text string | source string |
|---|---|
query via reasoning R1-Router IterableR1-Router R1-Router LRM LRM LRMFigure 1: Comparison of Different RAG Architectures. LLMs (MLLMs) to interact dynamically with different Knowledge Bases (KBs) and fails to fully meet the information needs of MLLMs during reasoning [41, 46, 59, 22]. To address this limitation, existi... | https://arxiv.org/abs/2505.22095v1 |
Related Work RAG methods [ 36,43,62] with multiple Knowledge Bases (KBs) extend RAG approaches [ 28,44,40] by retrieving knowledge from various sources, thereby enhancing their ability to address more complex and diverse real-world applications [ 1,5,58,26]. These methods typically route the query to a predefined KB–su... | https://arxiv.org/abs/2505.22095v1 |
answer a: (q0, D)R1-Router (R)− − − − − − − → a, (1) where the input query q0is from different modalities, such as text or image. The knowledge base D comprises heterogeneous sources, including image-caption pairs, textual documents, and tabular data, enabling MLLMs to integrate knowledge from different KBs for query a... | https://arxiv.org/abs/2505.22095v1 |
a more accurate answer aiwithin the exploration trajectory. First, r(1)encourages the model to ask more targeted queries and correctly route them to relevant KBs during the reasoning step Ri(1≤i≤n). This involves two components: (i) a query reward rask(qi), which measures the semantic similarity between the generated q... | https://arxiv.org/abs/2505.22095v1 |
for golden reasoning trajectory construction. Our experiments incorporate three QA scenarios: Text QA (2WikiMultihopQA [ 21]), Visual QA (InfoSeek [ 16], Dyn-VQA [ 32], and WebQA [ 6]), and Table QA (Open-WikiTable [ 25] and Tab- Fact [ 14]). Specifically, we use 2WikiMultihopQA, InfoSeek, and Open-WikiTable for traini... | https://arxiv.org/abs/2505.22095v1 |
the image and generates answers without any retrieval process. In contrast, for LLMs without multimodal capa- bilities, the input image is first converted into a detailed textual description via image captioning provided by MLLMs. The vanilla RAG model incorporates retrieved evidence as additional con- text to assist M... | https://arxiv.org/abs/2505.22095v1 |
consistently outperforms all baseline models, achieving an average performance gain of approximately 7%. Notably, R1-Router shows consistent improvements by adaptively routing queries to different KBs to collect information, highlighting its strong generalization ability and potential to serve as a universal solution f... | https://arxiv.org/abs/2505.22095v1 |
compared, including Prompt, SFT, and Step-GRPO. methods, suggesting that its benefits stem not only from routing precision but also from more effec- tive reasoning and intermediate query generation. By further incorporating self-routing, R1-Router (Step-GRPO) achieves an additional 3% improvement. This confirms the eff... | https://arxiv.org/abs/2505.22095v1 |
a VQA task. As illustrated in Figure 4, we show the step-by-step reasoning process of R1-Router for a query that asks for the closest parent taxonomy of the bird depicted in the image. In the first reasoning step, R1-Router initiates a deliberation process and decomposes the task into a subgoal: conducting a visual ana... | https://arxiv.org/abs/2505.22095v1 |
Recognition , pages 1818–1826, 2024. [6]Yingshan Chang, Mridu Narang, Hisami Suzuki, Guihong Cao, Jianfeng Gao, and Yonatan Bisk. Webqa: Multihop and multimodal qa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16495–16504, 2022. [7]Danqi Chen, Adam Fisch, Jason Weston, an... | https://arxiv.org/abs/2505.22095v1 |
multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060 , 2020. [22] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In Proceedings of the 2023 Conference on E... | https://arxiv.org/abs/2505.22095v1 |
Conference on Empirical Methods in Natural Language Processing: Industry Track , pages 1001–1015, 2024. [38] Xiaoye Qu, Yafu Li, Zhaochen Su, Weigao Sun, Jianhao Yan, Dongrui Liu, Ganqu Cui, Daizong Liu, Shuxian Liang, Junxian He, et al. A survey of efficient reasoning for large reasoning models: Language, multimodalit... | https://arxiv.org/abs/2505.22095v1 |
your llm? In Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 2572–2585, 2024. [54] Peng Xia, Kangyu Zhu, Haoran Li, Hongtu Zhu, Yun Li, Gang Li, Linjun Zhang, and Huaxiu Yao. Rule: Reliable multimodal rag for factuality in medical vision language models. In Proceedings of the 2024 Conferen... | https://arxiv.org/abs/2505.22095v1 |
dump 20241020 as the source. Following prior work [ 7,49], we extract the textual content from the dump and segment each article into multiple disjoint text blocks of 100 words each, which serve as the basic retrieval units. To ensure the quality of the text KB, we discard any text blocks containing fewer than 7 words,... | https://arxiv.org/abs/2505.22095v1 |
- VQA InfoSeek [16] 1,000 - VQA Dyn-VQA [32] 715 - VQA WebQA [6] 1,000 - TableQA Open-WikiTable [25] 1,000 - TableQA TabFact [14] 1,000 - reasoning trajectories for VQA (totaling 545 reasoning steps), and 139 reasoning trajectories for Table QA (totaling 865 reasoning steps). In addition, we utilized a development set ... | https://arxiv.org/abs/2505.22095v1 |
the performance of QA systems in existing works [2, 31]. A.4 Prompt Templates Used in R1-Router The 8-shot prompt templates used by the R1-Router (Prompt) and R1-Router (SFT) methods for intermediate query generation and retriever selection across three task scenarios are shown in Figure 6, Figure 7, and Figure 8. The ... | https://arxiv.org/abs/2505.22095v1 |
generation are shown in Figure 16 and Figure 17. The maximum number of iterations is set to 3. Search-O1. We implement Search-O1[ 30] based on their official codebase4and adopt R1-Distill- Qwen-7B [ 20] as the backbone reasoning model. When the model generates the special tokens <|begin_search_query|> and<|end_search_q... | https://arxiv.org/abs/2505.22095v1 |
of Emily Dickinson is Amherst, Massachusetts. Notably, R1-Router ceases generating further intermediate queries once the original question can be answered with the accumulated knowledge, demonstrating its ability to adaptively determine when sufficient information has been gathered for effective reasoning. The final ca... | https://arxiv.org/abs/2505.22095v1 |
Answer Generation. We use this prompt template for all R1-Router models, including R1-Router (SFT), R1-Router (Prompt) and R1-Router (Step-GRPO). 24 Q:{question}Q:{question}ImageCaption:{caption}Belowarerelatedinformationsearched,whichmaybehelpfulforansweringthequestionlateron:{document}Question:{question}Vanilla Model... | https://arxiv.org/abs/2505.22095v1 |
MemOS: An Operating System for Memory-Augmented Generation (MAG) in Large Language Models (Short Version) Zhiyu Li∗,Shichao Song1,3,∗,Hanyu Wang1,3,∗,Simin Niu3,∗,Ding Chen4,∗,Jiawei Yang1,3,Chenyang Xi1,Huayi Lai3,Jihao Zhao3,Yezhaohui Wang1,Junpeng Ren1, Zehao Lin1,Jiahao Huo1,Tianyi Chen2,Kai Chen1,Kehang Li2,Zhiqia... | https://arxiv.org/abs/2505.22101v1 |
perceive and generate to those that remember, adapt, and grow over time . 2 Memory in Large Language Models Stage 1Stage 2 Stage 3 MemOS ImplicitExplicitLong -termShort -term abcdea b c d eMem Hippocampus Neocortex Mindmap API Add Mem Delete MemLLM MemHuman -like LLM MemMem Management Modify MemUpdate Mem Figure 1 Memo... | https://arxiv.org/abs/2505.22101v1 |
this end, we introduce MemOS —a prototype system designed to support a new memory-centric training paradigm , where learning and inference are no longer separate phases but part of a unified, memory-driven process. MemOS not only enables structured memory storage, interface-level invoca- tion, and lifecycle management,... | https://arxiv.org/abs/2505.22101v1 |
Knowledge & Drug Information Activation Memory •Inference -Coupled •State -Responsive e.g. Doctor's Expressive Patterns Plaintext MemoryParameter Memory •Implicitly Embedded •Statically Encoded e.g. Symptom-to - Mechanism Reasoner Figure 3 Transformation paths among three types of memory, forming a unified, controllabl... | https://arxiv.org/abs/2505.22101v1 |
serves as the system entry point, responsible for parsing natural language requests, identifying memory-related intents, and invoking standardized Memory APIs. The built-in MemReader component translates user inputs into structured memory operation chains. The Operation Layer functions as the central controller, orches... | https://arxiv.org/abs/2505.22101v1 |
isolation, and extensibility to future multi-modal scenarios. 4.4 System Execution Flow As illustrated in Figure 6, a MemOS execution begins with a user prompt or triggered task, parsed by MemReader into a structured Memory API call. This call initiates a pipeline, where context and state are passed viaMemoryCube units... | https://arxiv.org/abs/2505.22101v1 |
forward to advancing the frontiers of MemOS in collaboration with the community, making memory a first-class computational resource in the age of general-purpose AI. References [1]Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, ... | https://arxiv.org/abs/2505.22101v1 |
and smaller model sizes. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Findings oftheAssociation forComputational Linguistics: ACL2023, pages 8003–8017, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-acl.507. URL https: //aclanthology.org/2023.fi... | https://arxiv.org/abs/2505.22101v1 |
2024. URL http://arxiv.org/ abs/2308.10248 . arXiv:2308.10248 [cs]. [28]Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma. Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories. In Proceedings... | https://arxiv.org/abs/2505.22101v1 |
arXiv:2505.22113v1 [cs.CL] 28 May 2025THINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models Zhiyuan Li1, Yi Chang1,2,3, Yuan Wu1* 1School of Artificial intelligence, JiLin University 2Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, JiLin Universit... | https://arxiv.org/abs/2505.22113v1 |
serves as a critical indicator of LRMs’ intrinsic reasoning 1 Dataset Overview Math Physics Chemistry Question Answer ★Key StepLet $\omega$ be a complex number such that $\omega^3 = 1.$ Find all possible values of ... 1 "step1: Identify that ω³ = 1 means ω is a cube root of unity.", "step2: The three cube roots of unit... | https://arxiv.org/abs/2505.22113v1 |
resource distribution during inference, consequently constraining their overall performance optimization. To bridge this critical research gap, we introduce Think-Bench , a multidisciplinary dataset specifi- cally designed for comprehensively evaluating the reasoning efficiency and accuracy of LRMs. Our benchmark featu... | https://arxiv.org/abs/2505.22113v1 |
Since a single question may admit multiple logically valid reasoning paths, all reasonable and logically consistent solutions are considered and included. 3 Evaluation Strategy Recent studies have revealed that LRMs frequently exhibit low reasoning efficiency (Sui et al., 2025; Chen et al., 2024). However, there is cur... | https://arxiv.org/abs/2505.22113v1 |
have attracted growing interest for their human-like capacity for extended, reflective rea- soning. Through advanced long CoT and test-time scaling methods, these models iteratively evalu- ate multiple reasoning paths before finalizing an- swers (Chen et al., 2025; Muennighoff et al., 2025). However, as test-time scali... | https://arxiv.org/abs/2505.22113v1 |
thought," and "wait a moment." A higher count may indicate instability in reasoning or a tendency toward exploratory behavior. 3.2 CoT Quality Evaluation As LLMs increasingly adopt CoT reasoning strate- gies, assessing the quality of their internal reason- ing processes has emerged as a critical research challenge (Jia... | https://arxiv.org/abs/2505.22113v1 |
its strong performance in multi-turn reasoning tasks. We also conduct a comprehensive assessment of the DeepSeek-R1 family (Guo et al., 2025), includ- ing the full-scale DeepSeek-R1 and its distilled Qwen-1.5-based variants at 1.5B, 7B, 14B, and 32B scales, all explicitly optimized for efficient multi- step reasoning. ... | https://arxiv.org/abs/2505.22113v1 |
quality. Regarding CoT quality, Qwen3-235b-a22b and Ernie-x1-turbo-32k stand out by achieving the high- est reflection quality scores, with values of 92.16% and 90.97%, respectively. They also demonstrate top-tier precision at 86.97% and 88.67%, and re- call rates of 85.80% and 82.03%. These impressive results highligh... | https://arxiv.org/abs/2505.22113v1 |
ERNIE-X1- Turbo-32K, exhibited an unusual issue of generat- ing empty outputs. This issue primarily manifested in the model generating only intermediate reason- ing content without producing a final answer for certain questions. See Figure 8 in Appendix C.2 for a concrete example. Potential causes include limitations i... | https://arxiv.org/abs/2505.22113v1 |
for 2+ 3=? on the overthinking of o1-like llms. arXiv preprint arXiv:2412.21187 . Alejandro Cuadron, Dacheng Li, Wenjie Ma, Xingyao Wang, Yichuan Wang, Siyuan Zhuang, Shu Liu, Luis Gaspar Schroeder, Tian Xia, Huanzhi Mao, and 1 others. 2025. The danger of overthinking: Exam- ining the reasoning-action dilemma in agenti... | https://arxiv.org/abs/2505.22113v1 |
Yuan, Hongyi Liu, An- drew Wen, Shaochen Zhong, Hanjie Chen, and 1 others. 2025. Stop overthinking: A survey on ef- ficient reasoning for large language models. arXiv preprint arXiv:2503.16419 . Mirac Suzgun, Nathan Scales, Nathanael Schärli, Se- bastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V Le... | https://arxiv.org/abs/2505.22113v1 |
neural information processing systems , 36:11809–11822. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023b. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR) . Zhuosheng Zhang, Aston Zhang, Mu Li, and... | https://arxiv.org/abs/2505.22113v1 |
logic, and non-monotonic logic (Patel et al., 2024). Additionally, large reasoning models tend to exhibit overthinking behavior during chain-of- thought reasoning, where excessively long and un- necessary reasoning steps are generated even for simple or ill-posed problems (Sui et al., 2025; Chen et al., 2024; Fan et al... | https://arxiv.org/abs/2505.22113v1 |
235b-a22b. Model nameThought Num Efficiency Useful Tokens Reflection Tokens Chemistry Physics Math Chemistry Physics Math Chemistry Physics Math Chemistry Physics Math C-3-7-sonnet 0.16 0.23 0.58 0.47 0.52 0.5 339.85 411.6 695.93 404.35 425.03 798.55 Ds-r1-distill-qwen-1.5b 8.2 8.65 6.35 0.31 0.35 0.53 842.08 1243.98 2... | https://arxiv.org/abs/2505.22113v1 |
elimination. Let's see. If I subtract the first equation from the second, maybe that will eliminate the eraser?\n\nLet me try that. So, subtract (3p + e) from (5p + e):\n\n(5p + e) - (3p + e) = 1.82 - 1.24\n\nSimplifying the left side: 5p - 3p + e - e = 2p. The right side is 1.82 - 1.24, which is 0.58. So, 2p = 0.58. T... | https://arxiv.org/abs/2505.22113v1 |
\\tan A \\) because they are involved in the expression \\ ( \\sec^n A + \\tan^n A \\). Let me start by using the given equation \\( \\ tan A = 2 \\cos A \\).\n\nI know that \\( \\tan A = \\frac{\\sin A}{\\cos A} \\), so substituting that into the equation gives ... Response None Figure 8: Failure Case Example in the D... | https://arxiv.org/abs/2505.22113v1 |
Task Overview Given a solution with multiple reasoning steps for a text problem, reformat it into well-structured steps and evaluate their correctness. # Step 1: Reformatting the Solution Convert the unstructured solution into distinct reasoning steps while: - Preserving all original content and order - Not adding new ... | https://arxiv.org/abs/2505.22113v1 |
format, such as "First/Second/Third." This is the **ONLY** content modification allowed. * Avoid making paragraphs too lengthy, as long paragraphs might contain multiple reasoning steps that should be paragraphed separately. * Disregard the accuracy of the solution content. Do **NOT** alter any of the original solution... | https://arxiv.org/abs/2505.22113v1 |
arXiv:2505.22118v1 [cs.CL] 28 May 2025Multilingual vsCrosslingual Retrieval of Fact-Checked Claims: A Tale of Two Approaches Alan Ramponi,1*Marco Rovera,1∗Robert Moro,2Sara Tonelli1 {alramponi,m.rovera,satonelli}@fbk.eu ,robert.moro@kinit.sk 1Fondazione Bruno Kessler, Trento, Italy 2Kempelen Institute of Intelligent Te... | https://arxiv.org/abs/2505.22118v1 |
a narrative arising in their country was already present and debunked in other countries (or other languages of their country), so they need to retrieve fact-checked claims in a language that is 1 different from their input claim. In this case the required approach is crosslingual . Figure 1 illus- trates this last cas... | https://arxiv.org/abs/2505.22118v1 |
the relative popularity of the task, there is a range of existing relevant datasets as summarized in recent surveys by Panchendrarajan and Zubiaga (2024) and Srba et al. (2024). These resources differ in the number of included languages, data volume, means of identification of pairs between in- put social media posts a... | https://arxiv.org/abs/2505.22118v1 |
from OCRed images as it is too noisy) in their original languages. We include only languages represented with at least 180 posts and keep only the fact-checked claims that have at least one paired post. Next, we split posts, fact-checked claims, and pairs into training, development, and test sets strat- ified by langua... | https://arxiv.org/abs/2505.22118v1 |
and fact- checks in the dataset, there is not a single and well- established way to select negative examples for training. Previous work has mainly focused on ran- dom sampling (Pikuliak et al., 2023), i.e., creating a negative example by pairing a given post with a fact-checked claim randomly picked from those not ass... | https://arxiv.org/abs/2505.22118v1 |
and further compare performance for 14 addi- tional models in the unsupervised setting.8All configurations are evaluated using Pair Success at 10 (S@10) and Mean Reciprocal Rank at 10 (MRR@10). The former measures, for each post, whether the paired claim appears in the top-10 retrieved results, while the latter also co... | https://arxiv.org/abs/2505.22118v1 |
in pure re- trieval, and to enable comparison with the supervised approach (multilingual-e5-large ). 13We rely on a single Tesla V100-SXM2-32GB GPU.12345 100.85 0.84 without fine-tuning k(# negative examples)S@10 random similarity topic Figure 2: Multilingual S@10 performance across neg- ative sampling strategies and n... | https://arxiv.org/abs/2505.22118v1 |
retrieval, in terms of S@10 (Figure 3 – top, orange bars), it boosts the performance of weaker and mod- erately increases performance in mid-performing models, but it reduces the performance of high- performing models. This is further confirmed by looking at MRR@10, which shows that cross- encoder re-ranking, while inc... | https://arxiv.org/abs/2505.22118v1 |
cases. 6.2 Supervised Results For supervised experiments, we observe that fine- tuning models using similarity as a negative 6 0.60.81 0.763 0.5762 0.7687 0.7516 0.6416 0.7756 0.7912 0.7951 0.7769 0.7971 0.7525 0.8462 0.7879 0.7863 0.7881 0.81680.7731 0.695 0.7731 0.7811 0.6796 0.7938 0.8055 0.8039 0.7922 0.8075 0.7692... | https://arxiv.org/abs/2505.22118v1 |
0.5955 ±0.0002 0.4758 ±0.0008 topic 0.6206 ±0.0003 0.5014 ±0.0012 similarity 0.6320 ±0.0002 0.5157 ±0.0000 retrieve 0.6416 0.5188 c-e re-rank 0.6796 0.5779 Table 3: Multilingual andcrosslingual S@10 per- formance across negative sampling strategies ( random , topic ,similarity ) for fine-tuned models on the test set co... | https://arxiv.org/abs/2505.22118v1 |
differ- ent, with crosslingual retrieval being much more challenging. Unsupervised learning with LLM- based re-ranking yields the best results, even out- performing the best supervised approach. Overall, our study highlights the importance of a thorough evaluation of embeddings models and the impact of re-ranking and n... | https://arxiv.org/abs/2505.22118v1 |
Slovak Research and Develop- ment Agency under the project Modermed, GA No. APVV-22-0414, and the PNRR project FAIR – Fu- ture AI Research (PE00000013), under the NRRP MUR program funded by NextGeneration EU. The authors also wish to acknowledge the TAI- LOR project funded by the European Union under the EU Horizon 202... | https://arxiv.org/abs/2505.22118v1 |
Carolina Scarton. 2024. Euvsdisinfo: A dataset for multilingual detection of pro-kremlin disinformation in news articles. In Proceedings of the 33rd ACM International Conference on Informa- tion and Knowledge Management , CIKM ’24, page 5380–5384, New York, NY , USA. Association for Computing Machinery. Chaofan Li, Zhe... | https://arxiv.org/abs/2505.22118v1 |
Processing . Associa- tion for Computational Linguistics. Shaden Shaar, Nikolay Babulkov, Giovanni Da San Mar- tino, and Preslav Nakov. 2020. That is a known lie: Detecting previously fact-checked claims. In Pro- ceedings of the 58th Annual Meeting of the Asso- ciation for Computational Linguistics , pages 3607– 3618, ... | https://arxiv.org/abs/2505.22118v1 |
Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. 2024. Qwen2 techni- cal report. Preprint , arXiv:2407.10671. Puxuan Yu, Luke Merrick, Gaurav Nuti, and Daniel Campos. 2024. Arctic-embed 2.0: Multilingual retri... | https://arxiv.org/abs/2505.22118v1 |
( cs), Danish (da), German ( de), Modern Greek ( el), English (en), Spanish ( es), Persian ( fa), Finnish ( fi), French ( fr), Hindi ( hi), Croatian ( hr), Hungarian (hu), Indonesian ( id), Italian ( it), Kazakh ( kk), Korean ( ko), Macedonian ( mk), Malayalam ( ml), Malay ( ms), Burmese ( my), Nepali ( ne), Dutch ( nl... | https://arxiv.org/abs/2505.22118v1 |
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 bn 3 0786 0 0 0 151 0 2 125 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 bs 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ca 0 0 0 0 0 0 5 75 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 cs 1 0 0246 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 11 0 0 0 0 0 da 2 0... | https://arxiv.org/abs/2505.22118v1 |
2 0 0 5 0 81 7 4 1 0 1 01,154 0 0 0 0 2 4 0 3 9 0 0 1 0 0 0 0 kk 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 ko 0 0 0 0 0 0 4 0 0 0 0 0 0 0410 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 mk 0 1 0 0 0 0 1 0 0 0 3 0 0 0 01,137 0 0 0 0 0 0 29 0 0 0 0 0 0 0 ml 0 0 0 0 0 0 60 0 0 11 0 0 0 0 0 0482 0 0 0 0 0 0 0 0 4 0 0 1 ... | https://arxiv.org/abs/2505.22118v1 |
0 0 sl 0 0 0 0 0 0 1 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 sr 0 0 0 0 0 0 0 1 0 0 31 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 te 3 0 1 0 0 0 14 1 0 21 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 th 0 0 0 0 2 0 6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0449 0 0 0 2 tl 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0... | https://arxiv.org/abs/2505.22118v1 |
0.3547 0.3315 0.2048 0.5779 0.4482 0.4137 0.3002 KaLM-m-instruct-v1 0.5272 0.3365 0.3162 0.1936 0.6178 0.49 0.3883 0.2903 multilingual-e5-large-instruct 0.5395 0.3318 0.3047 0.1764 0.6332 0.4998 0.4029 0.2962 SFR-Embedding-Mistral 0.5687 0.3744 0.363 0.2094 0.6562 0.4968 0.4413 0.3177 gte-Qwen2-7B-instruct 0.6301 0.438... | https://arxiv.org/abs/2505.22118v1 |
0.1664 0.0935 0.1535 0.168 0.1224 0.1119 0.0351 0.0299 0.0381 0.0385 0.0149 0.0162Retrieval - re-ranking (MRR@10) Figure 12: Test set performance difference (re-ranking - retrieval), MRR@10 , incrosslingual setting with bge-reranker-v2-m3 . 929 pairs, 850 posts, 901 claims. Top- n= 30 . 16 Linq-Embed-Mistral multilingu... | https://arxiv.org/abs/2505.22118v1 |
arXiv:2505.22120v1 [cs.CL] 28 May 2025LoKI: Low-damage Knowledge Implanting of Large Language Models Runyu Wang1, Peng Ping∗1, Zhengyu Guo2, Xiaoye Zhang3, Quan Shi1, Liting Zhou4, Tianbo Ji1 1Nantong University2South China University of Technology 3China Southern Power Grid Company Limited4Dublin City University 24303... | https://arxiv.org/abs/2505.22120v1 |
be implanted without a significant impact on the general performance that LMs obtained during pre-training. By combining these two principles—allocating updates to carefully selected FFN weights—LoKI redefines the paradigm for sustainable LLM customization, ensuring models evolve without losing their essence. In summar... | https://arxiv.org/abs/2505.22120v1 |
output distribution of transformer-based language models is gradually constructed in a bottom-up fashion [35]–[37]. However, prior studies have demonstrated that LLMs contain a substantial number of parameters whose removal has minimal impact on the model’s performance on its primary tasks [26], [28], [38], particularl... | https://arxiv.org/abs/2505.22120v1 |
scores of each FFN’s knowledge output nodes using the top 50 examples from each subject in the MMLU benchmark. We denote the recorded IG scores as IG∈RN×L×D, where Nrepresents the number of examples, Ldenotes the number of layers in the model, and Dindicates the dimensionality of the knowledge output nodes. We conducte... | https://arxiv.org/abs/2505.22120v1 |
governs the percentage of trainable nodes selected from all Wdown matrices . Given Ninference samples, we denote the integrated gradients for node i in layer lon sample tas: IG(t) l,i, i = 1, . . . , D, t = 1, . . . , N The implementation proceeds as follows: 1.Quota Allocation: Calculate the total trainable slots: T=q... | https://arxiv.org/abs/2505.22120v1 |
state-of-the-art performance in function- calling benchmarks [52], even for 8B-parameter models. Consistent with the official model, we fine-tuned Llama3.1-8B-Instruct [53] on this dataset. The details of the experiment setup can be found in Appendix E. For both tasks, to systematically evaluate CF during fine-tuning, ... | https://arxiv.org/abs/2505.22120v1 |
LoKI( q=30) achieves the highest overall accuracy ( 58.93% ), surpassing the official ToolACE baseline. Notably, LoKI( q=20) exhibits the strongest multi-turn reasoning capability with a 17.75% success rate, significantly outperforming ToolACE’s mere 7.62%. All LoKI variants reduce indicators of Irrelevance, which may ... | https://arxiv.org/abs/2505.22120v1 |
Performance of two suppression strategies on benchmarks when q=1. Model TriviaQA GSM8K HellaSwag WinoGrande HumanEval IFEval Avg(%, ↓) Llama3.1 65.77 84.46 73.85 62.98 68.29 79.76 NaN S-H( q=1) 62.32 15.16 23.54 60.93 43.29 59.73 36.73 S-L( q=1) 62.05 57.85 73.84 60.3 54.88 74.15 11.35 5.2 Effect of Layer-Balanced Stra... | https://arxiv.org/abs/2505.22120v1 |
and tool-use tasks, we demonstrate that LoKI achieves competitive task adaptation while significantly reducing catastrophic forgetting compared to full parameter fine-tuning and LoRA. By connecting insights from mechanistic interpretability research to practical fine-tuning goals, our study emphasizes the promising pot... | https://arxiv.org/abs/2505.22120v1 |
Austria, May 7-11, 2024 , 2024. [19] J. Ma et al. , “Untying the reversal curse via bidirectional language model editing,” CoRR , vol. abs/2310.10322, 2023. 10 [20] J. Fang et al. , “AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models,” The Thirteenth International Conference on Learning Representat... | https://arxiv.org/abs/2505.22120v1 |
Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA, 2025. [36] I. Tenney et al. , “BERT Rediscovers the Classical NLP Pipeline,” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019. [37] J. Wallat et al. , “BERTnesia:... | https://arxiv.org/abs/2505.22120v1 |
Reading Comprehension,” Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers , 2017. [55] K. Cobbe et al. , “Training verifiers to solve math word problems,” arXiv preprint arXiv:2110.14168 , 2021. [56] R. Zellers... | https://arxiv.org/abs/2505.22120v1 |
“P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,” CoRR , vol. abs/2110.07602, 2021. [72] Y . Leng and D. Xiong, “Towards Understanding Multi-Task Learning (Generalization) of LLMs via Detecting and Exploring Task-Specific Neurons,” Proceedings of the 31st International C... | https://arxiv.org/abs/2505.22120v1 |
using only the first 50 samples for each MMLU subset and using all MMLU samples. Note that we compare at m=7and q=10. Following the definition of similarity in the previous section, we report the node similarity obtained by two methods on Qwen2.5-0.5B-Instruct and Llama3.1-8B-Instruct in Tables 8 and 9, respectively. 1... | https://arxiv.org/abs/2505.22120v1 |
3. C Implementation of LoKI Linear We provide the PyTorch implementation code for LoKI Linear, and readers can access the complete project code on our public GitHub repository. Based on this implementation, we performed all the LoKI-based model training in the paper. Our code has been validated for use with Llama-Facto... | https://arxiv.org/abs/2505.22120v1 |
,SCIDOCS , SciFact ,TREC-COVID-v2 ,ViHealthQA , and Webis-Touche2020 . For each dataset, we evaluate on a subset of the queries (the first 250). E Experiment Setups for the ToolACE Function-Calling Dataset Model Learning Rate Batch Size Epochs LR Scheduler WarmUp Ratio LoKI( q=10) 9.0e-06 4 3 cosine 0.1 LoKI( q=20) LoK... | https://arxiv.org/abs/2505.22120v1 |
improving both the accuracy and efficiency of these attribution methods. I Broader Impacts Positive Societal Impacts. LoKI enables LLMs to be adapted to new tasks with minimal loss of pre-trained general knowledge. By preserving general capabilities, LoKI promotes the development of multi-domain models that retain vers... | https://arxiv.org/abs/2505.22120v1 |
are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. •The authors should reflect on the ... | https://arxiv.org/abs/2505.22120v1 |
and data are provided or not. •If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. •Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, descr... | https://arxiv.org/abs/2505.22120v1 |
intermediate data, and generated data, etc. •The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. •At submission time, to preserve anonymity,... | https://arxiv.org/abs/2505.22120v1 |
[Yes] Justification: All these information can be found in Appendix D and Appendix E. Guidelines: • The answer NA means that the paper does not include experiments. •The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. •The paper s... | https://arxiv.org/abs/2505.22120v1 |
for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: We do not release any models. Guidelines: • The answer NA means that the paper poses no such risks. •Released models that have a high risk for... | https://arxiv.org/abs/2505.22120v1 |
compensation (if any)? Answer: [NA] Justification: The paper does not involve crowdsourcing nor research with human subjects. Guidelines: •The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. •Including this information in the supplemental material is fine, but if the main... | https://arxiv.org/abs/2505.22120v1 |
EULER: Enhancing the Reasoning Ability of Large Language Models through Error-Induced Learning Zhuoyang Wu1*, Xinze Li1*, Zhenghao Liu1†, Yukun Yan2, Zhiyuan Liu2, Minghe Yu1, Cheng Yang3, Yu Gu1, Ge Yu1, Maosong Sun2 1School of Computer Science and Engineering, Northeastern University, China 2Department of Computer Sc... | https://arxiv.org/abs/2505.22131v1 |
2023a; Chen et al., 2022; Liu et al., 2023). However, the effectiveness of mathematical problem-solving and step-by-step reasoning heav- ily depends on the capabilities of the LLMs. Some open-source LLMs, such as Llama (Touvron et al., 2023), still leave room fo improvement (An et al., 2023; Fu et al., 2023). To enhanc... | https://arxiv.org/abs/2505.22131v1 |
to baseline models. Further analysis reveals that this improvement stems from the error expo- sure model, which helps LLMs avoid repeating the same solution errors, particularly reasoning errors. Specifically, the exposure model is trained using an error-induced learning approach, enabling it to gen- erate erroneous so... | https://arxiv.org/abs/2505.22131v1 |
for the problem is 20%. Error Example 3 ：Finally... Supervised fine -tuning(B) Error -Enhanced SFT𝐌𝐚𝐱𝑷(𝒚−>𝒚+|Question ) Superior LLMSampling Figure 2: The Framework of Our EULER Model. EULER uses the error exposure model to synthesize educational solution errors. These synthesized solution errors not only enhance... | https://arxiv.org/abs/2505.22131v1 |
SFT method, EULER introduces the error-enhanced SFT method, which incorporates erroneous solution examples gener- ated by the LLM itself during the fine-tuning pro- cess, to make the LLM learn from the mistakes. Learning from errors has demonstrated its ef- fectiveness in synthesizing high-quality data which has an err... | https://arxiv.org/abs/2505.22131v1 |
whileMerrorrefers to the model being optimized. In DPO training, the loss function is regularized using the reference model ( M). Specifically, this objective considers relative probabilities by com- paring the outputs of Merrorto those of the ref- erence model M, thereby mitigating the risk of overfitting to specific ... | https://arxiv.org/abs/2505.22131v1 |
gradient accumulation step of 8, and each model is trained for 5 epochs. We also use LoRA (Hu et al., 2022) for efficient training. Additionally, our EULER model is built on TRL1and LLaMA Efficient Tuning2. More experimental details of prompt templates are provided in Appendix A.4. 5 Evaluation Result In this section, ... | https://arxiv.org/abs/2505.22131v1 |
SFT (Superior) on both Llama3-8b and Qwen2.5-7b models, showing its effectiveness in enhancing the mathematical rea- soning capability of LLMs of different scales. 5.2 Ablation Study The ablation study is conducted to evaluate the effectiveness of different strategies for synthesizing and incorporating solution errors.... | https://arxiv.org/abs/2505.22131v1 |
the Reasoning Error involves mistakes in the logical process or approach taken by the model; the Calculation Error pertains to arithmetic mistakes or errors involving incorrect numbers or values; and the Other category covers any errors that do not fit into the previous categories. The experimental details are provided... | https://arxiv.org/abs/2505.22131v1 |
The evaluation results show that EULER consis- tently produces significantly lower PPL scores than the Direct Prompt method across various datasets. This indicates that EULER is better at synthesizing more confusing solution errors, instead of some simple errors that are easy to address. Educational Value. Next, we ass... | https://arxiv.org/abs/2505.22131v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.