stanisław Jastrz˛ ebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fi...
https://arxiv.org/abs/2505.18642v1
Linguistics.Subhro Roy and Dan Roth. 2015. Solving general arith- metic word problems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Lan- guage Processing , pages 1743–1752, Lisbon, Portu- gal. Association for Computational Linguistics. Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal ...
https://arxiv.org/abs/2505.18642v1
sampling with T=0.7 to gener- ate diverse rationales for each sample. A.3 Rationale generation of SLM The student model predictions are limited to a se- quence length of 1024 and greedy decoding is ap- plied for SLM across all benchmarks. A.4 Training datails For SLM training, we configure a batch size of 2, an initial...
https://arxiv.org/abs/2505.18642v1
first solution involves in- creasing the loss weight for core reasoning tokens in the rationale, while the second solution focuses on designing prompts to guide the LLM in generat- ing refined rationales with minimal non-reasoning content. In this work, we evaluate the feasibility of these 11 Question: Alyssa picked 17...
https://arxiv.org/abs/2505.18642v1
failure cases for these schemes. It can be observed that these schemes often generate repetitive reasoning steps until the maximum generation length is reached. This oc- curs because the number of chunks resulting from sentence- or step-based segmentation is typically variable, making it challenging for the SLM to de- ...
https://arxiv.org/abs/2505.18642v1
are linear complexity, they will bring a little additional training time compared to the baseline. Second, chunking increases the amount of training data. In theory, compared to other methods, CWT requires approximately M+ 1times the training time, while STT requires about num i×(M+ 2) times the training time, where nu...
https://arxiv.org/abs/2505.18642v1
is dancing with Patrick. Throughout the song, the dancers often trade partners. First, Bob and Alice switch partners. Then, Claire and Bob switch partners. Finally, Claire and Alice switch partners. At the end of the dance, Bob is dancing with Which choice is true? Answer choices: (A) Rodrigo. (B) Ophelia. (C) Patrick....
https://arxiv.org/abs/2505.18642v1
arXiv:2505.18644v1 [eess.AS] 24 May 2025Enhancing Generalization of Speech Large Language Models with Multi-Task Behavior Imitation and Speech-Text Interleaving Jingran Xie1,2, Xiang Li1,2, Hui Wang2, Yue Yu2, Yang Xiang2, Xixin Wu3, Zhiyong Wu1,3,∗ 1Shenzhen International Graduate School, Tsinghua University, China 2P...
https://arxiv.org/abs/2505.18644v1