text
string
source
string
Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. [79] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W Cohen. Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. arXiv preprint arXi...
https://arxiv.org/abs/2505.20128v1
SIGIR International Conference on Theory of Information Retrieval , 2024. [96] Zhengliang Shi, Weiwei Sun, Shuo Zhang, Zhen Zhang, Pengjie Ren, and Zhaochun Ren. Rade: Reference-assisted dialogue evaluation for open-domain dialogue. ArXiv , 2023. 15 Part I Appendix Contents A Practical Considerations and Societal Impac...
https://arxiv.org/abs/2505.20128v1
E E XSEARCH -Zoo: Extending E XSEARCH for Diverse Scenarios 34 E.1 Diverse Model Families and Scales . . . . . . . . . . . . . . . . . . . . . . . . 34 E.2 Extended Retrieval Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 F Prompt and Case Study 35 F.1 System Prompt in E XSEARCH . . . . . . . . . ....
https://arxiv.org/abs/2505.20128v1
search-augmented reasoning systems. A.2 Ethics Statement The research conducted in this paper centers around the development of a reasoning-augmented search framework. The proposed method enables Language Models (LLMs) to dynamically retrieve and reason over external information. In the process of conducting this resea...
https://arxiv.org/abs/2505.20128v1
encourage future work to explore safeguards, such as automated monitoring and ethical auditing mechanisms, in high-stakes applications. B Comparison with Prior Work In this work, the proposed method enables LLMs to perform interleaved reasoning and search (also referred to as agentic search in this work) and optimizes ...
https://arxiv.org/abs/2505.20128v1
on the generated trajectories, showing promising results on mathematical reasoning tasks. Similar ideas have also been applied in combination with REINFORCE Leave-One-Out (RLOO) methods [ 88,91,92]. More recently, other work [ 89,93] proposes self-rewarding methods, where the LLM itself is used via LLM-as-a-Judge promp...
https://arxiv.org/abs/2505.20128v1
a sequence z={(xi,di, ei)|i∈ [|z|]}, with the joint likelihood: p(z|x;θ) =|z|Y i=1p((xi,di, ei)|x,z<i;θ) (12) After the interleaved search and reasoning process, the LLM aggregates information from zto generate the final answer y∼p(y|x,z;θ). 19 Training Objective. The goal of EXSEARCH is to improve the LLM’s ability to...
https://arxiv.org/abs/2505.20128v1
ELBO (θt, θt) = log p(y|x;θt). By combining these two equations, we have: logp(y|x;θt+1)≥logp(y|x;θt). (23) Therefore, we have completed the proof of non-decreasing improvement for each training iteration. Lemma C.2 (Boundedness ).The sequence {logp(y|x;θt)}∞ t=1is upper-bounded. Proof. Since p(y|x;θ)∈[0,1], we natural...
https://arxiv.org/abs/2505.20128v1
Here, p(y,z|x)denotes the LLM generating a reasoning path zfollowed by a final answer y. Marginalizing over all possible (z, y)is typically intractable due to the large action space of the LLM. We now derive a variational surrogate for optimizing such a goal-oriented objective through a tractable lower bound. Propositi...
https://arxiv.org/abs/2505.20128v1
adopt two rule-based metrics, namely Exact Match (EM) and Accuracy (Acc.), as the function r(y)in Proposition 1. These metrics are commonly used in open-domain QA and provide direct, interpretable supervision. Figure 6 presents the performance of models trained using both the vanilla objective and the extended goal-ori...
https://arxiv.org/abs/2505.20128v1
logJ(ˆθ)−ELBO (ˆθ, q) =KL(q∥q∗) (37) withq∗(z, y)∝r(y)p(z, y|x;ˆθ). Putting it all together, we can obtain the following equation: logJ(θ∗)−ELBO (ˆθ, q)≤L 2∥θ∗−ˆθ∥2+KL(q∥q∗) (38) By rearranging this, we have: ∥θ∗−ˆθ∥2≤2 L· logJ(θ∗)−ELBO (ˆθ, q) ≤2 L·KL(q∥r·pˆθ) (39) 3https://en.wikipedia.org/wiki/Lipschitz_continuity...
https://arxiv.org/abs/2505.20128v1
and GPT-3.5 [ 39], as well as strong open-source models, such as DeepSeek-R1 [ 37], Qwen2.5 [ 40], QwQ-32B [ 41], LLaMA-3.3-70B [ 42], and Mistral-8x7B [ 43]. All of these models ex- hibit strong instruction-following and chain-of-thought reasoning capabilities, achieving remarkable performance on a wide range of natur...
https://arxiv.org/abs/2505.20128v1
process of a real system rather than directly relying on officially annotated golden documents. To synthesize pseudo training trajectories, we leverage existing datasets, such as HotpotQA [ 33], to generate pseudo training data for our method, as it is a widely used multi-hop QA dataset similar to the setting of our ag...
https://arxiv.org/abs/2505.20128v1
by Bauer Media Group in the USA . The magazine was started in 1989. It is based in Englewood Cliffs , New Jersey . In 2011 , the circulation of the magazine was 1 ,310 ,696 copies . Your Output : <THINK > When did the magazine " Arthur 's Magazine " start ? <SEARCH > [1] <RECORD > 1844 <THINK > When did the magazine " ...
https://arxiv.org/abs/2505.20128v1
step Tfor the thinking →search →recording iteration is set to 5. Table 6: Experimental Settings for Model Training ModelBatch SizeLearning RateCutoff LengthSchedulerGradient Accumulation Qwen-2.5-3B-instruct 4 2×10−68192 tokens Cosine 16 Qwen-2.5-7B-instruct 4 2×10−68192 tokens Cosine 16 Llama-3.2-3B-instruct 4 2×10−68...
https://arxiv.org/abs/2505.20128v1
34.11 17.23 13.13 25.34 18.24 31.75 25.62 Search-r1 [52] 40.17 36.86 27.12 32.67 4.07 8.30 16.25 23.49 21.90 25.33 Ours-Qwen2.5-7B 43.50 37.58 44.47 37.49 18.85 15.91 29.60 23.33 34.11 28.58 Ours-Llama3.1-8B 43.56 37.49 43.76 36.91 19.33 16.30 28.22 22.20 33.72 28.22 D.6 Supplementary Experimental Results Supplementary...
https://arxiv.org/abs/2505.20128v1
correct; 0 for incorrect). Each query is paired with the corresponding golden documents and ground truth answers from the original datasets, which serve as references for the human evaluators. We ask at least two annotators to evaluate the same case repeatedly. If there is a discrepancy between two annotators, ask a th...
https://arxiv.org/abs/2505.20128v1
Iteration5057647178 Training Performance w/ K=800 examples 0 1 2 3 4 Training Iteration5057647178 Training Performance w/ K=900 examples 0 1 2 3 4 Training Iteration5057647178 Training Performance w/ K=1000 examplesFigure 10: F1 score for EXSEARCH -Qwen-2.5-7B, which is initially empowered by varying amounts of warm-up...
https://arxiv.org/abs/2505.20128v1
retrieval and evidence selection, resulting in a four-step reasoning pattern: think→seek→rank→ record . Specifically, we implement this re-ranking following generative re-ranking techniques [ 6,64], where the LLM reads the retrieved documents and autoregressively generates a ranked list of selected identifiers (e.g., [...
https://arxiv.org/abs/2505.20128v1
model to simulate an intelligent search agent. The detailed content is shown below. This prompt instructs the model to decompose a complex query into sub-queries (thinking ), retrieve relevant documents from Wikipedia based on each sub-query ( search ), and extract factual answers from the retrieved content ( recording...
https://arxiv.org/abs/2505.20128v1
quantify the reliability of the annotations, we compute Cohen’s Kappa score to measure inter- annotator agreement. A Kappa value close to 1 indicates strong agreement, while a value near 0 suggests random agreement. In our human evaluation, the overall Kappa value is 0.771, demonstrating substantial agreement among the...
https://arxiv.org/abs/2505.20128v1
> four ============ Turn 3 ( answer aggregation ) ============ <FINAL > <SEARCH > end search . The final answer to the input question " Navarone Garibaldi is the half - brother of a singer who has been married how many times ?" Output : four Bad Case: Over-Searching due to Missed Reasoning. We present a concrete exampl...
https://arxiv.org/abs/2505.20128v1
director , playwright , film director , producer , screenwriter , poet , writer , and cartoonist . ============ Turn 5 ============ <THINK > Are there any shared occupations between Maher Sabry and Til Schweiger ? <SEARCH > [1] Title : Maher Sabry . Content : Maher Sabry is an Egyptian theater director , playwright , f...
https://arxiv.org/abs/2505.20128v1
city. This example highlights the need for disambiguation-aware reasoning mechanisms that delay answer generation until entity identity is sufficiently grounded. USER QUERY : Into Dust is a song by the alternative rock band formed in which city ? ( Golden answer : Santa Monica , California ) ========== <THINK > Which c...
https://arxiv.org/abs/2505.20128v1
searching after gathering sufficient evidence, while the failure case redundantly queries instead of reasoning. Step Successful Reasoning Over-Searching Failure Input Query Navarone Garibaldi is the half-brother of a singer who has been married how many times?What occupations are shared by Maher Sabry and Til Schweiger...
https://arxiv.org/abs/2505.20128v1
arXiv:2505.20133v1 [cs.CL] 26 May 2025AWEDIST: Attention-aware Embedding Distillation for New Input Token Embeddings Konstantin Dobler Hasso Plattner Institute konstantin.dobler@hpi.deDesmond Elliott University of Copenhagen de@di.ku.dkGerard de Melo Hasso Plattner Institute gerard.demelo@hpi.de Abstract Current langua...
https://arxiv.org/abs/2505.20133v1
token embedding (only this is trained):= original subtoken embeddings of new token:= context sequence token embeddings:= hidden states using original subtokens:= hidden states using single new token:= hidden states of original subtokens (except for last one); discarded in loss computationFigure 1: Illustration of AWEDI...
https://arxiv.org/abs/2505.20133v1
initialization for new tokens proposes some variation of a weighted mean or copy of existing embedding vectors [Mosin et al., 2023, Zeng et al., 2023, Mundra et al., 2024, Yamaguchi et al., 2024a,b, Remy et al., 2024]. Nonetheless, the resulting embeddings still require further tuning to achieve good results [Minixhofe...
https://arxiv.org/abs/2505.20133v1
and can require manual inspection [Elhage et al., 2022]. Instead, we propose to distill [Hinton et al., 2015, Snell et al., 2022] the impact that the multiple subtokens t1, . . . , t nhave on other tokens attending to them into a single token embedding e⋆. Our intuition is as follows: If we identify an embedding e⋆fort...
https://arxiv.org/abs/2505.20133v1
log-probability in the case of output embeddings. As we aim to have a fast method, we need a better approach. We propose two different approaches: (1) Our main approach is to simply retrieve snippets that contain our target tokens from a domain-specific or general corpus. This can be implemented efficiently using the a...
https://arxiv.org/abs/2505.20133v1
et al., 2020, Gee et al., 4 2023, Hasan et al., 2024, Singh et al., 2024]. This is because it is a particularly challenging domain with highly complex domain-specific terminology, which can serve as a benchmark to stress test embedding initialization methods. Therefore, we evaluate our method on a collection of standar...
https://arxiv.org/abs/2505.20133v1
[Sachidananda et al., 2021, Koto et al., 2021, Gee et al., 2022] as initialization for a new token, which has been shown to perform similarly to more sophisticated initialization methods [Minixhofer et al., 2024] that also use a weighted average of existing embeddings. Since our proposed method conducts a short optimiz...
https://arxiv.org/abs/2505.20133v1
al., 2022] 2.5 4.3 16.8 16.2 17.6 20.7 25.6 25.6 23.2 28.3 15.2 15.2 12.1 14.6 20.1 24.2 16.6 18.6 NTP 26.8 30.3 33.2 39.5 34.4 43.4 28.5 35.0 41.4 50.0 19.7 20.7 18.6 21.1 22.5 26.6 28.1 33.3 NTP++ [Lampinen and McClelland, 2018] 49.8 58.2 45.3 49.6 48.4 58.0 58.6 65.6 60.2 69.9 51.6 58.4 50.0 58.8 52.1 57.0 52.0 59.4...
https://arxiv.org/abs/2505.20133v1
future work. 6 Note that in Table 2, only for Llama-3.2-3B ,AWEDISThas performance on par with random initialization. Llama-3.2-3B(-i) are the only models in our lineup with tied embedding weights. Our objective does not explicitly enforce a bound on the norm of the new embedding, which in this case led to the failure ...
https://arxiv.org/abs/2505.20133v1
each new model and – in some cases – target domain. Also, we can actually use ZeTT-generated embeddings as a starting point and further tune them using our method for even better results. We analyze this possibility in Section 5.2. 5.2 In-Depth Analysis Multi-word tokens. In Table 3, we report LLM-as-a-Judge results on...
https://arxiv.org/abs/2505.20133v1
initialization (see “Random + AWEDIST” in Table 6). If a pretrained embedding prediction hyper-network is available, we can instead use these predicted embeddings as the starting point. We run this experiment using ZeTT; this combination further improves performance compared to using the subtoken mean (see Table 6). We...
https://arxiv.org/abs/2505.20133v1
65.9 65.4 64.3 ZeTT + NTP++ 62.8 66.1 66.8 65.3 ⋆Random + A WEDIST 62.5 65.3 65.9 64.6 ⋆(Mean +) A WEDIST 63.5 67.3 68.0 66.3 ⋆ZeTT + A WEDIST 64.2 67.9 68.4 66.8 ⋆(Mean +) A WEDIST-αNTP++ 63.2 67.2 68.4 66.3 ⋆ZeTT + A WEDIST-αNTP++ 64.1 67.9 68.7 66.9 Table 6: Biomedical benchmark results with dif- ferent starting poi...
https://arxiv.org/abs/2505.20133v1
In particular, a more localized identification of subtoken contextualization (e.g., specific attention heads that aggregate subtokens) is a promising direction for a more targeted distillation objective. 9 Acknowledgements We thank the German Federal Ministry for Education and Research (BMBF) for their compute grant th...
https://arxiv.org/abs/2505.20133v1
, pages 13440–13454, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.829. URL https://aclanthology.org/2023.emnlp-main. 829/ . C. M. Downey, Terra Blevins, Nora Goldfine, and Shane Steinert-Threlkeld. Embedding structure matters: Comparing methods to adapt multiling...
https://arxiv.org/abs/2505.20133v1
Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoychev, Niladri Chatterji, Ning Zhang, Olivier Duchenne, Onur Çelebi, Patrick Alrassy, Pengchuan Zhang, Pengwei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan, Punit Singh Koura, Puxin Xu, ...
https://arxiv.org/abs/2505.20133v1
Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kiran Jagadeesh, Kun Huang, Kunal Chawla, Kyle Huang, Lailin Chen, Lakshya Garg, Lavender A, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu, Liron Moshkovich, Luca Wehrstedt, Madian Khabsa, Manav Avalani, Manish Bhatt, Martynas Mankus, Matan Ha...
https://arxiv.org/abs/2505.20133v1
Dean. Distilling the Knowledge in a Neural Network, March 2015. URL http://arxiv.org/abs/1503.02531 . arXiv:1503.02531 [stat]. Hongzhi Huang, Defa Zhu, Banggu Wu, Yutao Zeng, Ya Wang, Qiyang Min, and Xun Zhou. Over-Tokenized Transformer: V ocabulary is Generally Worth Scaling, January 2025. URL http://arxiv.org/abs/250...
https://arxiv.org/abs/2505.20133v1
the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 3992–4006, Seattle, United States, July 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-main.293. URL https: //aclanthology.org/2022.naacl-main.293/ . Benjamin Minixhofer, Edoard...
https://arxiv.org/abs/2505.20133v1
Xia, Wenjie Li, and Roberto Navigli, editors, Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 3118–3135, Online, August 2021. Association for Computational Linguistics. doi:...
https://arxiv.org/abs/2505.20133v1
2020.emnlp-demos.6 . Pages: 38–45 original-date: 2018-10-29T13:56:00Z. Atsuki Yamaguchi, Aline Villavicencio, and Nikolaos Aletras. An Empirical Study on Cross- lingual V ocabulary Adaptation for Efficient Language Model Inference, September 2024a. URL http://arxiv.org/abs/2402.10712 . arXiv:2402.10712 [cs]. Atsuki Yam...
https://arxiv.org/abs/2505.20133v1
distillation training on samples from the original model’s learned data distribution of samples containing the new word. B Implementation Details B.1 Token Selection for Domain Adaptation When using benchmarks to evaluate the quality of new token embeddings, it is crucial to ensure that the new tokens actually occur fr...
https://arxiv.org/abs/2505.20133v1
roughly matches our measured run times. Note however that – as demonstrated in Section 5.2 – AWEDISTcan be significantly sped up while potentially even improving results by choosing an earlier target layer than the last one, yielding even faster run times than NTP++. However, our main version of AWEDISTis slightly more...
https://arxiv.org/abs/2505.20133v1
hyper-network pretraining would be beneficial, as our experiments show that this significantly outperforms next-token prediction for learning new embeddings (ZeTT uses next-token prediction). 6Note that we do not anneal the learning rate to zero, so this is not an artifact of the learning rate schedule. 18 Mistral-7B L...
https://arxiv.org/abs/2505.20133v1
conveys a similar enough meaning to the ground truth definition provided for a given word.<|eot_id|><|start_header_id|>user<|end_header_id|>,→ ,→ ,→ 4 5Remember to answer with one word either "Yes" or "No". 6 7### Instruction: 8Determine if the predicted definition conveys a similar meaning to the ground truth definiti...
https://arxiv.org/abs/2505.20133v1
arXiv:2505.20139v1 [cs.SE] 26 May 2025StructEval: Benchmarking LLMs’ Capabilities to Generate Structural Outputs Jialin Yang∗†1Dongfu Jiang∗†1Lipeng He1Sherman Siu1Yuxuan Zhang2 Disen Liao1Zhuofeng Li4Huaye Zeng1Yiming Jia1Haozhe Wang3 Benjamin Schneider1Chi Ruan5Wentao Ma1Zhiheng Lyu1Yifei Wang1 Yi Lu2Quy Duc Do1Ziyan...
https://arxiv.org/abs/2505.20139v1
A: 16px Query Metric Expert Review Generate Query & Metric GPT-4.1 - Keywords - Syntax Rules Query Metric 2 Rounds Manual Review Label Studio Query Metric & &Figure 2: The overall designed annotation pipeline of S TRUCT EVAL dataset et al., 2024), with limited emphasis on their ability to produce format-conforming stru...
https://arxiv.org/abs/2505.20139v1
StructEval Dataset In this section, we first present an overview of our STRUCT EVAL dataset and statistical analysis in subsection 2.1. Next, we elaborate on how we de- 2 Subset# Total Tasks# Total Examples# Avg Keywords# Avg VQA pairs SE-T-gen 5 250 7.9 - SE-T-conv 14 700 17.5 - SE-V-gen 13 650 11.1 7.9 SE-V-conv 12 4...
https://arxiv.org/abs/2505.20139v1
class is applied to the second table row? A:highlight •Q:What text is on the button at the bottom? A:Export PDF Figure 4: Example question, keywords, and VQA pairs for S TRUCT EVAL-V generation task 3 Rule Type Example Description Literal key access planet.name Checks if key name exists as a child of object planet . Ne...
https://arxiv.org/abs/2505.20139v1
for human refinement. Expert Review To ensure quality and correct- ness, we employ a two-pass human review process. Annotators first validate and refine the generated task queries and associated metrics. They are al- lowed to freely modify, add, or remove any part of the synthesized content to ensure task clarity, comp...
https://arxiv.org/abs/2505.20139v1
depth, list membership). Annotators translate each require- ment into a concrete dot-path rule using the syn-VQA Prompt Template You are given an image and a list of question-answer pairs. •For each pair, verify if the image content supports the expected answer based on the corresponding question. •Base your judgment s...
https://arxiv.org/abs/2505.20139v1
parameters. Inference Setup. All model generations are per- formed using LLM-Engine (Jiang, 2024), a uni- fied inference framework that supports both open- source backends (e.g., VLLM, SGLang, Together), and commercial APIs (e.g., OpenAI, Claude, Gem- ini). For open-source models, we specifically uti- lize the vLLM eng...
https://arxiv.org/abs/2505.20139v1
72.39 51.62 52.48 56.97 Qwen2.5-7B-Instruct (Team, 2024) 59.21 62.18 53.28 61.43 59.03 Qwen3-4B (Yang et al., 2025) 64.95 81.13 57.00 65.08 67.04 Close Source Gemini-1.5-pro (Team et al., 2023) 88.07 74.24 58.11 66.59 71.75 Gemini-2.0-flash (Team et al., 2023) 72.42 72.20 53.62 51.97 62.55 GPT-4.1-mini (OpenAI, 2025) 9...
https://arxiv.org/abs/2505.20139v1
and Markdown →Angular(see scores in subsec- tion A.2). Both closed-source and open-source models achieve low scores on these tasks, which typically require complex structural or visual rea- soning. Notably, the performance gap between closed-source and open-source models is even wider on these challenging subtasks, sug...
https://arxiv.org/abs/2505.20139v1
include gener- ating JSON for API responses (Geng et al., 2025), YAML or TOML for configuration files (Ped- direddy, 2024), HTML or React for UI compo- nents (Si et al., 2024), and LaTeX or Markdown for technical writing (Wen et al., 2024). Moreover, in data science, models are used to transform unstruc- tured descript...
https://arxiv.org/abs/2505.20139v1
highly capa- ble language model locally on your phone. Preprint , arXiv:2404.14219. Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, An...
https://arxiv.org/abs/2505.20139v1
, abs/2403.07974. Dongfu Jiang. 2024. Llm-engines: A unified and par- allel inference engine for large language models. https://github.com/jdf-prog/LLM-Engines . Max W.F. Ku, Thomas Chong, Jonathan Leung, Krish Shah, Alvin Yu, and Wenhu Chen. 2025. Theorem- explainagent: Towards multimodal explanations for llm theorem ...
https://arxiv.org/abs/2505.20139v1
Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Mil- lican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 . 10 Qwen Team. 2024. Qwen2.5: A party of foundation models. Maxim Tkachenko, Mikhail Malyuk, Andrey Holmanyuk, and Nikolai Liub...
https://arxiv.org/abs/2505.20139v1
HTML →Vue 40 Markdown →React 30 HTML →Angular 30 Markdown →Vue 25 Vue→React 15 Markdown →Angular 10 React→Angular 5 Table 6: Statistics of number examples for each task in all the 4 subsets of S TRUCT EVAL. 12 A.2 Subtask Performance Model T→JSON T→CSV T→TOML T→XML T→YAML Avg. Llama-3.1-8B-Instruct 78.82 81 .68 6 .76 5...
https://arxiv.org/abs/2505.20139v1
Table 9: StructEval-V Generation Scores (Part 2) 13 Model C→JSON J→CSV X→JSON J→XML Y→JSON J→YAML X→CSV Llama-3.1-8B-Instruct 34.14 95 .96 68 .62 56 .02 94 .00 92 .52 98 .98 Meta-Llama-3-8B-Instruct 31.40 48 .00 69 .24 55 .40 90 .00 74 .00 48 .26 Phi-3-mini-128k-Instruct 24.88 87 .28 8 .00 12 .40 23 .20 32 .80 33 .92 P...
https://arxiv.org/abs/2505.20139v1
56 .91 Phi-3-mini-128k-Instruct 65.60 42 .16 34 .65 33 .00 25 .10 41 .60 41 .23 Phi-4-mini-Instruct 92.44 57 .11 41 .05 55 .87 26 .50 71 .20 52 .48 Qwen-2.5-7B-Instruct 85.16 69 .20 80 .02 50 .87 35 .00 84 .60 61 .43 Qwen-3-4B 90.20 65 .31 83.05 68.13 34 .50 85 .00 65 .08 Gemini-1.5-pro 95.28 40 .62 86.65 64.00 49.80 8...
https://arxiv.org/abs/2505.20139v1
arXiv:2505.20144v1 [cs.CL] 26 May 2025SeMe: Training-Free Language Model Merging via Semantic Alignment Jian Gu Monash University jian.gu@monash.eduAldeida Aleti Monash University aldeida.aleti@monash.edu Chunyang Chen Technical University of Munich chun-yang.chen@tum.deHongyu Zhang Chongqing University hyzhang@cqu.edu...
https://arxiv.org/abs/2505.20144v1
tasks Orgad2024LLMsKM . In this work, we propose a new paradigm for model merging that eliminates the dependency on data and training. Leveraging semantic alignment across models—derived from their latent repre- sentations—we develop a data-free, computation- efficient merging method that is robust across tasks and fin...
https://arxiv.org/abs/2505.20144v1
in the latent space, termed as “semantic basis”, they share the same semantic meaning, as shown in Figure 1. Figure 1: Semantic association of vocabulary and latent space. For each color label on the vocabulary (left), there is a color semantic basis in the latent space (mid- dle). The semantics of the dark dot (indica...
https://arxiv.org/abs/2505.20144v1
modularization. Leveraging the semantics isotropy of LM latent Figure 2: Empirical Validation of Semantics Decompo- sition (CodeGen). space, we discuss two situations of stitching (het- erogeneous) modules, with progressive difficulties: (1) LM vocabularies are the same but LM-head matrices are different; (2) LM vocabu...
https://arxiv.org/abs/2505.20144v1
of ⃗ rymay require additional calibrations (to be confirmed by further exper- iments). 4.3 Semantic Alignment Due to differences in tokenizer vocabularies and response sequences across models, direct compari- son of these distributions is not straightforward. To address this, we realizes semantic alignment to solve two...
https://arxiv.org/abs/2505.20144v1
share the same vocabulary. Between modules, latent representations are the intermediate results and the medium of computa- tion. Assume the dimension of a latent space is d, the representation will be: r1,r2, ...,rd. Compared with the probabilities on vocabulary, the represen- tations in latent space have a much smalle...
https://arxiv.org/abs/2505.20144v1
chitecture, while merging seeks a compact, unified model in parameter space. In this work, we focus exclusively on parameter- 5 level model merging due to its scalability, inference efficiency, and growing importance for model reuse and multi-task generalization. 7 Conclusion In this paper, we have proposed the concept...
https://arxiv.org/abs/2505.20144v1
arXiv:2505.20152v1 [cs.CV] 26 May 2025Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models Kai Sun∗, Yushi Bai∗, Zhen Yang, Jiajie Zhang, Ji Qi, Lei Hou, Juanzi Li Tsinghua University Abstract Benefiting from contrastively trained visual encoders on large-scale natural ...
https://arxiv.org/abs/2505.20152v1
all hallucinate non-existent elements or misinterpret spatial relationships (e.g., △ABC , △ABE , and the concept of similar triangles), ex- 1 Model Training Training Data MM-MATH AltCLIP No No 23.8 AltCLIP In-Batch Randomly-sampled Negative (400K) 24.6 AltCLIP MMCLIP 1⃝: Retrieval-based Negative (100K) 26.6 AltCLIP MMC...
https://arxiv.org/abs/2505.20152v1
the pro- posed model outperforms all existing open-source models on GeoQA and MathVISTA, and achieves state-of-the-art performance on MM-MATH, sur- passing GPT-4o by 7.5%. We further analyze the impact of different types of hard negative samples. Notably, we find that negative samples constructed from authentic, exam-b...
https://arxiv.org/abs/2505.20152v1
2.5 Pro to modify the Python scripts based on these perturbed captions. These modified scripts produce diagrams that closely resemble the originals while aligning with the intended constraints. 2.2 Negative Captions Construction We propose two methods for constructing text- based hard negative samples. The first method...
https://arxiv.org/abs/2505.20152v1
fine-grained image understanding. While meth- ods like NegCLIP (Yuksekgonul et al., 2022) in- troduce hard negatives, they still rely heavily on in-batch samples. Considering the methods in Sec- tion 2.2 have allowed us to construct a large num- ber of hard negatives for each image, we propose MMCLIP training strategy,...
https://arxiv.org/abs/2505.20152v1
alignment dataset (Zhang et al., 2024c) to im- prove the model’s generalization for geometric problems. Next, we fine-tune the model follow- ing the MMCLIP training strategy, using 100K hard caption negative samples, each consisting of 10 rule-based and 30 retrieval-based negatives to enhance the model’s understanding ...
https://arxiv.org/abs/2505.20152v1
behind Chimera-Reasoner- 8B. As Chimera-Reasoner-8B was trained on GeoQA (Peng et al., 2024b), MMGeoLM- Qwen2.5-7B achieves the best performance among other models that were not trained on this dataset. The improved performance in geometric problem- solving demonstrates the effectiveness of our training approach. For t...
https://arxiv.org/abs/2505.20152v1
from 5 to 50. 4.2.2 Results As shown in Figure 3, LMMs trained with hard neg- ative samples consistently improve performance on the GeoQA and MM-MATH benchmarks. No- tably, on MM-MATH, Negative-images AltCLIP , trained with only 4K image negative samples, sur- passes Retrieved AltCLIP and Rule-based AltCLIP 6 Model Geo...
https://arxiv.org/abs/2505.20152v1
realistic images. Notably, the dataset includes only 4,021 image 7 Random RetrievalRule-based Negative-images Negative Sampling StrategyOriginal AltCLIP Random AltCLIP Negative-images AltCLIP Retrieved AltCLIP Rule-based AltCLIP Random + Retrieved Random + Rule-basedModel Variants0.09 0.01 0.04 0.10 0.95 0.35 0.31 0.06...
https://arxiv.org/abs/2505.20152v1
for geometric elements while enabling arbitrary ex- pansion, thereby improving geometric understand- ing in multimodal tasks. 6 Conclusion This paper introduced MMCLIP, a novel hard neg- ative contrastive learning framework to enhance LMMs’ geometric reasoning by refining their vi- sion encoders with systematically gen...
https://arxiv.org/abs/2505.20152v1
Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. 2023b. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition , pages 2657–2668. Jiahui Gao, Renjie Pi, ...
https://arxiv.org/abs/2505.20152v1
2025-05-19. Maitreya Patel, Abhiram Kusumba, Sheng Cheng, Changhoon Kim, Tejas Gokhale, Chitta Baral, and Yezhou Yang. 2024. Tripletclip: Improving composi- tional reasoning of clip via synthetic vision-language negatives. arXiv preprint arXiv:2411.02545 . Maitreya Patel, Naga Sai Abhiram Kusumba, Sheng Cheng, Changhoo...
https://arxiv.org/abs/2505.20152v1
ative contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808 . Kaicheng Yang, Jiankang Deng, Xiang An, Jiawei Li, Ziyong Feng, Jia Guo, Jing Yang, and Tongliang Liu. 2023. Alip: Adaptive language-image pre- training with synthetic caption. In Proceedings of the IEEE/CVF International Conference ...
https://arxiv.org/abs/2505.20152v1
elements; do not use any colored elements.2. Do not include any title in the figure.3. Ignore any auxiliary or extended lines that may be relevant for reasoning but are not explicitly part of the original figure.4. Only annotate the geometric elements (such as coordinates, lengths, and angles) explicitly mentioned in t...
https://arxiv.org/abs/2505.20152v1
image, leading to an incorrect final result. In contrast, MMGeoLM produces a correct solution, albeit through a rea- soning process different from the ground-truth an- swer. This highlights MMGeoLM’s capability to generate diverse problem-solving approaches. Fig- ure 11 illustrates errors in image-based reasoning for b...
https://arxiv.org/abs/2505.20152v1
arXiv:2505.20154v1 [cs.CL] 26 May 2025UORA: Uniform Orthogonal Reinitialization Adaptation in Parameter Efficient Fine-Tuning of Large Models Xueyan Zhang1,∗, Jinman Zhao2,†,∗, Zhifei Yang3, Yibo Zhong4, Shuhao Guan5, Linbo Cao1,Yining Wang2 1University of Waterloo,2University of Toronto, 3Peking University,4Independen...
https://arxiv.org/abs/2505.20154v1
(VeRA) freezes the two low-rank matrices via randomization and introduces two trainable scal- ing vectors, which again reduces the number of trainable parameters (Kopiczko et al., 2023). As a result of randomized frozen matrix initialization, VeRA requires a higher rank for comparable per- formance, which increases the...
https://arxiv.org/abs/2505.20154v1
zero inference latency. A collection of methods extends the LoRA con- cept. Dynamic LoRA (DyLoRA) allows for a range of training ranks, rather than a fixed one, to avoid the exhaustive search for the optimal rank (Valipour et al., 2022). Adaptive LoRA (AdaLoRA) adjusts the rank adaptively across different layers within...
https://arxiv.org/abs/2505.20154v1
2023), Linear Representation-Steering (Bi et al., 2025a),chunk-wise gradient computation (Li et al., 2025a) and fine-tune partial layers of the model (Fan et al., 2025). 2.3 Parameter Modeling It has been widely studied to model the param- eter distribution using randomized matrices for efficiency. Given the model spar...
https://arxiv.org/abs/2505.20154v1
the following formula. :vrand=Random (vold) (3) vnew=αvold+ (1−α)vrand (4) 3.2 Details Figure 1 right panel illustrates the pipeline of UoRA. Orthogonal Uniform Initialization. The impact of initialization methods is studied in VeRA. The initialization method preserves matrix expressivity while maintaining a well-condi...
https://arxiv.org/abs/2505.20154v1
•Adapter tuning - Following the naming convin- tion, AdapterHinjects adapters, with two fully connected layers and activation functions, be- tween the attention module, multi-layer percep- tron (MLP) module and the following residual connection (Houlsby et al., 2019). AdapterLre- duces the number of parameters by injec...
https://arxiv.org/abs/2505.20154v1
vector ⃗dand⃗bwith 0.1 and 0 respectively. Following prior work, UORA is applied to query and value projection matrices in the transformer blocks. Similar to VeRA, we use separate learning rate for the head and UORA layers to mitigate the additional scaling hyperparameter introduced in LoRA. Refer to Table 6 in Appendi...
https://arxiv.org/abs/2505.20154v1
of arithmetic reasoning problems in various formats: 1) the AddSub dataset contains simple math word problems on addition and subtraction, typically in- volving one-step arithmetic reasoning (Hosseini et al., 2014); 2) the MultiArith dataset is a collec- tion arithmetic word problems that require multiple operation ste...
https://arxiv.org/abs/2505.20154v1
both ViT base and large models, pre- trained on ImageNet-21K, on a wide range of datasets, including RESIS45 (Cheng et al., 2017), CIFAR100 (Krizhevsky et al., 2009), FOOD101 (Bossard et al., 2014), FLOWER102 (Nilsback and Zisserman, 2008). We include full fine-tuning and training classi- fication head only as baseline...
https://arxiv.org/abs/2505.20154v1
in the VeRA paper. 4.5.3 Impact of Threshold τ, Count k, and Interpolation Factor α The frequency of interpolation-based reinitializa- tion depends on the threshold τand the count k. The extent of reinitialization depends on the inter- polation factor α. After empirical study, we found that even though ( τ, k, α ) are ...
https://arxiv.org/abs/2505.20154v1
reasoning distractor. In Pro- ceedings of the 32nd ACM International Conference on Multimedia , MM ’24, page 8209–8218. ACM. Jiali Chen, Xusen Hei, Yuqi Xue, Zihan Wu, Jiayuan Xie, and Yi Cai. 2025. Classic4Children: Adapting Chinese literary classics for children with large lan- guage model. In Findings of the Associa...
https://arxiv.org/abs/2505.20154v1
Han Fang, Zhifei Yang, Yuhan Wei, Xianghao Zang, Chao Ban, Zerun Feng, Zhongjiang He, Yongxiang Li, and Hao Sun. 2023. Alignment and generation adapter for efficient video-text understanding. In ICCV (Workshops) , pages 2783–2789. Yujie Feng, Bo Liu, Xiaoyu Dong, Zexin Lu, Li-Ming Zhan, Xiao-Ming Wu, and Albert Lam. 20...
https://arxiv.org/abs/2505.20154v1
Conferences on Artificial Intelligence Organi- zation. Main Track. Gaurav Iyer, Gintare Karolina Dziugaite, and David Rolnick. 2024. Linear weight interpolation leads to transient performance gains. Transactions on Ma- chine Learning Research . Chen Jia. 2024. Adversarial moment-matching distilla- tion of large languag...
https://arxiv.org/abs/2505.20154v1