text
string
source
string
et al. Pangu- π: Enhancing language model architectures via nonlinearity compensation. arXiv preprint arXiv:2312.17276 , 2023. [40] Tianwen Wei, Jian Luan, Wei Liu, Shuang Dong, and Bin Wang. Cmath: Can your language model pass chinese elementary school math test?, 2023. [41] Guangxuan Xiao, Yuandong Tian, Beidi Chen, ...
https://arxiv.org/abs/2505.21411v2
Liqun Deng, Miao Rang, Minghui Xu, Nianzu Zheng, Pengfei Xia, Shixiong Kai, Tao Lü, Tianyu Guo, Tiezheng Yu, Wei He, Weizhe Lin, Wenjie Liu, Xing Li, Xiang Lu, Xinduo Liu, Xing Huang, Xu He, Xuan Li, Yao Wang, Yasheng Wang, Ye Tian, Yichun Yin, Yihan Hu, Yinfei Pan, Yixian Ren, Yongbing Huang, Yunsheng Ni, Yuxuan Sun, ...
https://arxiv.org/abs/2505.21411v2
will proceed step by step, interpreting the percentage as 35% of the preceding number (a common convention in many calculators and practical contexts). Step 1: Calculate 35% of 28.97 35% of28.97 = 0 .35×28.97 = 10 .1395 ... Step 3: Subtract 58.87 −3261.1695−58.87 =−3320.0395 Final Step: Rounding The original numbers (2...
https://arxiv.org/abs/2505.21411v2
−3310.25.</think> To solve the expression 28.97−35%−82×40−58.87, we follow these steps: 1. Convert the percentage to a decimal: 35% = 0.35 2. Perform the multiplication: 82×40 = 3280 3. Substitute the converted percentage and multiplication result back into the expression: 28.97 - 0.35 - 3280 - 58.87 4. Perform the sub...
https://arxiv.org/abs/2505.21411v2
list the flavors, and end with a strong call to action. Maybe add some emojis for visual appeal. Let me check if I have six exclamation points. Let’s count: "Introducing..." has one, "crafted..." has two, "No more..." has three, "Try..." has four, "Join..." has five, and "Sunnis..." has six. Wait, that’s only six. Mayb...
https://arxiv.org/abs/2505.21411v2
with a call to action. Sprinkle those exclamation marks throughout to keep the energy high. Wait, the user specified "short" pitch. So keep it concise but impactful. Maybe three to four sentences. Let me draft a sample: "Introducing Sunnis Ice Cream—finally, a treat that’s as kind to your tummy as it is to your taste b...
https://arxiv.org/abs/2505.21411v2
Towards Better Instruction Following Retrieval Models Yuchen Zhuang†∗, Aaron Trinh†∗, Rushi Qiang†∗, Haotian Sun†, Chao Zhang†, Hanjun Dai‡, Bo Dai† †Georgia Institute of Technology,‡precur.ai Dataset: https://huggingface.co/datasets/InF-IR/InF-IR Code: https://github.com/night-chen/InF_Embed Abstract Modern informatio...
https://arxiv.org/abs/2505.21439v1
the pressing need for an effective embedding-based instruction-aware retrieval model that simultaneously excels in both efficiently encoding and accurately interpreting complex instruction-query-passage interactions. Addressing this challenge requires high-quality training resources specifically tailored for instructio...
https://arxiv.org/abs/2505.21439v1
datasets. "I", "Q", and "P" denote "instruction", "query", and "passage", respectively. "−" denotes negative samples; for example, "I−" indicates contrasting instruction for negative sample generation. Notations are consistent across tables. Datasets Eval. Train (Q, P)+I+I−Q−P−Quality Check #I #Q #P Avg. |I| Avg. |Q| A...
https://arxiv.org/abs/2505.21439v1
hoc search systems to retrieve with user instructions when responding to complex queries (Wang et al., 2023a; Moreira et al., 2024; Su et al., 2023; Asai et al., 2023). Early attempts to incorporate instructions into retrieval systems have often relied on decoder-only LLMs, formulating the retrieval task as a specializ...
https://arxiv.org/abs/2505.21439v1
(Weller et al., 2024; Wang et al., 2022b; Oh et al., 2024). By incorporating instructions, the retrieval model flexibly adapts to diverse user intents, thereby enhancing personalization and utility of retrieved passages P+. 4InF-IR : Instruction-Following Information Retrieval Training Corpus 4.1 Data Curation In this ...
https://arxiv.org/abs/2505.21439v1
Additional details are available in appendix B. 0.700.750.800.850.900.95Cohen's Kappa0.84 0.780.800.92FollowIR-7B gpt-4o-mini gpt-4o o3-mini Figure 4: Cohen’s kappa from 100 random samples.Data Quality Check. To ensure the quality and semantic consis- tency of our synthetic data, we employ an advanced reasoning model, ...
https://arxiv.org/abs/2505.21439v1
entities within a shared d-dimensional embedding space: pi=g(Pi;θP),ii=g(Ii;θI,Q),qi=g(Qi;θI,Q), (2) where pi,ii,qi∈Rddenote the embedding for the passage, instruction, and query, respectively. Instruction-Aware Query Representation. Our primary goal is to improve the instruction-awareness of retrieval models by explic...
https://arxiv.org/abs/2505.21439v1
, (6) which flexibly enables any combination of univariate conditional modeling by selectively retaining the desired contrastive terms. ⋄Objective II (Multivariate Conditional Modeling) : Alternatively, we can ensure instruction- following by keeping instructions as part of the contrasting inputs. This naturally leads ...
https://arxiv.org/abs/2505.21439v1
49.2 30.1 34.0 53.2 75.3 10.5 Qwen2.5-3B-Inst 5.0 -1.3 9.7 2.4 6.6 -0.4 7.1 0.2 1.3 3.1 2.2 1.7 8.8 0.3 +InF-Embed 19.6 3.3 22.4 1.8 14.6 3.7 18.9 2.9 45.3 29.2 35.0 55.4 72.7 10.6 6 Experiments 6.1 Experiments Setup Evaluation Datasets. We conduct a comprehensive evaluation across the following representative instruct...
https://arxiv.org/abs/2505.21439v1
on Follow-IR (Weller et al., 2024) benchmarking multiple loss function designs and varying sizes of backbone LMs. Category ( →) Encoder Decoder Base Model ( →) ModernBERT Llama-3.2 Llama-3.2 Qwen2.5 Qwen2.5 Qwen2.5 Qwen2.5 Model Size ( →) 109M 1B 1B-Instruct 1.5B 1.5B-Instruct 3B 3B-Instruct Config. ( ↓) score p-MRR sc...
https://arxiv.org/abs/2505.21439v1
Qwen2.5-1.5B ✓ last 1 -0.06 e5-base qwen2.5-1.5B qwen2.5-1.5B -InstructLlama-3.2-1B Llama3.2-1B -Instruct012345Avg. p-MRRInF-IR InF-IR w/o filterFigure 6: Effect of quality filtering. Effect of Negative Pairs Synthesis. We analyze the impact of various training configurations in Table 4, using Qwen2.5-1.5B as the base ...
https://arxiv.org/abs/2505.21439v1
2018. URL https://arxiv.org/abs/1611 .09268 . 10 Chen, M., Li, T., Sun, H., Zhou, Y ., Zhu, C., Yang, F., Zhou, Z., Chen, W., Wang, H., Pan, J. Z., et al. Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470 , 2025. Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tay, Y ., F...
https://arxiv.org/abs/2505.21439v1
llama for multi-stage text retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pp. 2421–2425, 2024. Ma, Z. and Collins, M. Noise contrastive estimation and negative sampling for conditional models: Consistency and statistical efficiency. arXiv ...
https://arxiv.org/abs/2505.21439v1
and beyond. Foundations and Trends® in Information Retrieval , 3(4):333–389, 2009. Robertson, S. E., Walker, S., Jones, S., Hancock-Beaulieu, M. M., Gatford, M., et al. Okapi at trec-3. Nist Special Publication Sp , 109:109, 1995. Soboroff, I. Overview of trec 2021. In Proceedings of the Thirtieth Text REtrieval Confer...
https://arxiv.org/abs/2505.21439v1
S., Reddy A, S., Patro, S., Dixit, T., and Shen, X. Super-NaturalInstructions: Generalization via declarative instructions on 1600+ NLP tasks. In Goldberg, Y ., Kozareva, Z., and Zhang, Y . (eds.), Proceedings of the 2022 Conference on Empirical Methods in Natu- ral Language Processing , pp. 5085–5109, Abu Dhabi, Unite...
https://arxiv.org/abs/2505.21439v1
16 A.4 Ethical Statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 BInF-IR : Additional Data Curation Details 17 B.1 Additional Data Quality Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 B.2 Additional Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....
https://arxiv.org/abs/2505.21439v1
Impacts. Improved instruction-following retrieval systems could inadvertently amplify existing biases or misinformation if the training data inherently contains biased or incorrect information. Given that InF-IR andInF-Embed leverage synthetic generation techniques and LLMs trained on web-scale data, there remains a ri...
https://arxiv.org/abs/2505.21439v1
to differentiate samples. All other data curation procedures remain consistent across datasets. CInF-Embed : Additional Method Details C.1 Univariate Contrastive Loss Details Here are the detailed definitions of univariate contrastive objective functions: ℓuni P=−Ei∼Bh logexp( sim(pi,iqi,i))P m∼Bexp( sim(pm,iqi,i))i , ...
https://arxiv.org/abs/2505.21439v1
generate relevance judgments through prompting without retrieval-specific fine-tuning. •E5(Wang et al., 2022a) models (base and large) are dual-encoder retrieval systems fine-tuned on extensive weakly supervised contrastive pairs. They excel in embedding-based retrieval tasks, explicitly leveraging query and passage in...
https://arxiv.org/abs/2505.21439v1
as BM25, particularly in instruction-following tasks, highlighting the advantage of semantic embedding-based approaches. •Model Size and Effectiveness. Larger model sizes generally exhibit stronger retrieval and instruction-following performance. Models in the XL size category (over 5B parameters), such asGritLM-7B and...
https://arxiv.org/abs/2505.21439v1
24.6 -0.0 52.3 36.0 36.3 58.3 82.7 53.1 NV-Embed-v1 (7B) (2024) – – – – – – – – 45.0 31.5 30.8 43.0 84.7 47.0 repllama-v1-7b (2024) 24.0 -8.9 24.5 -1.8 20.6 1.3 23.0 -3.1 – – – – – – promptriever-llama2-7b (2025) 28.3 11.7 28.5 6.4 21.6 15.4 26.1 11.2 – – – – – – OpenAI-v3-large 27.2 -5.8 27.2 -2.0 21.6 -0.2 25.3 -2.7 ...
https://arxiv.org/abs/2505.21439v1
-1.05 11.04 1.67 13.27 0.09 16.77 4.95 14.11 2.71 12.68 1.95 14.52 3.20 w/ℓmulti P,IQ 10.36 -0.92 6.36 0.39 10.42 1.41 9.05 0.30 15.69 3.32 11.85 1.43 11.99 1.47 13.18 2.07 w/ℓmulti I,IQ 9.10 -1.97 6.83 0.90 9.15 0.72 8.36 -0.12 15.18 1.51 15.34 0.62 11.68 1.13 14.07 1.09 w/ℓmulti P,I,IQ 9.96 0.28 5.99 0.06 9.79 2.92 8...
https://arxiv.org/abs/2505.21439v1
3.62 16.41 3.00 w/ℓmulti P,IQ 18.12 4.42 21.80 2.82 13.20 3.23 17.71 3.49 15.60 1.93 19.11 1.41 11.76 1.14 15.49 1.49 w/ℓmulti I,IQ 19.03 -0.99 24.23 2.44 14.40 3.43 19.22 1.63 15.45 0.33 15.68 1.81 10.86 1.82 14.00 1.32 w/ℓmulti P,I,IQ 17.88 2.51 26.86 1.77 14.55 2.65 19.76 2.31 17.72 2.92 18.35 0.37 12.67 1.03 16.25 ...
https://arxiv.org/abs/2505.21439v1
which document is the original vs. synthetic based on these key distinctions . The goal is that when your NEW query is used with the ORIGINAL instruction , they should produce documents that are clearly distinguishable from the original document (at least 3 significant differences ). Please provide your answer in the f...
https://arxiv.org/abs/2505.21439v1
arXiv:2505.21451v1 [cs.CL] 27 May 2025Words Like Knives : Backstory-Personalized Modeling and Detection of Violent Communication Jocelyn Shen♣Akhila Yerukola♢Xuhui Zhou♢Cynthia Breazeal♣ Maarten Sap♢♠Hae Won Park♣ ♣Massachusetts Institute of Technology, Cambridge, MA, USA ♢Carnegie Mellon University, Pittsburgh, PA, US...
https://arxiv.org/abs/2505.21451v1
et al., 2023). While promis- ing, most existing AIMC systems are developed for public, often anonymized contexts—peer sup- port platforms or online debates—where speakers are strangers and little is known about each partic- ipant’s background. As such, these systems tend to operate without modeling interpersonal histor...
https://arxiv.org/abs/2505.21451v1
affect (Javed et al., 2024) or use linguistic and pragmatic features to detect an- ti/pro - social features in conversation (Zhang et al., 2018; Bao et al., 2021; Kasianenko et al., 2024). However, none of the aforementioned works ex- plore how breakdowns between close social part- ners are tailored to the relationship...
https://arxiv.org/abs/2505.21451v1
and violent communication types to obtain fine-grained labels for problematic or constructive communica- tion turns during human annotation. 4 P ERSONA CONFLICTS CORPUS We introduce the PERSONA CONFLICTS CORPUS , a dataset of realistic conflict and non-conflict sce- narios simulated using LLMs (Figure 3). Given the sen...
https://arxiv.org/abs/2505.21451v1
relationship backstories for each conver- sation: a positive backstory, which paints a chosen character in a less problematic light, and a neg- ative backstory, which paints the same character in a more problematic light. In particular, certain scenarios and ways of conveying backstory can in- fluence empathy towards a...
https://arxiv.org/abs/2505.21451v1
potential harm towards the listener (2) fine-grained labels ofNVC or VC communication types depending on problematic rating (3) how the turn will make the other character feel if they heard the statement (bet- ter/worse/the same). 5.1 Believability For believability of our simulated conversations and backstories across...
https://arxiv.org/abs/2505.21451v1
= 4 .18,p <0.0001 , Cohen’s d= 0.55). Additionally, with negative backstories, partici- pants found the character’s communication less understandable (t(232) = −4.70,p <0.0001 , Co- hen’s d=−0.61) and interpreted their behavior as expressing more negative intent towards the other character (t(232) = −4.95,p <0.0001 , C...
https://arxiv.org/abs/2505.21451v1
average for Likert ratings within each backstory condition (pos- itive/negative). We compute the F1 score between model outputs and human ratings. 7.2 Results and Discussion Table 2 reports model performance across vary- ing context conditions and relationship backsto- ries. Overall, models performed comparably on the ...
https://arxiv.org/abs/2505.21451v1
positive backstory condition: GPT- 4o (p < 0.0001 ,∆M= +0 .12), Gemini ( p < 0.0001 ,∆M= +0 .08), and LLaMA ( p <0.0001 , ∆M= +0 .08). Interestingly, only Gemini re- verses this pattern in the negative backstory condi- tion, significantly underpredicting emotional posi- tivity ( p <0.0001 ,∆M=−0.07), while GPT-4o and L...
https://arxiv.org/abs/2505.21451v1
through methodological improve- ments (Finch and Choi, 2024) as well as evaluating how relationship-informed mediation tools perform in live settings with human users. Second, we did not analyze every possible out- come measure collected during the human study. Our focus was primarily on a subset of key social- interpr...
https://arxiv.org/abs/2505.21451v1
Shah, Junjie Hu, and Timothy Rogers. 2024. Simulating Opinion Dynamics with Networks of LLM-based Agents. In Findings of the Association for Computational Linguistics: NAACL 2024 , pages 3326–3346, Mexico City, Mexico. Association for Computational Linguistics. Jean Decety and Claus Lamm. 2006. Human Empathy Through th...
https://arxiv.org/abs/2505.21451v1
pages 3605–3627, Mexico City, Mexico. Association for Computational Linguistics. Gauri Kambhatla, Matthew Lease, and Ashwin Ra- jadesingan. 2024. Promoting Constructive Delibera- tion: Reframing for Receptiveness. In Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 5110–5132, Miami, Florida...
https://arxiv.org/abs/2505.21451v1
Nature Machine Intelligence , 5(1):46–57. Number: 1 Pub- lisher: Nature Publishing Group. Jocelyn Shen, Joel Mire, Hae Won Park, Cynthia Breazeal, and Maarten Sap. 2024. HEART-felt Nar- ratives: Tracing Empathy and Narrative Style in Per- sonal Stories with LLMs. In Proceedings of the 2024 Conference on Empirical Metho...
https://arxiv.org/abs/2505.21451v1
Akhila Yerukola, Thomas David- son, Jena D. Hwang, Swabha Swayamdipta, and Maarten Sap. 2023b. COBRA Frames: Contextual Reasoning about Effects and Harms of Offensive Statements. In Findings of the Association for Com- putational Linguistics: ACL 2023 , pages 6294–6315, Toronto, Canada. Association for Computational Li...
https://arxiv.org/abs/2505.21451v1
Turn #1 (speaker 1 's first name): dialogue Turn #2 (speaker 2 's first name): dialogue A.3 Simulation Prompt (Conflict) Let's think step by step. Generate a 10 to maximum 15 turn conversation between {speaker} and {nonviolent_speaker}. The conversation can be shorter than 15 turns if the characters decide to leave the...
https://arxiv.org/abs/2505.21451v1
characters should use pet names or titles (e.g. "babe", "honey", "sweetie", "Mom", "Dad") ### ### Format the output as: Turn #1 (speaker 1 's first name): dialogue Turn #2 (speaker 2 's first name): dialogue A.4 Backstory Generation Let's think step by step. We are analyzing a conversation between two people, {speaker}...
https://arxiv.org/abs/2505.21451v1
state what the relationship between the characters is like, but SHOW through narrative and life events. 6. Each backstory should be at most a paragraph. This conversation happened in this scenario: {original_scenario} ### Speakers: {agent_1_name} and {agent_2_name} --- {agent_1_name} Profile --- {agent_1_data} --- {age...
https://arxiv.org/abs/2505.21451v1
try helping unpacking?” →[4] {backstory_instruction} DO NOT OUTPUT YOUR REASONING PROCESS, just output the result as a JSON object like below. Return just the JSON object. {{ "response": list of numbers (e.g., [2, 4]) }} """ LISTENER_IMPACT_INSTRUCTIONS = """How would the provided conversation turn (CONVERSATION TURN T...
https://arxiv.org/abs/2505.21451v1
arXiv:2505.21458v1 [cs.CL] 27 May 2025Do LLMs Need to Think in One Language? Correlation between Latent Language and Task Performance Shintaro Ozakiα,βTatsuya HiraokaγHiroto Otakeα,β Hirki Ouchiα,ηMasaru Isonumaβ,δ,ϵBenjamin Heinzerlingη,δKentaro Inuiγ,δ,η Taro WatanabeαYusuke Miyaoϵ,βYohei OsekiϵYu Takagiθ αNAISTβNII ...
https://arxiv.org/abs/2505.21458v1
deliberately disrupt the internal linguistic consistency of models by injecting adversarial prompts composed of multiple languages into the inputs and quantitatively evaluate how these changes affect downstream task performance. To understand this correlation, we evaluate several LLMs known to possess their proficient ...
https://arxiv.org/abs/2505.21458v1
languages that models rely on during reasoning. In contrast to those contributions, our study aims to systematically analyze how the latent language of LLMs influences output robustness and downstream task performance, addressing a critical gap in our understanding of language dependent reasoning. By exploring the corr...
https://arxiv.org/abs/2505.21458v1
in consistency for v. The function Score (v;θ)averages the degree of such disruptions: Score (v;θ) =L−1X l=1 P(θ) l,v·KL(θ) l,l+1+KL(θ) l,l+1·P(θ) l+1,v · 1(v∗(θ) l+1̸=v) L−1X l=1 P(θ) l,v+P(θ) l+1,v · 1(v∗(θ) l+1̸=v)(1) We define P(θ) l,vas the probability that the model uses latent language vat layer l. The term ...
https://arxiv.org/abs/2505.21458v1
consistency on model robustness have been developed. To address this issue, we construct a dataset, following prior studies [ 6,7] that demonstrate the feasibility of semi- automatically generating QA pairs using LLMs. Our work focuses on translation and geo-culture domains, which are particularly sensitive to the choi...
https://arxiv.org/abs/2505.21458v1
↓) Robustness ( ↑)r 0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0 LLM-jp-3JaJa Ja 0.06 En0.08 Ja0.09 Ja0.10 Ja0.11 0.27 0.26 0.27 0.24 0.24 -0.82 En Ja 0.09 Ja0.07 Ja0.11 Ja0.10 Ja0.11 0.13 0.22 0.13 0.06 0.04 -0.83 EnJa Ja 0.06 Ja0.07 En0.12 En0.12 En0.13 0.15 0.11 0.10 0.15 0.13 -0.17 En En 0.10 En0.11 En0.11 En0.12 En0.13...
https://arxiv.org/abs/2505.21458v1
injecting adversarial prompts. We systematically examine the proportion of the input length occupied by these adversarial prompts, as detailed in Appendix B.4, using ratios of 20%, 40%, 60%, 80%, and 100% of the model ’s maximum input token length (e.g., 32,768 tokens for Qwen2.5). By varying these ratios incrementally...
https://arxiv.org/abs/2505.21458v1
Language Figures 2, 3, and 4 illustrate the correlation between consistency and robustness in translation and geo-culture tasks. The results tend to align with the ideal line y=−x, indicating that a loss of robustness is often accompanied by a disruption in consistency. For example, in Figure 4, when the adversarial pr...
https://arxiv.org/abs/2505.21458v1
view that models do not necessarily need to operate in their preferred latent language to maintain internal consistency. 8 0.2 (en) 0.4 (en) 0.6 (en) 0.8 (en)1.0 (en)0.2 (en) 0.4 (en) 0.6 (en)0.8 (en)1.0 (en)0.2 (en) 0.4 (en) 0.6 (en) 0.8 (en)1.0 (en) 0.088 0.09 0.092 0.094 0.0960.250.260.270.280.290.30.310.320.33 ja (...
https://arxiv.org/abs/2505.21458v1
perturbations than previously thought. However, tasks may be heavily dependent on precise linguistic alignment, such as translation, are particularly sensitive to disruptions by adversarial prompts. In these cases, the introduction of adversarial prompts from less familiar languages largely degrades performance, likely...
https://arxiv.org/abs/2505.21458v1
Ruder, Denny Zhou, et al. Language models are multilingual chain-of-thought reasoners. arXiv preprint arXiv:2210.03057 , 2022. [13] Huiyuan Lai and Malvina Nissim. mCoT: Multilingual instruction tuning for reasoning con- sistency in language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings...
https://arxiv.org/abs/2505.21458v1
agents and llms. arXiv preprint arXiv:2407.18416 , 2024. [25] Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval , pages 1097–110...
https://arxiv.org/abs/2505.21458v1
experimental results, and reflect how much the results can be expected to generalize to other settings. •It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2.Limitations Question: Does the paper discuss the limitations of the work performed by t...
https://arxiv.org/abs/2505.21458v1
relies upon should be properly referenced. 4.Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and ...
https://arxiv.org/abs/2505.21458v1
The answer NA means that paper does not include experiments requiring code. •Please see the NeurIPS code and data submission guidelines ( https://nips.cc/ public/guides/CodeSubmissionPolicy ) for more details. •While we encourage the release of code and data, we understand that this might not be possible, so “No”is an ...
https://arxiv.org/abs/2505.21458v1
one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. •For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are...
https://arxiv.org/abs/2505.21458v1
could enable people to train models that generate Deepfakes faster. •The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following ...
https://arxiv.org/abs/2505.21458v1
4 and Appendix B.7. Guidelines: • The answer NA means that the paper does not release new assets. •Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. •The paper should discuss whethe...
https://arxiv.org/abs/2505.21458v1
the dominant language. In our study, we defined the LLC Score and identified the language with the lowest score as the dominant language, treating it as the latent language. However, other, potentially better, approaches may exist. Our approach was selected because it showed the strongest correlation with other potenti...
https://arxiv.org/abs/2505.21458v1
(2), M>ℓ(hℓ) =W⊤ ULayerNorm" hℓ+LX ℓ′=ℓFℓ′(hℓ′)# M>ℓ(hℓ)denotes the final logits if forward computation continues from layer ℓ.WU∈Rd×|V|is the unembedding matrix, and |V|is the vocabulary size. The sum represents all residual updates after layer ℓ. In Equation (3), LogitLens( hℓ) =W⊤ ULayerNorm( hℓ) The LogitLens remov...
https://arxiv.org/abs/2505.21458v1
question end with ", answer: " even though language is Japanese or Chinese. # Example: What is the capital of {country}?, answer: What is the highest mountain in {country}?, answer: What is the official language of {country}?, answer: What is the currency of {country}?, answer: Prompt for creating question with related...
https://arxiv.org/abs/2505.21458v1
are additional experimental results. As these results indicate, it is not possible to determine whether the inclusion of adversarial prompts directly affects consistency based on this table. However, it is evident that robustness and confidence tend to decline. 0.05 0.1 0.15 0.20.10.20.30.40.50.60.70.80.9 Incorrect Cor...
https://arxiv.org/abs/2505.21458v1
En0.08 En0.00 0.68 0.61 0.57 0.65 0.75 -0.86 Zh En 0.00 En0.00 En0.00 En0.06 En0.07 0.85 0.83 0.78 0.73 0.74 -0.88 Gemma3Ja EnJa En 0.02 En0.02 En0.02 En0.02 En0.02 0.78 0.77 0.76 0.30 0.59 0.76 En En 0.06 En0.06 En0.03 En0.06 En0.03 0.80 0.82 0.78 0.73 0.68 0.49 Zh En 0.03 En0.08 En0.07 En0.24 En0.10 0.80 0.79 0.69 0....
https://arxiv.org/abs/2505.21458v1
En En 0.08 En0.09 En0.09 0.23 0.17 0.20 -0.95 Zh En 0.09 En0.09 En0.09 0.28 0.27 0.29 -0.95 ZhEn En 0.10 Zh0.04 En0.10 0.01 0.01 0.01 0.62 Zh Zh 0.04 Zh0.04 Zh0.04 0.00 0.01 0.01 0.94 27 Table 8: Correlation between LLC Score and robustness in translation tasks. The value rrepresents the correlation coefficient for eac...
https://arxiv.org/abs/2505.21458v1
arXiv:2505.21465v1 [cs.CV] 27 May 2025ID-Align: RoPE-Conscious Position Remapping for Dynamic High-Resolution Adaptation in Vision-Language Models Bozhou Li Peking University Beijing, China libozhou@pku.edu.cnWentao Zhang Peking University Beijing, China wentao.zhang@pku.edu.cn Abstract Currently, a prevalent approach ...
https://arxiv.org/abs/2505.21465v1
scores between query and key diminish as their relative distance increases. Although generally as- sumed to be valid, some researchers have contested this property (Barbero et al., 2024). Our further analysis reveals that, based purely on RoPE’s math- ematical formulation, its effective behavior (e.g., long-term decay,...
https://arxiv.org/abs/2505.21465v1
encoded image features Fimage : Pimage =Projector (Fimage, Itext) (2) where Itextrepresents the text input. In certain ar- chitectures, such as BLIP-2 (Li et al., 2023a), Itext also interacts with Fimage at this stage. Following this, the LLM backbone processes Itextalongside Pimage , generating the corresponding outpu...
https://arxiv.org/abs/2505.21465v1
this set of resolutions could be defined as [(672, 672), (336, 672), (672, 336), (1008, 336), (336, 1008)]. •Select Appropriate Resolution. Given an input image with dimensions (H0, H0), the most suitable resolution is selected from a set of predefined resolutions based on its aspect ra- tio. •Adjust Input Image Resolu...
https://arxiv.org/abs/2505.21465v1
of long-range decay: for a query qat position mand a key kat position n, after encoding with RoPE, the dot product (Rmq)T(Rnk)generally de- creases as the absolute value of |m−n|in- creases. However, this property of RoPE is partially controversial, which we will discuss further in Section 3.1. •The value of θcontrols ...
https://arxiv.org/abs/2505.21465v1
representative of the actual situation. To investigate whether RoPE exhibits a long- range decay property, we adopted an empirical ap- proach. Specifically, we randomly sampled several data sequences from the WikiText (Merity et al., 2016) dataset. Then, for each layer, we randomly selected several q-k pairs before app...
https://arxiv.org/abs/2505.21465v1
user instruction to attend more to the bottom-left cor- ner of the image. The dynamic high-resolution method exacerbates this problem by increasing the difference in position IDs between the top-left and bottom-right tokens. Furthermore, studies have shown that in VLMs, image tokens inherently receive less attention (C...
https://arxiv.org/abs/2505.21465v1
Analysis From the perspective of attention distribution, in Figure 4c compared to 4b, the attention correspond- ing to the red region is no longer confined to certain unrelated areas but can focus on the magnets in the image. In Figure 4e compared to Figure 4d, the attention of ‘each pair’ can focus on the correspond- ...
https://arxiv.org/abs/2505.21465v1
: a method that aligns the position IDs of high-resolution embeddings with their correspond- ing low-resolution embeddings, preserving their relationship and constraining excessive growth in position IDs. We conducted experiments on the LLaV A-Next architecture, demonstrating the effec- tiveness of our approach. 8 7 Li...
https://arxiv.org/abs/2505.21465v1
beyond a fixed-length context. arXiv preprint arXiv:1901.02860 . Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. 2024. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv pr...
https://arxiv.org/abs/2505.21465v1
model an all-around player? In European conference on computer vi- sion, pages 216–233. Springer. Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. 2024f. Nvila: Effi- cient frontier visual language models. arXiv preprint arXiv:2412.04468...
https://arxiv.org/abs/2505.21465v1
preprint arXiv:2306.13549 . Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 11975–11986. Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and D...
https://arxiv.org/abs/2505.21465v1
their cor- responding high-resolution tokens is reduced. This adjustment not only brings related embeddings closer in terms of positional encoding but also effec- tively restricts the growth of position IDs. Conse- quently, this approach prevents the issue of position IDs increasing by thousands when processing a sin- ...
https://arxiv.org/abs/2505.21465v1
18 -- group_by_modality_length False \ 19 --bf16 True \ 20 -- output_dir ./ checkpoints /${ RUN_NAME } \ 21 -- num_train_epochs 1 \ 22 -- per_device_train_batch_size 8 \ 23 -- per_device_eval_batch_size 4 \ 24 -- gradient_accumulation_steps 4 \ 25 -- evaluation_strategy "no" \ 26 -- image_grid_pinpoints " [(336 , 672) ...
https://arxiv.org/abs/2505.21465v1
} \ 6 -- version plain \ 7 -- data_path ${ DATA_PATH } \ 8 -- image_folder ${ IMAGE_FOLDER } \ 9 -- vision_tower ${ VISION_TOWER } \ 10 -- mm_projector_type mlp2x_gelu \ 11 -- tune_mm_mlp_adapter True \ 12 -- unfreeze_mm_vision_tower False \ 13 -- mm_vision_select_layer -2 \ 14 -- mm_use_im_start_end False \ 15 -- mm_u...
https://arxiv.org/abs/2505.21465v1
\ 43 -- run_name ${ RUN_NAME } C.2 Benchmarks Focusing on the overall and various hierarchical capabilities of models, we primarily adopted three benchmarks—MMBench (Liu et al., 2024e), MME (Yin et al., 2023), and MMStar (Chen et al., 2024b). Additionally, SeedBench-2-Plus (Li et al., 2024) and AI2D (Kembhavi et al., 2...
https://arxiv.org/abs/2505.21465v1
and the benchmark data distribution. Overall, in the context of dynamic high-resolution, our method is superior. D.3 MMBench Leaf Tasks Coarse Perception: • Image Style • Image Topic• Image Scene • Image Mood • Image Quality Fine-grained Perception (Single-instance): • Attribute Recognition • Celebrity Recognition • Ob...
https://arxiv.org/abs/2505.21465v1
arXiv:2505.21467v1 [cs.CL] 27 May 2025Accelerating Diffusion Language Model Inference via Efficient KV Caching and Guided Diffusion Zhanqiu Hu∗Jian Meng∗Yash Akhauri Mohamed S. Abdelfattah Jae-sun Seo Zhiru Zhang Udit Gupta Cornell University {zh338, jm2787, ya255, ma839, js3528, zz284, ug28}@cornell.edu Abstract Diffu...
https://arxiv.org/abs/2505.21467v1
DLM is caused by the incompatibility with the caching strategy, which has been widely used in auto regressive models (ARM). Recently proposed Block Diffusion [ 1] have shown promising results on diffusion model with Key-Value caching strategy, whereas the overhead of additional fine-tuning and training hinder the fast ...
https://arxiv.org/abs/2505.21467v1
practicality and scalability of DLM. More importantly, diffusion and autoregressive-based language models are not orthogonal. The potential compatibility between DLM and Autoregressive Model is worth further exploration. 2.2 KV Caching for Autoregressive LLM The overhead of KV Cache has been one of the most critical me...
https://arxiv.org/abs/2505.21467v1
L) MHAWQ,WK,WVprojections O(d2) O(Ld2) Query ×Key (QK⊤) O(l2d/h) O(L2d/h) Attention Score ×Value O(ld/h) O(L2d/h) Output projection ( Wout) O(d2) O(Ld2) FFNW1projection O(d˙dff) O(Ldd ff) W2projection O(ddff) O(Ldd ff) Table 2: Compute complexity of Transformer modules: AR decodes a prefix of length lper token vs. DLM ...
https://arxiv.org/abs/2505.21467v1
proposed FreeCache: The variation of QPKPis minimal throughout the entire diffusion process with 256 steps. caching the Key and Value states of the “clean” tokens and reuse them across future steps without noticeable quality degradation, namely, the FreeCache . Initial clean token caching : Naturally, the input sequenc...
https://arxiv.org/abs/2505.21467v1
a lightweight coherence prior, enabling the model to safely unmask multiple tokens in parallel, without requiring any additional training or fine-tuning. Guided diffusion performs iterative unmasking by coordinating predictions from a diffusion language model (DLM) and a frozen autoregressive model (ARM). At each denoi...
https://arxiv.org/abs/2505.21467v1
fully unmasked x 1:while∃i:x[i] =mask do 2: logitsDLM←softmax( fθ(x)) 3: tokDLM←arg max logitsDLM 4: logitsAR←softmax( gϕ(tokDLM)) 5: tokAR←arg max logitsAR 6: letM= [i1, . . . , i m]be masked indices 7:k←size({tokDLM=tokAR}) 8: Unmask i1:ikwith tok DLM[i1:ik] 9: Correct ik+1with tok AR[ik+1] 10:end while 4 Experiments...
https://arxiv.org/abs/2505.21467v1
negligible accuracy degradation. For the tasks with long-context input prompt (e.g., GSM8K with 8-shot), the proposed method accelerates the diffusion process with 34.1×end-to-end average speedup on all the math problems, as presented in Table 3. With guided unmasking, the diffusion model exhibits stronger reasoning po...
https://arxiv.org/abs/2505.21467v1
or additional finetuning effort, the proposed method enables the end-to-end speedup by using the off-the-shelf diffusion model only. More importantly, for the first time, our method enables diffusion model to achieve the comparable and even better performance compared to the autoregressive language models while maintai...
https://arxiv.org/abs/2505.21467v1
Learning (ICML) , 2023. [15] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 , 2024. [16] Fangcheng Liu, Yehui Tang, Zhenhua Liu, Yunsheng Ni, Duyu Tang, Kai Han, and Yunhe W...
https://arxiv.org/abs/2505.21467v1
demonstrated the consistent speedup on the GSM8K dataset with 8-shot chain-of-thought (CoT). The baseline LLaDA [ 19] can optionally support semi-autoregressive unmasking by dividing the sequence into several blocks and unmasking them from left to right. The user-defined hyperparameter block length defines the size of ...
https://arxiv.org/abs/2505.21467v1