text string | source string |
|---|---|
the given segment lengths, we determine the coordinates of points D,E,F, andG. Points: - Dis on ABat(4,0). -Eis on ABat(20,0). -Fis on ACat p 7,q 7 . -Gis on ACat 5p 7,5q 7 . Reflecting Dthrough Fgives M, and reflecting Gthrough E gives N. Using the midpoint formula, we find: - Mhas coordinates 2p 7−4,12 . -Nhas coo... | https://arxiv.org/abs/2505.18237v1 |
variables and use ratios? Alternatively, since the area of quadrilateral DEGF is given as 288, maybe I can find relations based on that? Hmm, but first, perhaps assigning coordinates would be better. Let me fix coordinate system: Let ´s set point A at (0,0). Let ´s place side AB along the x-axis, so point B will be at ... | https://arxiv.org/abs/2505.18237v1 |
degenerate. Hmm, so 40 instead of that, let me just let C have coordinates (0, 91)? Wait, no, that might not work either. Alternatively, perhaps it’s better to assign coordinates such that calculations become easier. Let me think. Alternatively, set coordinates so that AC is along the y-axis. Wait maybe: Let me choose ... | https://arxiv.org/abs/2505.18237v1 |
0 + 0 + 52 = 52 Area = |Sum1 - Sum2| / 2 = |1300 -52| /2 = 1248 / 2 = 624 But in the problem statement, quadrilateral DEGF has area 288. That’s a problem, since in my coordinate system, it’s 624. So either this coordinate system isn’t valid, or perhaps I made a mistake. Hmm, so my assumption placing C at (0,91) leads t... | https://arxiv.org/abs/2505.18237v1 |
D = (4, 0) E = 41 (20, 0) G = (5 cx/7,5cy/7 ) F = ( cx/7,cy/7 ) To compute the area of quadrilateral DEGF. Let’s list the coordinates in order for the quadrilateral. It should be D-E-G-F-D, right? Yes: D to E to G to F to D. So applying shoelace formula: List of vertices: 1. D: (4,0) 2. E: (20,0) 3. G: (5cx/7,5cy/7 ) 4... | https://arxiv.org/abs/2505.18237v1 |
y-coordinate of point C is 42. That’s good. The coordinates of point C are ( cx,42). The value of cxmight not matter for some areas, especially if they can be expressed using coordinates with cxterms, but perhaps since we need to find the area of heptagon AFNBCEM, which includes points related to M and N whose coordina... | https://arxiv.org/abs/2505.18237v1 |
perhaps drawing this is complicated, but maybe area can be calculated by adding areas of various components? Alternatively, perhaps express everything in terms of cx, and see that when you compute the shoelace formula, the variables cancel out. Let me attempt the shoelace formula for the heptagon. Order of points: A = ... | https://arxiv.org/abs/2505.18237v1 |
arXiv:2505.18240v1 [cs.CL] 23 May 2025Taming LLMs with Negative Samples: A Reference-Free Framework to Evaluate Presentation Content with Actionable Feedback Ananth Muppidi1,2,+, Tarak Das1,3,+, Sambaran Bandyopadhyay1, Tripti Shukla1, Dharun D A1,3,+ 1Adobe Research,2IIIT Hyderabad,3IIT Madras Correspondence: sambaran... | https://arxiv.org/abs/2505.18240v1 |
than 50% of cases on average, with worse performance in the reference- free evaluation. More inconsistencies on direct LLM evaluators are found in other downstream tasks such as story summarization (Subbiah et al., 2024) and coherence (Mu and Li, 2024). Specifically, the task of automatic evaluation with actionable fee... | https://arxiv.org/abs/2505.18240v1 |
only have positive samples (i.e., human generated presenta- tions) while our task is to evaluate the quality of presentations of different qualities. Further, a pre- sentation Pi= (Si,1, Si,2,···, Si,mi)is a sequence (ordered) of mislides. A slide contains a slide title and optionally text and images. Number of slides ... | https://arxiv.org/abs/2505.18240v1 |
list of metrics is not exhaustive, and one can include new metrics in our framework with some additional work. 4.1 Coverage This metric is applicable when the source document is given along with the generated presentation+. We define coverage as a measure of how well cru- cial information from the source document is ca... | https://arxiv.org/abs/2505.18240v1 |
2024), but our work is the first to evaluate the flow (or narrative) of a presentation. 3 5 Detailed Pipeline of REFLEX 5.1 Content Extraction and Summarization First, we explain the extraction of multimodal con- tent from a presentation and the source document if present, and the preprocessing of them. 5.1.1 For Prese... | https://arxiv.org/abs/2505.18240v1 |
a good +https://developer.adobe.com/ document-services/apis/pdf-extract 4 quality presentation (positive sample) to generate negative presentations along with the correspond- ing scores and explanations to complete the training data. In this work, we consider perturbations up to degree 4 and the ground truth scores is ... | https://arxiv.org/abs/2505.18240v1 |
redundant information”. Text-Image Alignment : Perturbations for the metric m=Text-Image Alignment are created by interchanging the images within the slides of the presentation. Based on the degree dof perturbation hm d,x%of slides (where, x=d×20) with images are selected at random. Then, we create a random permutation... | https://arxiv.org/abs/2505.18240v1 |
set ˆTm. We use standard CrossEn- tropy loss over the tokens generated by the model and the tokens in the pseudo ground truth explana- tions. We use LoRA (Hu et al., 2021) to fine-tune our base LLM. This choice was motivated by the fact that since there are multiple metrics, we will create multiple explanation modules.... | https://arxiv.org/abs/2505.18240v1 |
overall, while also maintaining sensitivity to detecting smaller degrees of perturbations, where other baselines show a performance decline. slides within the presentation. Similarly for re- dundancy, we calculate the average sentence em- beddings of any two slides within a presentation and subtract that from 1 as foll... | https://arxiv.org/abs/2505.18240v1 |
indicating that our model is sensitive to nu- ances in the input presentation that the baselines 7 ROUGE Metric Few-Shot GPT-4o Few-Shot Phi3-Mini REFLEX ROUGE-1 ↑ 0.472 0.341 0.658 ROUGE-2 ↑ 0.284 0.132 0.361 ROUGE-3 ↑ 0.131 0.087 0.214 ROUGE-L ↑ 0.411 0.314 0.623 Table 2: Comparison of ROUGE F1 scores between Few-Sho... | https://arxiv.org/abs/2505.18240v1 |
and maximum was 0.78. The distribution of the scores are shown in Figure 3. The annotation shows that REFLEX, on average scored higher than the other baselines, achieving the highest number of "5" and "4" scores. 7 Discussions In this study, we introduced REFLEX, a method to evaluate the multimodal content quality of g... | https://arxiv.org/abs/2505.18240v1 |
Can Xu, Jiahang Xu, Weijian Xu, Sonali Yadav, Fan Yang, Jianwei Yang, Ziyi Yang, Yifan Yang, Donghan Yu, Lu Yuan, Chengruidong Zhang, Cyril Zhang, Jian- wen Zhang, Li Lyna Zhang, Yi Zhang, Yue Zhang, Yunan Zhang, and Xiren Zhou. 2024. Phi-3 technical report: A highly capable language model locally on your phone. Prepri... | https://arxiv.org/abs/2505.18240v1 |
Computational Linguistics. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human align- ment. Preprint , arXiv:2303.16634. Zheheng Luo, Qianqian Xie, and Sophia Ananiadou. 2023. Chatgpt as a factual inconsistency evaluator for text summa... | https://arxiv.org/abs/2505.18240v1 |
Proceedings of the 2021 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies , pages 1405–1418, Online. As- sociation for Computational Linguistics. Simeng Sun and Ani Nenkova. 2019. The feasibility of embedding based automatic evaluation for sin- ... | https://arxiv.org/abs/2505.18240v1 |
step-2, we convert each presenta- tion Piinto a set of images of slides Pi= [Si,1, Si,2,···, Si,mi]∀i∈[N], Pi∈ P. This array is then parsed into batched of 5slide-images concatenated vertically and fed into GPT-4o using a suitable prompt to evaluate for each of the above criteria in step-2. After conducting the steps a... | https://arxiv.org/abs/2505.18240v1 |
helpful language model and your task is to judge a presentation based on whether it contains redundant infor- mation. This can include any repetitive text/images. The input will be in the form of a list of slides, where each slide will contain a textual description of both the text and the image present in the slide. <... | https://arxiv.org/abs/2505.18240v1 |
Is the explanation easy to understand? •Actionability: Does the explanation provide clear steps or suggestions for improving the slide or presenta- tion? •Relevance: Is the explanation di- rectly related to the issues identified in the slide? Score: •High if the explanation is clear, ac- tionable, and relevant. •Low if... | https://arxiv.org/abs/2505.18240v1 |
evaluations. Polo et al. (2024) estimates performance across multiple prompts, enabling ro- bust metrics like top quantiles. Lan et al. (2024) assesses critique abilities across diverse tasks us- ing annotated references for reliability. These ap- proaches highlight the need for scalable, multi- dimensional, and reliab... | https://arxiv.org/abs/2505.18240v1 |
Multi-Scale Probabilistic Generation Theory: A Hierarchical Framework for Interpreting Large Language Models Yukun Zhang The Chinese University of Hong Kong 215010026@link.cuhk.edu.cnQi Dong Fudan University 19210980065@fudan.edu.cn Abstract Large Transformer-based language models achieve remarkable performance but rem... | https://arxiv.org/abs/2505.18244v1 |
[ 11,12]. Recent work in NLP has documented striking emergent abilities in domains from mathematical reasoning [ 13] to theory of mind [ 14], but explaining their mechanistic origins remains challenging. Our approach builds on information-theoretic foundations established by the information bottleneck principle [ 15,16... | https://arxiv.org/abs/2505.18244v1 |
upper layers [ 27], a finding corroborated across architectures [ 28]. Tenney et al. further showed that contextual representations mirror a linguistic pipeline—from part-of-speech tagging to semantic role labeling [ 29]—and Hewitt and Manning proposed that Transformers implicitly perform tree-like computations akin to... | https://arxiv.org/abs/2505.18244v1 |
factorization: P(X|C)≈P(G|C)P(I|G, C)nY t=1P xt|Lt, I, G, C , (3.2) where •P(G|C)captures the global context generated from C, •P(I|G, C)encodes intermediate semantic coherence conditioned on both GandC, and •P(xt|Lt, I, G, C )describes local token-level decisions Ltthat, together with I,G, andC, govern the generatio... | https://arxiv.org/abs/2505.18244v1 |
parsing); intermediate probes , which test inter-sentential or paragraph-level skills (e.g., next-sentence prediction); and global probes , which evaluate discourse-level understanding (e.g., topic classification, genre identification). Theoretical expectations are as follows: local layers should peak on local tasks, i... | https://arxiv.org/abs/2505.18244v1 |
information) that flag structural change points, and (iii) layer-wise functional probes (POS for local, next-sentence for intermediate, topic for global)—into a single score,which consistently yields crisp local, intermediate, and global partitions across all models. Intervention Framework To test causal predictions we... | https://arxiv.org/abs/2505.18244v1 |
8 →9. Mutual-information matrices corroborate these clusters; in BERT the three blocks {0–4, 5–8, 9–12} each maintain within-scale MI above 8.3, 9.4 and 10.0 respectively. Robustness checks confirm that these boundaries persist across input length and noise, indicating that they arise from the model’s intrinsic archite... | https://arxiv.org/abs/2505.18244v1 |
become robust; it complements fine-grained circuit analysis by shifting the lens from neurons and heads to whole-layer collaboration, with our perturbation study confirming that macro-scale shifts interact with diverse micro-circuits; and it deepens classic scaling laws, explaining super-linear gains not by parameter c... | https://arxiv.org/abs/2505.18244v1 |
of the Association for Computational Linguistics (pp. 4166–4176). [9]Belinkov, Y ., & Glass, J. (2019). Analysis methods in neural language processing: A survey. Transactions of the Association for Computational Linguistics , 7, 49–72. [10] Azaria, A., & Mitchell, T. (2023). The internal state of an LLM knows when it’s... | https://arxiv.org/abs/2505.18244v1 |
Andonian, A., & Belinkov, Y . (2022). Locating and editing factual associations in GPT. InAdvances in Neural Information Processing Systems , 35, 17359–17372. [27] Jawahar, G., Sagot, B., & Seddah, D. (2019). What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Associat... | https://arxiv.org/abs/2505.18244v1 |
L. (2022). Improving language models by retrieving from trillions of tokens. In Proceedings of ICML (pp. 2206–2240). [43] Izacard, G., & Grave, E. (2021). Leveraging passage retrieval with generative models for open domain question answering. In Proceedings of the 16th EACL (pp. 874–880). [44] Bai, Y ., Kadavath, S., K... | https://arxiv.org/abs/2505.18244v1 |
and Global (9–12, >30) tiers. This monotonic increase indicates that as information flows upward, attention expands from token-level relations to broader context. BERT and RoBERTa exhibit a smoother but still clear rise, beginning at an average span of 17.3 in layers 0–4 and ending above 30 in layers 9–12, reflecting t... | https://arxiv.org/abs/2505.18244v1 |
exhibit a clear three-tier pattern: layers 0–4 excel at local tasks (F1 rising from 0.18 to 0.77), layers 5–8 peak on intermediate tasks, and layers 9–12 achieve the highest global accuracy ( >0.82). This distribution confirms that lower layers focus on lexical and syntactic features, middle layers on sentence-level re... | https://arxiv.org/abs/2505.18244v1 |
+5.84 0.211 0.016 * GPT-2 Intermediate Reduce Max Dependency Depth -11.11 -0.192 0.037 * GPT-2 Intermediate Attention V ocabulary Diversity +4.55 0.195 0.028 * GPT-2 Intermediate Attention Sentiment Score -80.09 -0.246 0.004 ** GPT-2 Local Shift Coherence Score 0.00 -0.180 0.020 * GPT-2 Local Amplify V ocabulary Divers... | https://arxiv.org/abs/2505.18244v1 |
Baseline vs. Intervention the use of small satellites with large payloads. These small satellites are designed to provide high-speed data and communication with Earth orbit, but also provide reliable, high-resolution images and data on the environment and the physical environment. The new technologies and technologies ... | https://arxiv.org/abs/2505.18244v1 |
messages, and software that allows users to decrypt messages. The US Department of Defense (DoD) recently announced that it will build a new, more secure, and B Detailed Analysis of Scale Distributions Across Architectures This section examines how Transformer architecture and pre-training objectives shape the distribu... | https://arxiv.org/abs/2505.18244v1 |
processing is a fundamental adaptation to natural language. These insights guide model selection: choose encoders for robust local analysis, decoders for coherent generation, and encoder–decoders for complex transformations. C Mathematical Proofs and Derivations of Multi-Scale Probability Generation Theory MSPGT provid... | https://arxiv.org/abs/2505.18244v1 |
Theorem 1. There exist mappings fG, fI, fLsuch that, for any ϵ >0, P(X|C)−X G,I,LP(G|C)P(I|G, C)P(L|I, G, C )P(X|L, I, G, C ) < ϵ. (C.4) Proof. LetZ= (G, I, L )be the joint latent variables. By total probability, P(X|C) =X ZP(X, Z|C) =X ZP(Z|C)P(X|Z, C). (C.5) We show that this sum can approximate the standard chain ru... | https://arxiv.org/abs/2505.18244v1 |
Theorem 3 : For any Transformer architecture A, there exists a mapping function ϕAsuch that: ϕA:A7→ {LA, IA, GA} where LA,IA, andGAare the semantic layer partitions of that architecture, satisfying Propositions 1–3. Proof : We consider three cases: 1. Decoder-only Architecture (e.g., GPT series): For autoregressive dec... | https://arxiv.org/abs/2505.18244v1 |
1 Will Large Language Models Transform Clinical Prediction? Yusuf Yildiz yusuf.yildiz@postgrad.manchester.ac.uk Faculty of Biology, Medicine and Health, School of Health Sciences, Division of Informatics, Imaging and Data Sciences, University of Manchester, Manchester, United Kingdom, Goran Nenadic gnenadic@manchester.... | https://arxiv.org/abs/2505.18246v1 |
or machine learning methods that apply a given model to the data in a single step. In contrast, LLMs use a two-step development process: pretraining on a large portion of the data, followed by fine-tuning, often on a smaller subset. Pretraining serves as general feature extraction, while fine-tuning is tailored to the ... | https://arxiv.org/abs/2505.18246v1 |
and clinical utility still need investigation. In the next section we discuss the limitations and current challenges. Table 1 summarises the current state of play and existing gaps, offering suggestions as a general pathway for developing CPMs using LLMs. 6 Challenges for LLMs in Clinical Prediction We have identified ... | https://arxiv.org/abs/2505.18246v1 |
of clinical queries or interventions. Significant work is 8 needed to develop complex, clinically relevant time-based prediction tasks. Beyond accuracy, new validation methods are essential to assess the usefulness of multi-outcome predictions in clinical settings. This emphasises the need for improved methods of time ... | https://arxiv.org/abs/2505.18246v1 |
the studies that used the LLM architecture for clinical prediction has cited any reporting guidelines. A lack of appropriate reporting increases concerns about transparency and thoroughness and reduces the degree of trust and reproducibility of the studies. The last critical regulatory challenge is determining liabilit... | https://arxiv.org/abs/2505.18246v1 |
multioutcome usage • No standard way of evaluation (i.e.inconsistent metrics) • Suboptimal data split for validation • Extended validation. Methods for multioutcome and prognostic prediction • Bias evaluation (Tests on different subgroups e.g. Race, age groups etc.) • Standard evaluation metrics on public test dataset ... | https://arxiv.org/abs/2505.18246v1 |
K, Gutierrez L, Tan TF, Ting DSW. Large language models in medicine. Nat Med. 2023 Aug;29(8):1930–40. 5. Truhn D, Eckardt JN, Ferber D, Kather JN. Large language models and multimodal foundation models for precision oncology. Npj Precis Oncol. 2024 Mar 22;8(1):1–4. 6. Wornow M, Xu Y, Thapa R, Patel B, Steinberg E, Flem... | https://arxiv.org/abs/2505.18246v1 |
METAGENBLENDED RAG: H IGHER ACCURACY FOR DOMAIN -SPECIFIC Q&A W ITHOUT FINE-TUNING Kunal Sawarkar IBM Kunal@ibm.comShivam R Solanki IBM Shivam.Raj.Solanki@ibm.comAbhilasha Mangal IBM Abhilasha.Mangal@ibm.com May 27, 2025 ABSTRACT Despite the widespread exploration of Retrieval-Augmented Generation (RAG), its deployment... | https://arxiv.org/abs/2505.18247v1 |
high term overlap. However, its reliance on exact term matching limits its ability to capture semantic relationships, making it less effective for queries with paraphrased or conceptually related terms, a common challenge in domain-specific corpora. Semantic indexing, in contrast, leverages dense or sparse vector repre... | https://arxiv.org/abs/2505.18247v1 |
section titles are limited—fine-tuning alone cannot compensate for poor retrieval quality. Moreover, these methods often require considerable computational & human resources and may struggle to generalize outside the specific distributions they were tuned on. This limits their scalability and maintainability, especiall... | https://arxiv.org/abs/2505.18247v1 |
Following ingestion, the pipeline executes a critical metadata enrichment stage . In this phase, metadata streams are conditionally extracted and systematically calibrated based on domain-specific criteria and optimized according to performance metrics described in Section 3.2. This selective approach to metadata enric... | https://arxiv.org/abs/2505.18247v1 |
the predefined threshold, an additional high-precision enrichment step employing a large instruction- tuned LLM (e.g., meta-llama/llama-3-70B-instruct) is activated. This LLM-based enrichment phase entails prompting the model to extract comprehensive, domain-specific tags including topical keywords, significant phrases... | https://arxiv.org/abs/2505.18247v1 |
retrieval framework with both existing metadata and our enriched metadata streams. Incorporating existing metadata fields—such as Medical Subject Headings (MeSH terms) —yielded a notable accuracy improvement, elevating performance to 78.8%. When further supplemented 5 APREPRINT - M AY27, 2025 Figure 3: Impact of Metada... | https://arxiv.org/abs/2505.18247v1 |
enriched metadata 79.7 Hybrid (boosted) with enriched metadata 80.3 Hybrid (boosted) with existing metadata 80.6 Hybrid (boosted) with existing + enriched metadata 82.1 NQWithout metadata 49.99 With existing metadata 59.49 Existing + enriched metadata 60.48 SQuADWithout metadata 93.30 With existing metadata 93.58 Exist... | https://arxiv.org/abs/2505.18247v1 |
significantly enhances retrieval effectiveness, emphasizing the critical role metadata quality plays in RAG systems’ downstream generative accuracy. These empirical results compellingly demonstrate that a systematic and targeted metadata enrichment strategy not only significantly boosts retrieval precision but also dir... | https://arxiv.org/abs/2505.18247v1 |
Dai, Jiawei Sun, Qianyu Guo, Meng Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997 , 2023. [3]Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333–... | https://arxiv.org/abs/2505.18247v1 |
information 6.1.1 PubMedQA Dataset We downloaded the PubMedQA dataset from huggingface. PubMedQA is a biomedical question answering dataset created to advance research in scientific and medical QA. The dataset is built from PubMed abstracts, where each instance consists of a research question (often derived from articl... | https://arxiv.org/abs/2505.18247v1 |
sample data 6.2 Detailed results 6.2.1 Retriever Results The NQ evaluation as shown in Figure 10 reveals that augmenting documents with the existing title field alone lifts retrieval accuracy from ≈50% to≈59.5%(a+9.5-point,∼19% relative gain). Adding MetaGen’s enriched KeyBERT tags delivers an additional boost, reachin... | https://arxiv.org/abs/2505.18247v1 |
arXiv:2505.18279v1 [cs.MA] 23 May 2025Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control Alireza Rezazadeh Zichao Li Ange Lou Yuying Zhao Wei Wei Yujia Bao Center for Advanced AI, Accenture {alireza.rezazadeh, zichao.li, ange.lou, yuying.d.zhao, wei.h.wei, yujia.bao}@accenture.com... | https://arxiv.org/abs/2505.18279v1 |
environment where all users share unrestricted access to all agents. Scenario 2 (top-right): An asymmetric collaborative memory setup with heterogeneous privilege levels. Scenario 3 (bottom): Dynamically changing access, where permissions are granted or revoked over time. memory of entities and relations to enable stru... | https://arxiv.org/abs/2505.18279v1 |
work on augmenting LLMs with persistent context treats memory as a flat retrieval table, but recent systems introduce hierarchical or structured representations that better support long-horizon reasoning. MemGPT leverages operating-system abstractions to emulate virtual memory and extend the effective context window of... | https://arxiv.org/abs/2505.18279v1 |
when multiple human users interact with collaborative multiple LLM agents . Our contribution is to bridge these threads by embedding formal access-control graphs and policy-conditioned read/write transformations directly into the memory substrate, enabling safe and auditable cross-user knowledge transfer. 3 Collaborati... | https://arxiv.org/abs/2505.18279v1 |
history with other agents in A(u, t); 2. Fragments created by agent aitself while serving other users; 3.Fragments created by other users with agents that ucan invoke, subject to a’s resource-based permission R(a, t). In this way, the system enforces agent- and resource-level constraints even while enabling cross-user ... | https://arxiv.org/abs/2505.18279v1 |
as query →top_k_results . The underlying LLM performs function-calling to infer the arguments and invokes these functions via the standard OpenAI interface. Memory Encoder After completing a subtask, the conversational trace are mapped into candi- date LLM-generated key–value fragments. Each fragment is annotated with ... | https://arxiv.org/abs/2505.18279v1 |
confidentiality guarantees. 5.1 Scenario 1: Fully Collaborative Memory Cross-user collaborative memory significantly reduces overhead and improves efficiency. Task We use the MultiHop-RAG dataset [ 18], which contains 609 English news articles spanning six domains (technology, entertainment, sports, science, business, ... | https://arxiv.org/abs/2505.18279v1 |
portion of redundant work. 5.3 Scenario 3: Dynamically Evolving Collaborative Memory Adaptive permissions allow real-time updates without compromising security. Task We use the SciQAG dataset [ 21], a scientific question-answering benchmark spanning fields such as biology, chemistry, and physics. Five scientific catego... | https://arxiv.org/abs/2505.18279v1 |
R1: Chemistry, R2: Energy,R3: Ceramics, R4: Paper,R5: Physics Figure 6: Scenario 3 (Dynamically Evolving Collaborative Memory). Agent and resource usage across user queries from different categories. Yellow rectangles indicate granted access, with values representing the corresponding usage counts. simulated multi-user... | https://arxiv.org/abs/2505.18279v1 |
Ferraiolo, Rick Kuhn, Arthur R Friedman, Alan J Lang, Margaret M Cogdell, Adam Schnitzer, Kenneth Sandlin, Robert Miller, Karen Scarfone, et al. Guide to attribute based access control (abac) definition and considerations (draft). NIST special publication , 800(162):1–54, 2013. [8] Edwin Hutchins. Cognition in the Wild... | https://arxiv.org/abs/2505.18279v1 |
is organized as follows. •Section § A: Discussion of the Broad Impact of Collaborative Memory •Section § B: From Query to Response – Section § B.1 : The detailed pipeline. – Section § B.2 : The read/write policies. •Section § C: Scenario 1: Fully Collaborative Memory – Section § C.1 : Dataset details, including descrip... | https://arxiv.org/abs/2505.18279v1 |
–The selected agent gathers relevant memory fragments and accessible resources. –It generates a response, denoted as yu1,a1,t(in the case of agent a1). –The response is used to update the collaborative memory. •Aggregation | Aggregator : After all eligible agents have responded, their intermediate outputs are aggregate... | https://arxiv.org/abs/2505.18279v1 |
, a shared tier size kcross = 10 . C.2.1 Access Graph Configuration All five users ( user_1 through user_5 ) have permission to invoke each of the six domain-specialist agents. Conversely, each agent is restricted to a single, domain-aligned resource. These fixed access graphs—complete user-to-agent and one-to-one agen... | https://arxiv.org/abs/2505.18279v1 |
News specialist with deep exper- tise in: •Reporting on breakthroughs across physical and life sciences •Interpreting new research findings and method- ologies •Explaining fundamental scientific concepts clearly •Exploring implications for technology, environ- ment, and society Always check relevant memories first. Whe... | https://arxiv.org/abs/2505.18279v1 |
TechCrunch? C.4 Performance Breakdown by Query Type Figure 8 breaks down performance by query type—(top row) inference, (middle) comparison, and (bottom) temporal—across three key metrics: average accuracy (left), agent utilization (center), and resource utilization (right). In each plot, solid lines show the shared-me... | https://arxiv.org/abs/2505.18279v1 |
Asia )", " users ": [ {" user_id ": " user_1 ", " role ": " market "}, {" user_id ": " user_2 ", " role ": " finance "}, {" user_id ": " user_3 ", " role ": " logistics "}, {" user_id ": " user_4 ", " role ": " decision_maker "} ] }} 20 Output ONLY valid JSON . No explanation . Avoid markdown syntax . """ Dataset examp... | https://arxiv.org/abs/2505.18279v1 |
potential profitability for skincare products that cater to these preferences . finance_agent: Query: Analyze the economic indicators and purchasing power in Germany and other identified high-demand European markets to assess the financial viability and potential profitability of entering these skincare markets. 21 Res... | https://arxiv.org/abs/2505.18279v1 |
#### Step 3: Logistical and Supply Chain Evaluation ... ### Final Recommendation: 1. **Market Entry Strategy**: - **Germany**: Serve as a central hub for distribution across Europe, leveraging its efficient infrastructure and strong regulatory support. ... 2. **Product Strategy**: ... 3. **Operational Strategy**: ... B... | https://arxiv.org/abs/2505.18279v1 |
’logistical and supply chain requirements for high-demand regions,’ there are several factors to take into account, including infrastructure efficiency, existing trade networks, regulatory environment, and geographic location. Here’s a ranking based on these considerations: 1. **Germany**: Germany is known for its high... | https://arxiv.org/abs/2505.18279v1 |
D.2.1, and a full description of the agents and resources is provided in Section D.2.2. During the experiments, low-level users—assigned to market researcher ,financial analyst and logistics and operations lead —are each responsible for a subtask decomposed from the main query. Thestrategy director , acting as the deci... | https://arxiv.org/abs/2505.18279v1 |
reports and strategic case studies to identify trends, con- sumer behaviors, and competition analysis. Always consult memory first. If needed, call tools to re- trieve new insights. finance_agent Performs financial modeling and cost projections.You are a financial planning agent. Evaluate revenue forecasts, pricing str... | https://arxiv.org/abs/2505.18279v1 |
financial projections. Shared Memory Discussion: This is leveraged in the second decision_agent ofuser_4 entry to synthesize insights from the market analysis and financial modeling when crafting a growth plan for smart home devices in privacy-regulated markets under strict resource constraints Memory ID 54214a00-8806-... | https://arxiv.org/abs/2505.18279v1 |
scientific paper along with corresponding question-answer pairs. The passages serve as the reference corpus for answering the associated questions. Table 12 reports the detailed breakdown of document counts, average lengths, and query-type distributions. Table 12: Statistics of the subset of SCIQAG dataset for the dyna... | https://arxiv.org/abs/2505.18279v1 |
timestamp ti 6:end for The user-to-agent mappings at each timestamp during the granting process are summarized in Table 13, and the mappings for the revoking process are shown in Table 14. 31 Table 13: Users to their accessible agents at timestamp from t0tot4in the access granting stage. Timestamp User Accessible Agent... | https://arxiv.org/abs/2505.18279v1 |
knowledge_base tools. Prioritize in- formation from memories and the energy sector dataset over your parametric knowledge. Do not provide re- sponses to inquiries that are unrelated to the domain of energy and fuels. Continued on next page 34 Table 16: Domain agents and their system prompts in the dynamically evolving ... | https://arxiv.org/abs/2505.18279v1 |
ple specialized agents into a single, logically structured, and detailed response to the user query. E.3 Examples To illustrate how our system stores and subsequently reuses memory entries and how the saved memory facilitates solving subsequent queries, Table 18 presents two representative “write–then–read” cases using... | https://arxiv.org/abs/2505.18279v1 |
rates, which accelerates lignin breakdown. Controlling humidity can manage color alteration in wood. Generated at Query #21 (user_5): How does relative humidity affect the change in color of irradiated wood during heat treatment? Used at Query #66 (user_2): What role does moisture play in the degradation of insulating ... | https://arxiv.org/abs/2505.18279v1 |
arXiv:2505.18283v1 [cs.CL] 23 May 2025TAGS: A Test-Time Generalist–Specialist Framework with Retrieval-Augmented Reasoning and Verification Jianghao Wu1,2, Feilong Tang1,2, Yulong Li1,3, Ming Hu1,2, Haochen Xue1,3, Shoaib Jameel4, Yutong Xie1, Imran Razzak1† 1Mohamed bin Zayed University of Artificial Intelligence2Mona... | https://arxiv.org/abs/2505.18283v1 |
diverse and complementary reasoning strategies, and ensur- ing answer reliability under distribution shift, all without relying on any parameter updates. This mo- tivates our work: designing a structured, test-time reasoning framework that unifies generalist and specialist perspectives with retrieval-augmented promptin... | https://arxiv.org/abs/2505.18283v1 |
ios (Pal et al., 2022). Recent approaches leverage large language models (LLMs) or chain-of-thought prompting to enhance reasoning (Singhal et al., 2025). MedCoT (Liu et al., 2024) explicitly inte- grates multi-step rationale generation with hierar- chical expert feedback. Concurrently, biomedical LLMs such as MedLLaMA... | https://arxiv.org/abs/2505.18283v1 |
CoT similarity. This enables alignment in both problem context and reasoning structure, thereby improving downstream multi- agent reasoning. 2.3 Multi-Agent Systems for Reasoning Multi-agent systems (MAS) have emerged as a promising approach to enhance the robustness, di- versity, and reliability of reasoning in comple... | https://arxiv.org/abs/2505.18283v1 |
Oare concatenated in standard order (A, B, C, D). Cosine similarity is computed against all candidate embeddings E(Qi⊕Oi)in the corpus. The top- Kretrieved examples form: T1= Top -K di∈ D: sim( z,E(Qi⊕Oi)) .(1) Stage 2: Rationale-guided retrieval. After Round-1 reasoning yields preliminary rationales RG,1andRS,1, we r... | https://arxiv.org/abs/2505.18283v1 |
path, UAAA miti- gates hallucination propagation and stabilizes final predictions, all without any parameter updates. 4 Experiments 4.1 Experimental Setting Retrieval Dataset. We use the MedReason dataset (Wu et al., 2025) as our external retrieval corpus. It contains 32,682 medical QA pairs with clinically validated, ... | https://arxiv.org/abs/2505.18283v1 |
+ Self-Refine 41.0 13.0 34.0 28.1 34.2 34.0 22.0 17.0 19.0 26.9 + MedAgents 43.0 15.0 30.0 27.0 28.8 8.0 19.0 3.0 6.0 20.0 + MDAgents 36.0 11.0 22.0 21.3 24.7 8.0 13.0 4.0 5.0 16.1 + MedPrompt 34.0 11.0 26.0 22.5 26.0 22.0 16.0 14.0 9.0 20.1 + Ours 54.0 13.0 32.0 33.7 45.2 47.0 17.0 22.0 22.0 31.8 Table 1: Performance ... | https://arxiv.org/abs/2505.18283v1 |
(7) OpenBioLLM-8B (Pal and Sankarasubbu, 2024): An open-source 8B biomed- ical language model fine-tuned for healthcare and life sciences applications. Evaluation Metrics Following (Tang et al., 2023), we report Pass@1 Accuracy as the evalua- tion metric, which measures whether the model’s first generated answer exactl... | https://arxiv.org/abs/2505.18283v1 |
ac- curacy from 14.1% to 23.9%. In particular, TAGS demonstrates robust performance gains on difficult benchmarks such as MedQA (+12.0 percentage points), MMLU (+21.9 percentage points), and MedXpert-U (+21.0 percentage points). Notably, our inference-only strategy even outperforms sev- eral models that have been fine-... | https://arxiv.org/abs/2505.18283v1 |
improves accuracy from 34.2% to 45.2% on MMLU and from 16.0% to 22.0% on MedXpert-U. However, adding more special- ists brings limited or no further gains, likely due to redundancy or conflicts in reasoning paths. Fig- 2 3 4 5 01240 30 20 10 0 (b) Retrieval Size (a) Specialist CountAccuracy (%) Accuracy (%)40 30 20 10 ... | https://arxiv.org/abs/2505.18283v1 |
from human evaluation or rationale consistency metrics to fur- ther assess clinical applicability. Moreover, our specialty inference component can occasionally misclassify the most relevant do- main, which, although gracefully handled, may still introduce suboptimal reasoning contexts. Finally, our evaluation is confin... | https://arxiv.org/abs/2505.18283v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.