text
string
source
string
conducted an additional ablation study testing debates w ith four rounds instead of three (adding a second rebuttal round). Due to technical limitations - spe cifically, poor instruction-following and XML formatting issues that caused systematic parsing failu res - we were only able to successfully run this experiment w...
https://arxiv.org/abs/2505.19184v2
threshold. The opposition’s focus on the California recall was effective in grounding their ar gument in a real-world example, even though I tried to counter it. Looking at the judging criteria: Direct Clash Analysis: I think I did a better job of directly addressing the opponent ’s specific claims and quoting disagreem...
https://arxiv.org/abs/2505.19184v2
which are substantiated in §3 (Results) and §2 (Method ology). No claims beyond those sections appear in the discussion or conclusion 2.Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: The paper devotes a subsection (§ 4 "Limitati ons and Fut...
https://arxiv.org/abs/2505.19184v2
the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines ? Answer: [Yes] Justification: The work involves only synthetic LLM outputs , no personal data or human subjects, follows responsible-AI guidelines, and all pote ntially mis-informative findings are disclosed w...
https://arxiv.org/abs/2505.19184v2
or institution) were obtained? Answer: [NA] Justification: No human subjects were involved in this resea rch, as all experiments were conducted using language models. Therefore, IRB approval w as not required 16.Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original,...
https://arxiv.org/abs/2505.19184v2
arXiv:2505.19187v1 [cs.CL] 25 May 2025LIMOPro: Reasoning Refinement for Efficient and Effective Test-time Scaling Yang Xiao1Jiashuo Wang1Ruifeng Yuan1Chunpu Xu1 Kaishuai Xu1Wenjie Li1†Pengfei Liu2† 1The Hong Kong Polytechnic University2Shanghai Jiao Tong University yang-alan.xiao@connect.polyu.hk csjwang@comp.polyu.edu...
https://arxiv.org/abs/2505.19187v1
thus improving computational efficiency, demonstrating that selectively pruning low-importance functional steps produces more concise, faster, and more accurate reasoning chains. errors, revising the approach, and ultimately confirming the final answer. This thorough process generates lengthy reasoning chains with redu...
https://arxiv.org/abs/2505.19187v1
efficient reasoning without sacrificing solution quality. By systematically refining training data to preserve essential reasoning while eliminating redundant functional steps (verification, validation, and error correction processes), we enable LLMs to produce concise yet equally effective reasoning chains, advancing ...
https://arxiv.org/abs/2505.19187v1
Integral SetupStep 5[Progressive]Formulate the Area IntegralStep 3[Progressive]Solve for Intersection Points1.Raw Reasoning2.ClassifyStep Types3.Calculate PIR4.Refinement Figure 2: PIR framework pipeline for reasoning optimization: raw reasoning is segmented into logical steps, step is classified into reasoning pattern...
https://arxiv.org/abs/2505.19187v1
functional steps (verification, multi-method validation, and error correction) while preserving all progressive reasoning steps, which constitute the essential deductive core of the solution process. For each identified functional step, we compute its PIR value to quantify its importance based on the impact its removal...
https://arxiv.org/abs/2505.19187v1
9.36E-05 LIMO-P 63.3+6.610,588-15% 5.98E-05+32% 93.8+1.95,235-5% 1.79E-04+7% 71.2+46,969-3% 1.02E-04+9% LIMO-V2 66.3 13,896 4.77E-05 94.4 6,843 1.38E-04 70.2 8,035 8.74E-05 LIMO-V2-P 71.2+4.912,163-12% 5.86E-05+23% 96.6+2.26,348-7% 1.52E-04+10% 74.2+36,968-13% 1.07E-04+22% problem in our benchmark, we sample eight resp...
https://arxiv.org/abs/2505.19187v1
2,853 2.03E-04 S1-SPIRIT 4.32E+06 37.1 4,906 7.56E-05 81.3 3,517 2.31E-04 60.1 2,818 2.13E-04 S1-RULE 4.31E+06 36.7 4,807 7.63E-05 81.3 3,654 2.22E-04 58.1 3,837 1.51E-04 S1-32B-P 4.31E+06 42.1+4.24,716-29% 8.92E-05+56% 83.1+2.23,809-16% 2.18E-04+22% 61.6+0.92,472-41% 2.49E-04+71% 3.2 Main Results Performance Across Be...
https://arxiv.org/abs/2505.19187v1
highlights the robustness of our optimization methodology to different deployment scenarios. 7 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Ratio0.80.91.01.11.2Relative Value AIME 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Ratio0.80.91.01.11.2 AMC ACC TOK EFFFigure 3: Impact of pruning ratio on model performance. This figure displays relative perform...
https://arxiv.org/abs/2505.19187v1
learning (RL). SFT methods train on high-quality reasoning traces [ 20,37,16,15], with S1 and LIMO demonstrating improved performance through careful sample selection. RL approaches [24,10,12,29] have yielded exceptional results, with DeepSeek-R1 using GRPO and OpenAI’s o1 and QwQ enabling autonomous development of rea...
https://arxiv.org/abs/2505.19187v1
[1]Claude 3.7 Sonnet and Claude Code — anthropic.com. https://www.anthropic.com/ news/claude-3-7-sonnet . [Accessed 30-04-2025]. [2]Thinking | Generative AI on Vertex AI | Google Cloud — cloud.google.com. https: //cloud.google.com/vertex-ai/generative-ai/docs/thinking . [Accessed 30-04- 2025]. [3]Lingjiao Chen, Jared Q...
https://arxiv.org/abs/2505.19187v1
Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations , 2023. [19] Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuni...
https://arxiv.org/abs/2505.19187v1
11 [35] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 , 2024. [36] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thou...
https://arxiv.org/abs/2505.19187v1
= (ln a)/(ln b), so log_{1995} x is equal to (ln x)/(ln 1995). Also, x raised to log_{1995} x can be rewritten using exponentials. Wait, maybe I can express x^(log_{1995} x) as e^{(ln x) * (log_{1995} x)}. But that might not be helpful.Let me think differently.\n\nAlternatively, since the equation has x in the exponent...
https://arxiv.org/abs/2505.19187v1
- 10,000 + 25 = 990,025. Correct. So 990,025. The last three digits are indeed 025. So 025. Therefore, the answer should be 025. But leading zeros are usually not written, but since the question asks for the last three digits, we need to include the leading zero. Sothe last three digits are 025.\n\nBut let me confirm o...
https://arxiv.org/abs/2505.19187v1
\"We need to\", \"Given that\"\n\n**Example**:\n\n```\n\"Let's solve this step by step:\"\n\"First, we need to calculate n = 2^31 * 3^19\"\n\"So, we start by calculating this value.\"\n\"Now, let's compute the value of n.\"\n\"Therefore, let's find n = 2^31 * 3^19.\"\n\n```\n\n### 2. Verification\n\n**Definition**: Pro...
https://arxiv.org/abs/2505.19187v1
test-time efficiency to highlight the relationship between data efficiency and model performance. The baseline models—R1-Distill-Qwen-32B, Qwen2.5-32B-Instruct, and QWQ—establish performance references across the benchmarks. QWQ demonstrates superior accuracy but requires substantially longer responses, which impacts i...
https://arxiv.org/abs/2505.19187v1
should be categorized as one of the four patterns.\n3. **Pattern Identification**:\n -**Progressive Reasoning**: One continuous pattern from start to finish with logical steps\n -**Verification**: Moments where solver checks previous work (reference which Regular Reasoning ID is being verified)\n -**Multi-method Valida...
https://arxiv.org/abs/2505.19187v1
_steps, pattern ); threshold ←CalculatePruningThreshold (pattern _steps, ratio ); steps _to_prune ←SelectLowPIRSteps (pattern _steps, threshold ); solution opt←RemoveSteps (solution opt, steps _to_prune ); end return solution opt; 0.8 0.9 1.0 1.1 1.2 Relative Test Time Efficiency0.80.91.01.11.2Relative AccuracyBetter A...
https://arxiv.org/abs/2505.19187v1
9.08E-04 R1-Distill- Qwen-32BN/A 69.2 9,311 7.43E-05 N/A 94.4 5,561 1.70E-04 N/A 64.7 5,634 1.15E-04 QWQ N/A 81.7 12,234 6.68E-05 N/A 97.8 7,350 1.33E-04 N/A 70.2 7,483 9.38E-05 S1 S1-32B 4.51E+06 37.9 6,646 5.71E-05 4.51E+06 80.9 4,542 1.78E-04 4.51E+06 60.7 4,172 1.46E-04 S1-32B-P 4.31E+06 42.1 4,716 8.92E-05 4.31E+0...
https://arxiv.org/abs/2505.19187v1
maybe the problem says \"the walk takes her 4 hours\" including the coffee shop time, so perhaps the walking time is 4 hours minus t minutes. Wait, actually, that's exactly how we modeled it. The walking time is distance over speed, then add the coffee time to get the total time. So our model is correct.\n\nAlternative...
https://arxiv.org/abs/2505.19187v1
2 hours (120 minutes) plus 24 minutes is 144 minutes, which is 2 hours 24 minutes. So that's consistent.\n\nTherefore,whenwalkingat3km/h,walkingtimeis3hours(180minutes)plus24minutesinthecoffeeshop,totaling204minutes.Sotheanswershouldbe204minutes.\n\nBut just to double-check, let me go through the steps again.\n\nOrigin...
https://arxiv.org/abs/2505.19187v1
in practice and what the implications would be. •The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. •The authors should reflect on the fac...
https://arxiv.org/abs/2505.19187v1
suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also...
https://arxiv.org/abs/2505.19187v1
recommended, but including URLs to data and code is permitted. 6.Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: 3.1 Exp...
https://arxiv.org/abs/2505.19187v1
amount of compute required for each of the individual experimental runs as well as estimate the total compute. •The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9.Cod...
https://arxiv.org/abs/2505.19187v1
adhere to usage guidelines or restrictions to access the model or implementing safety filters. •Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. •We recognize that providing effective safeguards is challenging, and many pape...
https://arxiv.org/abs/2505.19187v1
of the data collector. 15.Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalen...
https://arxiv.org/abs/2505.19187v1
Misleading through Inconsistency: A Benchmark for Political Inconsistencies Detection Nursulu Sagimbayeva1, Ruveyda Bet ¨ul Bahc ¸eci2, Ingmar Weber1 1Saarland Informatics Campus, Saarland University 2Saarland University nusa00001@uni-saarland.de, ruba00002@uni-saarland.de, iweber@cs.uni-saarland.de Abstract Inconsiste...
https://arxiv.org/abs/2505.19191v1
prompt clarifications. How- ever, with the massive amount of digital content produced by parties and their members, it becomes more challeng- ing to detect inconsistencies manually. Moreover, parties use different platforms to communicate with their electorate, and can alter their messages based on the platform: for ex...
https://arxiv.org/abs/2505.19191v1
of arbitrary length (from one- liner social media posts to several-page manifestos). • Statements AandBmight be a part of the bigger doc- ument D, in which case self-inconsistency detection is performed for D. • To simplify the task, we assume that AandBwere said by the same actor on the same day. This is important bec...
https://arxiv.org/abs/2505.19191v1
be consistent or inconsistent, depending on their relationship to each other, but independent of an external truth value. Inconsistency Detection Vs Stance Detection Stance detection involves identifying an actor’s position or attitude toward a specific target topic. Usual labels for at- titude are {Favor, Against, Non...
https://arxiv.org/abs/2505.19191v1
is hard to account for in advance if we use preset targets. On the other hand, gener- ating such targets dynamically would create a huge universe of possible targets. Inconsistency Detection Vs Inconsistency Detection in Summarization The objective of Inconsistency Detection in Summarization (IDS) is to, given the orig...
https://arxiv.org/abs/2505.19191v1
from the Perspectrum dataset (Chen et al. 2019), although we do not directly use it in our sample generation pipeline. Sample generation for each class ForUnrelated class, we randomly sample N pairs of statements from Wahl-O- Mat dataset. For Consistent samples, we randomly select N statements and pair each with a comm...
https://arxiv.org/abs/2505.19191v1
types as one Inconsistent class; nominal metric); more on the metric choice in Appendix G. A small subset of samples had more than 5 annotations; in this case, we randomly sampled 5 out of N samples to calculate the agree- ment. While the score is generally not high, it is common in other subjective tasks, such as judg...
https://arxiv.org/abs/2505.19191v1
Model evaluation We evaluated four off-the-shelf models using the same instructions given to annotators, excluding visuals (see prompt in Appendix H). The following mod- els were considered: gpt-4-turbo-2024-04-09 , gpt-3.5-turbo-instruct (further referred to as ChatGPT-4 turbo and ChatGPT-3.5 turbo), LLaMA3.3 70B Inst...
https://arxiv.org/abs/2505.19191v1
overall performance in predicting the ma- jority label, sometimes even outperforming individual hu- man annotators. Model Unrel. Consist. Inconsist. ChatGPT-4 turbo 0.548 0.662 0.619 LLaMA 70B 0.525 0.707 0.633 Humans 0.503 0.637 0.617 Bootstrap humans 0.727 0.798 0.786 Table 3: Performance for 3 classes by MCC Model I...
https://arxiv.org/abs/2505.19191v1
sistency might be subjective and depend on factors such as personal knowledge, education, political preferences, and even factors such as concentration and attentiveness while judging the statements. Because it is hard to strictly define degrees of inconsistency, we believe aligning models’ un- derstanding of inconsist...
https://arxiv.org/abs/2505.19191v1
inconsistencies go unnoticed, and at least detecting them would be beneficial for both practical and research purposes. Regarding the content of the dataset, we realize that it contains some samples that some people might find offen- sive (for example, statements like ”Headscarves for female teachers in public schools ...
https://arxiv.org/abs/2505.19191v1
I.; and Glickman, O. 2004. Probabilistic textual en- tailment: Generic applied modeling of language variability. Proceedings of the PASCAL Workshop on Textual Entail- ment and Paraphrasing . This is a workshop paper and might not have a standard journal reference. Das, A.; Liu, H.; Kovatchev, V .; and Lease, M. 2023. T...
https://arxiv.org/abs/2505.19191v1
Detection in Summarization. Transactions of the Association for Computational Linguistics , 10: 163–177. Lattimer, B.; Chen, P. H.; Zhang, X.; and Yang, Y . 2023. Fast and Accurate Factual Inconsistency Detection Over Long Documents. In Bouamor, H.; Pino, J.; and Bali, K., eds., Proceedings of the 2023 Conference on Em...
https://arxiv.org/abs/2505.19191v1
Borgeaud, S.; Yogatama, D.; Bosma, M.; Zhou, D.; Met- zler, D.; Chi, E. H.; Hashimoto, T.; Vinyals, O.; Liang, P.; Dean, J.; and Fedus, W. 2022. Emergent Abilities of Large Language Models. ArXiv , abs/2206.07682. Wolfram, S. 1989. Philosophical Logic . Routledge. Wong, K.; Paritosh, P.; and Aroyo, L. 2021. Cross- repl...
https://arxiv.org/abs/2505.19191v1
(we anonymize names of politicians and parties) (i) Have you read the ethics review guidelines and en- sured that your paper conforms to them? Yes 2. Additionally, if your study involves hypotheses testing... (a) Did you clearly state the assumptions underlying all theoretical results? N/A (b) Have you provided justifi...
https://arxiv.org/abs/2505.19191v1
or releasing new datasets, did you discuss how you intend to make your datasets FAIR? No (g) If you are curating or releasing new datasets, did you create a Datasheet for the Dataset? Yes, we addressed the questions it asks in our document in section 5. Dataset, but we did not literally fill out the datasheet 6. Additi...
https://arxiv.org/abs/2505.19191v1
accord- ing to our taxonomy. C Scale visualization See Figure 7. D Overview of contradiction types See Figure 5 for an overview of contradiction types in other literature and its relation to our taxonomy. Works marked with an asterisk * name their category differently, but based on their meaning, we categorize them as ...
https://arxiv.org/abs/2505.19191v1
today. A B Indirect inconsistency V K - knowledge V - value, opinion, ideology If A is True, B is False , and vice versa. Understanding logical form/language in A and B is enough to see a contradiction. Example: a) All kikis are bobable. b) This kiki is not bobable. + ***************************************************...
https://arxiv.org/abs/2505.19191v1
do that, the door opens.(Li, Raheja, and Kumar 2024) Relation 1) Jane and Tom are a married couple. 2) Jane is Tom’s sister.(Li, Raheja, and Kumar 2024) Value inconsistency Violation of expectations I didn’t attend the funeral, but I sent a nice letter saying I approved of it.(Lin and Zhang 2023) Table 5: Types of Inco...
https://arxiv.org/abs/2505.19191v1
comprehension check consisting of four questions. Only participants who scored at least 2/4 were eligible for further consideration. To ensure a better understanding of the task before com- prehension checks, we also conducted practice sessions, where participants could annotate a set of questions and see the answers a...
https://arxiv.org/abs/2505.19191v1
beyond what is said in A and B is required to see inconsistency. This knowledge can include laws of physics, principles of economics, international rela- tions, etc., as well as real-world events and empirical evi- dence. Example 1: a) We will provide extensive social benefits. b) We will minimize all the taxes. Explan...
https://arxiv.org/abs/2505.19191v1
2: a) We believe that the current government is failing to address population decline effectively. b) We be- lieve the government should prioritize the traditional family model. Example 3: a) We believe that determining asylum eli- gibility before individuals reach the country would signif- icantly relieve the taxpayer...
https://arxiv.org/abs/2505.19191v1
arXiv:2505.19201v1 [cs.CL] 25 May 2025 DREAM: Drafting with Refined Target Features and Entropy-Adaptive Cross-Attention Fusion for Multimodal Speculative Decoding Yunhai Hu1, Tianhua Xia1, Zining Liu2, Rahul Raman1, Xingyu Liu1, Bo Bao3, Eric Sather3, Vithursan Thangarasa3, Sai Qian Zhang1 1New York University,2Univer...
https://arxiv.org/abs/2505.19201v1
accuracy. Finally, DREAM introduces a visual input compression scheme for the draft model, guided by the intermediate features from the target model, which substantially reduces processing latency without compromising accuracy. We evaluate DREAM on a diverse set of popular VLMs, including LLaV A-v1.6-Vicuna-7B/13B [ 29...
https://arxiv.org/abs/2505.19201v1
techniques include layer skipping [ 14, 2 61,8,28,32,56], which accelerates draft generation by selectively processing fewer transformer layers. Other strategies leverage token distillation [ 64], N-gram prediction [ 42,49,34], and retrieval- augmented drafting [57, 13, 54] to improve draft quality while minimizing com...
https://arxiv.org/abs/2505.19201v1
with parameters ranging from 256M to 2B, achieving exceptional performance while maintaining smaller model sizes. To quantify the computational cost introduced by visual inputs processing in VLMs, we profile the floating point operations (FLOPs) required by various models, including LLaV A-v1.6-Vicuna- 7B [31], LLaV A-...
https://arxiv.org/abs/2505.19201v1
visual input to accelerate output generation. Additionally, the intermediate features are adaptively selected to better guide the training of the draft model. Figure 3 presents the detailed training and inference scheme of DREAM. Specifically, let MtaandMdarepresent the target and draft models, respectively, with a tot...
https://arxiv.org/abs/2505.19201v1
the keys and values. With zdenoting the dimensionality of queries and keys, the cross-attention is then computed as follows: Q=EjWQ, K =SLWK, V =SLWV, F = softmaxQK⊤ √z V (1) where WQ, WK, WVdenote the weight matrices in the draft model. The fused features Freplace the original first-layer features and are propagated...
https://arxiv.org/abs/2505.19201v1
ℓ⋆ with the lowest average entropy, defined as ℓ∗= arg minℓ∈L[AE(ℓ)]. We then distill the information from sℓ⋆ iinto the initial decoder block of the draft model using a smooth ℓ1loss, guiding the draft model to align with the most informative intermediate representation of the target model. This adaptive feature disti...
https://arxiv.org/abs/2505.19201v1
output. This strategy significantly reduces the number of tokens while retaining the most important visual information. Although visual token subsampling may cause a slight decrease in the draft model’s accuracy, our evaluation in Section 4 shows that it can effectively reduce speculative decoding latency. 4 Empirical ...
https://arxiv.org/abs/2505.19201v1
speedup of 3.06×on LLaV A-v1.6-Vicuna-7B and 2.65×on Pixtral-12B, whereas the speedups are lower for smaller models, with 2.27×on SmolVLM-2B and 2.23×on LLaV A-v1.6-Vicuna-7B. This is because larger models experience more severe decoding bottlenecks, allowing the draft model to more effectively substitute the costly de...
https://arxiv.org/abs/2505.19201v1
2.12 1.31 2.09 1.17 1.89 1.18 1.98 1.15 1.86 1.24 2.01 Medusa [4] 1.58 2.88 1.59 3.01 1.44 2.77 1.22 2.33 1.25 2.41 1.22 2.34 1.38 2.62 Hydra [2] 1.78 3.86 1.72 3.88 1.68 3.79 1.41 3.21 1.35 3.11 1.42 3.25 1.56 3.52 EAGLE [25] 2.10 5.04 2.09 5.01 1.98 4.88 1.72 4.13 1.56 3.98 1.78 4.25 1.87 4.55 EAGLE-2 [24] 2.31 5.48 ...
https://arxiv.org/abs/2505.19201v1
4.13 1.76 4.56 1.95 4.91 DREAM 2.39 6.29 2.35 6.07 2.25 5.68 1.99 4.88 1.84 4.41 2.02 5.23 2.14 5.43 LLA V A-v1.6 Vicuna-13BSPD 0.88 1.22 0.84 1.25 0.84 1.32 0.79 1.18 0.81 1.14 0.88 1.24 0.84 1.22 Kangaroo 1.23 1.57 1.17 1.53 1.07 1.44 1.01 1.24 1.07 1.34 1.21 1.67 1.13 1.46 EAGLE-2 2.35 3.75 3.02 4.30 3.03 4.67 2.03 ...
https://arxiv.org/abs/2505.19201v1
token length τacross (a) intermediate feature selection strategies. (b) chain-based and tree-based decoding. (c) visual token compression ratios, where 1 and 3/4 denote 100% and75% of the visual tokens are retained, respectively. (d) loss weight settings, where the number is the value for λfeat andλintermed .λKLis fixe...
https://arxiv.org/abs/2505.19201v1
to 0.2 improves speedup and average accepted token length. However, at 0.4, both metrics drop, indicating that excessive feature supervision may harm generalization. 5 Conclusion and Limitation We present DREAM, a speculative decoding framework optimized for VLMs. By integrating visual token compression, cross-attentio...
https://arxiv.org/abs/2505.19201v1
Kathleen Kenealy, Lucas Beyer, Xiaohai Zhai, Anton Tsitsulin, Robert Busa-Fekete, Alex Feng, Noveen Sachdeva, Benjamin Coleman, Yi Gao, Basil Mustafa, Iain Barr, Emilio Parisotto, David Tian, Matan Eyal, Colin Cherry, Jan-Thorsten Peter, Danila Sinopalnikov, Surya Bhupatiraju, Rishabh Agarwal, Mehran Kazemi, Dan Malkin...
https://arxiv.org/abs/2505.19201v1
Shi, Jianfeng Gao, and Jianwei Yang. Ola-vlm: Elevating visual perception in multimodal llms with auxiliary embedding distillation. arXiv preprint arXiv:2412.09585 , 2024. [17] Byung-Kwan Lee, Ryo Hachiuma, Yu-Chiang Frank Wang, Yong Man Ro, and Yueh-Hua Wu. Vlsi: Ver- balized layers-to-interactions from large to small...
https://arxiv.org/abs/2505.19201v1
Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery of ocr in large multimodal models. Science China Information Sciences , 67(12), December 2024. [36] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishir...
https://arxiv.org/abs/2505.19201v1
and Jingjing Liu. Contrastive distillation on intermediate representations for language model compression. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 498–508, 2020. [52] Ziteng Sun, Ananda Theertha Suresh, Jae Hun Ro, Ahmad Beirami, Himanshu Jain, and Felix...
https://arxiv.org/abs/2505.19201v1
the most recently verified token provided by the target model. Here, the special prefix token “Fig.” denotes the visual tokens representation of the input image. From the root, the draft model expands a tree of candidate tokens across multiple depth layers. We expand the draft tree by sampling candidate tokens at each ...
https://arxiv.org/abs/2505.19201v1
inputs. During inference, all models are run with KV caching enabled to ensure efficient autoregressive decoding. For draft model training of the baseline models, including Kangaroo [ 28], Medusa [ 4], EAGLE [ 25] and EAGLE-2 [ 24], we follow their publicly available training scripts, introducing only minimal adjustmen...
https://arxiv.org/abs/2505.19201v1
SpeakStream: Streaming Text-to-Speech with Interleaved Data 1stRichard He Bai Apple United States richardbai@apple.com2ndZijin Gu Apple United States zijin@apple.com3rdTatiana Likhomanenko Apple United States antares@apple.com4thNavdeep Jaitly Apple United States njaitly@apple.com Abstract —The latency bottleneck of tr...
https://arxiv.org/abs/2505.19206v1
model that enables streaming TTS through modeling interleaved text-speech segments instead of a pair of text and speech. To create interleaved text-speech training data, we use a force-aligner [10] to align text and speech pairs from traditional TTS datasets. The model then trained like a standard LLM on the interleave...
https://arxiv.org/abs/2505.19206v1
distilling from a non-streaming TTS with limited access to future context. However, their architecture demon- strates limited zero-shot capability. Another work [6] upgraded LiveSpeech [22] from full-text audio synthesis to text-chunksynthesis. However, their model encounters misalignment is- sues between speech genera...
https://arxiv.org/abs/2505.19206v1
precise temporal correspondence between words and speech frames, we utilize the A3T’s alignment mechanism [10]. By training a decoder-only transformer model on such interleaved sequences, SpeakStream learns to synthesize the current speech segment Aiconditioned on the current text segment Ti, previous speech segments A...
https://arxiv.org/abs/2505.19206v1
updating the kv-cache state; 3) Repeats this process until the entire text is synthesized. The simplicity of dMel tokenization allows our model to generate high-quality speech in a streaming fashion without the complexity of managing multiple token types or separate acoustic and semantic representations. The primary la...
https://arxiv.org/abs/2505.19206v1
ocStream) with LibriTTS-R dataset [33]. It employs a generator comprising 30 WaveNet residual blocks (4.6M parameters) and a discriminator with 10 layers. The V ocStream vocoder takes the 120-channel, 6.25ms-resolution Mel features from the upsampler and synthesizes a 24kHz waveform. It consists of a 44-layer WaveNet-b...
https://arxiv.org/abs/2505.19206v1
2.3±0.1 2.2±0.1 XTTS 4.1±0.1 1.8±0.1 2.7±0.1 SpeakStream - 3.9±0.1 3.8±0.1 TABLE III LATENCY (MS)OFSPEAK STREAM WITH PARALLEL WAVEGAN (PWG, 1M) ORVOCSTREAM (11.9M) VOCODERS ,TESTED WITH APPLE SILICON (MAC MINI , M4 P RO, 64GB, 2024). PLAYER LATENCY IS LESS THAN 0.2MS. SpeakStream V ocoder TTS (ms) V ocoder (ms) Total (...
https://arxiv.org/abs/2505.19206v1
its first frame input and generating the first chunk of waveform. 3)TTS latency : Time elapsed between the TTS model receiving its first word and generating its first frame. Note, we are interested not in time when the first waveform outputted but rather in time of the first spoken phoneme.1 During inference we prompt ...
https://arxiv.org/abs/2505.19206v1
P ´erez, H. J ´egou, E. Grave, and N. Zeghidour, “Moshi: a speech-text foundation model for real-time dialogue,” arXiv preprint arXiv:2410.00037 , 2024. [3] J. Xu, Z. Guo, J. He, H. Hu, T. He, S. Bai, K. Chen, J. Wang, Y . Fan, K. Dang, B. Zhang, X. Wang, Y . Chu, and J. Lin, “Qwen2.5-omni technical report,” arXiv prep...
https://arxiv.org/abs/2505.19206v1
Jaitly, Z. Yang, Y . Xiao, Z. Chen, S. Bengio et al. , “Tacotron: Towards end- to-end speech synthesis,” arXiv preprint arXiv:1703.10135 , 2017. [16] OpenAI. (2024) Text-to-speech guide. [Online]. Available: https: //platform.openai.com/docs/guides/text-to-speech [17] G. Shopov, S. Gerdjikov, and S. Mihov, “Streamspeec...
https://arxiv.org/abs/2505.19206v1
Machine Learning . PMLR, 2023, pp. 28 492–28 518. [33] Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Yatabe, N. Morioka, M. Bacchiani, Y . Zhang, W. Han, and A. Bapna, “Libritts-r: A restored multi-speaker text-to-speech corpus,” in Proc. Interspeech 2023 , 2023, pp. 5496–5500. [34] V . Panayotov, G. Chen, D. Povey, and...
https://arxiv.org/abs/2505.19206v1
arXiv:2505.19209v1 [cs.CL] 25 May 2025MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search Zonglin Yang1,2, Wanhao Liu3,2, Ben Gao4,2, Yujie Liu2, Wei Li5, Tong Xie6, Lidong Bing7,Wanli Ouyang2,Erik Cambria1†,Dongzhan Zhou2† 1Nanyang Technological University2Shanghai...
https://arxiv.org/abs/2505.19209v1
hypothesis direction. We show that fine-grained scientific hypothesis discovery is a combinatorial search problem, as it requires selecting and composing a coherent set of concrete details from a vast space of plausible options—making it particularly challenging in practice. The difficulty is compounded by the fact tha...
https://arxiv.org/abs/2505.19209v1
reliability of the LLM’s internal reward signal in guiding fine-grained hypothesis discovery. Until now, the reward landscape guiding hypothesis search has been defined by a single LLM serving as the evaluator. We now turn to the third question ( Q3):whether defining this landscape with an ensemble of diverse LLMs of s...
https://arxiv.org/abs/2505.19209v1
. , d m}denotes the meaningful integration of edits d1, . . . , d mintohc, resulting in a coherent, fine-grained hypothesis. Each drepresents either: (1) the addition of a fine-grained detail to a concept iinhc, or (2) the deletion of a redundant concept from hc. We define two sets of edit candidates: D+, consisting of...
https://arxiv.org/abs/2505.19209v1
step : in the current search step, Recombination The th hierarchy, , The th hierarchy Add one at lev el ?no: yes: if no for a consecutive of times The th hierarchy feedback, ,, , , , Figure 1: Hierarchies designed for chemistry. Figure 1 illustrates an example hierarchical decomposition for chemistry, developed in coll...
https://arxiv.org/abs/2505.19209v1
Selection Instance-Level Component Choice Parametric Specification of Components Full Experimental ConfigurationsHierarchy 5 Hierarchy 3 Hierarchy 1Hierarchy 2Hierarchy 4: research background : hypothesis direction : final output from hierarchy : a candidate to consider for final output 𝐻𝑖−1=∅ if 𝑖≤1 {ℎ1,…,ℎ𝑖−1}if...
https://arxiv.org/abs/2505.19209v1
an aggregated estimate—approximating an average or soft maximum—of its lower-level subspace. For instance, when evaluating a coarse-grained concept like “hierarchical 3D copper,” the LLM may implicitly account for diverse fine-grained structural variants, some highly relevant, others ineffective. We hypothesize that th...
https://arxiv.org/abs/2505.19209v1
local optimum ( h1 i) is directly adopted as the output ( hi=h1 iin Figure 2). 3.2 Q1: How to Best Harness an LLM’s Internal Heuristics to Formulate the Fine-Grained Hypothesis It Itself Would Judge as the Most Promising Among All Possible Hypotheses It Might Generate? We frame this question as an optimization problem:...
https://arxiv.org/abs/2505.19209v1
(LLM) Detailedness (LLM) Feasibility (LLM) Overall (LLM) Overall (Expert) HHS v.s. Greedy Search Win 74.51% 41.18% 71.57% 67.65% 73.53% 76.47% Tie 18.63% 18.63% 28.43% 10.78% 18.63% 15.69% Lose 6.86% 40.20% 0.00% 21.57% 7.84% 7.84% HHS v.s. Greedy Search + Self-consistency Win 59.31% 42.16% 56.37% 48.53% 53.43% 74.51% ...
https://arxiv.org/abs/2505.19209v1
The hypothesis optimization in HHS depends on the “ hcur> h prev?” module (Figure 2), which provides the gradient signal. This raises the question: does a diverse ensemble of similarly capable LLMs enhance search performance compared to multiple instances of its strongest model? To answer this, we design three experime...
https://arxiv.org/abs/2505.19209v1
LLM, and a fourth instance of the same model aggregates these judgments by evaluating the underlying rationales and selecting the most justified preference. Crucially, this summarization step is not a simple majority vote. Instead, the LLM is explicitly instructed to assess the relative strength of reasoning across all...
https://arxiv.org/abs/2505.19209v1
finds higher-quality hypotheses than flat search methods. Experiments show that (1) HHS reliably discovers better local optima than baselines, (2) LLM-preferred hypotheses align more closely with expert ground truths, and (3) repeated use of the strongest model provides better reward landscapes than diverse ensembles. ...
https://arxiv.org/abs/2505.19209v1
Sergey Brin, Oliver Woodman, Marvin Ritter, Eric Noland, Minh Giang, Vijay Bolina, Lisa Lee, Tim Blyth, Ian Mackinnon, Machel Reid, Obaid 11 Sarvana, David Silver, Alexander Chen, Lily Wang, Loren Maggiore, Oscar Chang, Nithya Attaluri, Gregory Thornton, Chung-Cheng Chiu, Oskar Bunyan, Nir Levine, Timothy Chung, Evgeni...
https://arxiv.org/abs/2505.19209v1
and Xinya Du. Llm4sr: A survey on large language models for scientific research. arXiv preprint arXiv:2501.04306 , 2025. OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/ gpt-4o-mini-advancing-cost-efficient-intelligence/ , 2024. Accessed: 2025-05-16. Biqing Qi, Kaiyan Zhang, Haoxian...
https://arxiv.org/abs/2505.19209v1
performance? • Hypothesis Candidate (from HHS) : The development of a cost-effective N-type quasi-solid- state thermocell will be achieved through the strategic integration of three core components to enhance electricity production from low-grade heat ( ≤100°C): 1.Hierarchical Metal Electrodes : Constructed from a copp...
https://arxiv.org/abs/2505.19209v1
detailed and logically structured explanation, covering key aspects related to electrodes, redox pairs, and polymer gel media. Specifically, the preparation of Hierarchical Metal Electrodes is highlighted, noting the primary use of copper-containing electrodes, with a clear principle of providing a stable interface, wh...
https://arxiv.org/abs/2505.19209v1