text string | source string |
|---|---|
processed using the same DRM-based screening and path construction as in the offline setting. The pseudo- code and prompting for unstructured retrieval are detailed in Algorithm 2 of Appendix A.1 and Ap- pendix F. By combining KG-based, document-based, and web-based retrieval, Hydra generates a rich and interpretable p... | https://arxiv.org/abs/2505.17464v1 |
pruning balances topic relevance with cross-modal agreement, producing a compact yet reliable evi- dence set for downstream reasoning3. Due to space constraints, the pseudo-code for evidence pruning is summarized in Algorithm 4 of Appendix A.2. Formally, let C={pi}N i=1as the candidate ev- idence paths, each associated... | https://arxiv.org/abs/2505.17464v1 |
prompts can be found in Appendix F. 5 Experiment In this section, we evaluate Hydra on seven bench- mark KBQA datasets. Besides the Hydra proposed in this paper, we introduce Hydra-E , which ran- domly selects one relation from each edge in the clustered question subgraph to evaluate the impact of graph structure on KG... | https://arxiv.org/abs/2505.17464v1 |
23.1 42 20 27.7 48.7 CoT (Wei et al., 2022) 38.8 62.2 30.8 42.9 20.3 28.8 48.5 SC (Wang et al., 2022) 45.4 61.1 34.4 45.3 18.9 45.4 50.3 Vanilla RAGWeb-basedGPT-3.541.2 56.8 28.9 36 26.9 62.2 46.8 Text-based 33.8 67.9 23.7 42.4 21.4 29.5 35.8 KG-centric RAGToG (Sun et al., 2024) GPT-3.5 58.9 76.2 26.3 50.2 53.6 88.0 54... | https://arxiv.org/abs/2505.17464v1 |
Document-verified Wiki-Web verified 3-source inspired LLMFigure 2: Accuracy and answer source composition by varying KG completeness on AdvHotpotQA and CWQ. issues, highlighting its adaptability. To further evaluate the performance of Hydra, we perform additional experiments, including addi- tional effectiveness evalua... | https://arxiv.org/abs/2505.17464v1 |
pathways. Journal of Machine Learn- ing Research , 24(240):1–113. Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint a... | https://arxiv.org/abs/2505.17464v1 |
In Proceedings of the ACM on Web Con- ference 2025 , pages 3505–3522. Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open founda- tion and fine-tuned chat models. arXiv preprint arXiv:230... | https://arxiv.org/abs/2505.17464v1 |
. . . . . . . . . . . . . . . . . . . . 12 Algorithm 3: Evidence_Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 A.2 Evidence pruning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... | https://arxiv.org/abs/2505.17464v1 |
. . . . . . . .15 How do different knowledge sources affect the performance of Hydra? . . . . . . . . . . . . . . . . . . . . . 15 Does search depth matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 How do path refinement prompts affect ... | https://arxiv.org/abs/2505.17464v1 |
. . . . . . 19 Evidence of answer exploration sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Overlap ratio between explored paths and ground-truth paths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .19 C.4 Error Analysis . . . . . . . . . . . . .... | https://arxiv.org/abs/2505.17464v1 |
baselines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Experiment implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 E. Case study: multi-source cr... | https://arxiv.org/abs/2505.17464v1 |
. . . . . . 26 From paragraph to knowledge path prompt template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 Refined exploration prompt template . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Predict exploration prompt template . ... | https://arxiv.org/abs/2505.17464v1 |
Eoutter′← 17 IntersectMatchUpdate (Paths, Eoutter′); Eoutter←Eoutter′;D←D+ 1; 18 Return Paths; 19 Algorithm 2 : Unstructured_Retrieval Input : Select source ( Sc), topic entities (Topic (q)), question ( q), skyline indicator ( I), width ( W) Output :Summarized wiki structured paths ( Paths Wiki), summarized web structu... | https://arxiv.org/abs/2505.17464v1 |
←get_source_prior( pi); 8 supporting_sources ← 9 get_supporting_sources (pi, Ctilde); source_agreement ← 10 min(|supporting_sources |, Wmax)/Wmax; entity_alignment ←|Ent(pi)∩Eq|/ |Ent( pi)|; 11 Sver[pi]←α1·source_prior + 12 α2·source_agreement + α3·entity_alignment; for each pi∈Ctildedo 13 Cross_Score[ pi]← 14 αcross·S... | https://arxiv.org/abs/2505.17464v1 |
3 phases. The initial phase utilizes agent-selected sourcesfor fast and efficient retrieval. If no answer is found, the refined exploration phase incorporates the LLM’s inherent knowledge and leverages on- line retrieval to update and expand the relevant knowledge corpus, with LLMs generating follow- up queries to addr... | https://arxiv.org/abs/2505.17464v1 |
(d), showing which exploration phase the answer is generated from, reveal that higher depths reduce the effectiveness of both refined and predicted exploration. Hence, we set Dmax= 3for optimal balance between per- formance and efficiency. Notably, even at lower depths, Hydra maintains strong performance by ef- fective... | https://arxiv.org/abs/2505.17464v1 |
strategy avoids the naïve composition of all sources and leads to more efficient and effective reasoning. Table 5: Performance comparison of Hydra and Hydra- E with and without agentic source selector on CWQ and WebQSP datasets. Method Evaluation CWQ WebQSP Hydra w/ agentic source selector Accuracy 87.0 95 Token Input ... | https://arxiv.org/abs/2505.17464v1 |
Accuracy 62.0 52.1 Relevant Only Token Input 157,850 24,193 LLM Calls 7.9 9.6 Table 7: Average number of entities from Freebase, WikiKG, and after graph fusion and reduction for three datasets. CWQ AdvHotpotQA QALD10-en Ave. Entity Number from Freebase 2,289,881 1,329,012 2,753,230 Ave. Entity Number from WikiKG 160,76... | https://arxiv.org/abs/2505.17464v1 |
are categorized as LLM-inspired. The analysis reveals that over 95% of answers are supported by external knowl- edge supervision, confirming that Hydra primarily grounds its reasoning in verifiable sources. Fur- thermore, up to 56% of correct answers are jointly verified by at least two distinct knowledge sources. This... | https://arxiv.org/abs/2505.17464v1 |
from it. Figure 10 shows the distribution of these error types. The results indicate that more advanced LLMs generally reduce the incidence of "other hal- lucination errors", "refusal errors", and "answer generation errors", as improved reasoning capabil- ities allow the model to make better use of the re- trieved data... | https://arxiv.org/abs/2505.17464v1 |
al., 2016), AdvHot- potQA (Ye and Durrett, 2022), and QALD10-en (Usbeck et al., 2024), a single-hop dataset: Sim- ple Questions (SimpleQA) (Petrochuk and Zettle- moyer, 2018), a slot filling dataset: ZeroShot RE (Petroni et al., 2020), and an open-domain QA dataset: WebQuestions (Berant et al., 2013), to examine Hydra ... | https://arxiv.org/abs/2505.17464v1 |
1,639 3,098 AdvHotpotQA Entity/Number CC BY-SA 4.0 308 2,312 QALD10-en Entity/Number MIT 333 – Simple Questions∗Entity/Number CC BY 3.0 1,000 14,894 Zero-Shot RE Entity/Number CC BY-SA 4.0 3,724 147,909 WebQuestions Entity/Number CC-BY 4.0 2,032 3,778 E Case study: multi-source cross-verified interpretable reasoning In... | https://arxiv.org/abs/2505.17464v1 |
− − − − − {Barry Greaves} Wiki Excerpt: “Lerman next starred in David Ayer’s World War II-set film, Fury; in the film, Lerman’s character is an American soldier fighting Nazi forces. ” Wiki (converted KG path): {Logan Lerman}starred in− − − − − → {Fury } (The Wikipedia article confirms Logan Lerman’s participation in F... | https://arxiv.org/abs/2505.17464v1 |
"Mariner Moose" – mascot of – team – division – answer(team) – located in – "American League West" Split Questions split_question 1: Which team has a mascot named "Mariner Moose"? split_question 2: Which team is in the "American League West" division? Source Evidence KG Path: {Mariner Moose}sports.mascot.team− − − − − ... | https://arxiv.org/abs/2505.17464v1 |
likely completeness of the Wiki documents. Clearly state your reasoning, and then indicate your decision using these actions: action1 for KG action2 for Wiki action3 for Web Noted, combinations allowed (e.g., [action1 + action2]). In-Context Few-shot Q: {Query} Provided sources: {Provided sources} Question analysis: {Q... | https://arxiv.org/abs/2505.17464v1 |
a few split questions that you can use one by one and finally obtain the final answer, and the associated retrieved knowledge graph path, {set of entities (with id start with "m.")} -> {set of relationships} -> {set of entities(with id start with "m.")}, Please score and give me the top three lists from the candidate p... | https://arxiv.org/abs/2505.17464v1 |
arXiv:2505.17465v1 [cs.CL] 23 May 2025A Position Paper on the Automatic Generation of Machine Learning Leaderboards Roelien C. Timmer♣Stephen Wan♣Yufang Hou♢ ♠ ♣CSIRO Data61, Australia ♢IT:U Interdisciplinary Transformation University Austria, Austria ♠IBM Research, Ireland {roelien.timmer, stephen.wan}@data61.csiro.au... | https://arxiv.org/abs/2505.17465v1 |
two SciLead (¸ Sahinüç et al., 2024) leader- board entries summarising Lewis et al. (2020) and Dong et al. (2019). cessing (NLP) methodologies has seen significant developments in recent years. Indeed, there are still many open research questions as exemplified by the 2024 shared task on ALG (D’Souza et al., 2024), und... | https://arxiv.org/abs/2505.17465v1 |
data scaling.4 We can characterise the key differences in the problem definition as concerning expectations about input and output data. Specifically, we dis- cuss: (1) reliance on domain knowledge, and (2) limited scope of extraction.5 2.1 Reliance on Domain Knowledge We observe that the ALG domains can be cate- goris... | https://arxiv.org/abs/2505.17465v1 |
con- cepts are considered for establishing new tuples. We define PTT as “A closed-domain problem for ALG, in which a system is only allowed to detect leaderboard entries composed of specific, prede- fined combinations of known scientific concepts rather than forming any new combination.” In PTT variants of ALG, only pr... | https://arxiv.org/abs/2505.17465v1 |
✗ ✓ ✗ □ NLP-TDMS Hou et al. (2019) v1-v3 ✓ ✓ ✓ ✓ ✗ □ □ ✗ ✗ ✓ □ PwC-LB Kardas et al. (2020) v1-v2 ✓ ✓ ✓ ✓ ✗ □ □ ✗ ✓ ✗ ✗ SciREX Jain et al. (2020) - ✓ ✓ ✓ ✗ ✓ ∼ ∼ ✓ ✓ ✗ ✗ TDMS-Ctx Kabongo et al. (2024) v1-v6 ✓ ✓ ✓ ✓ ✗ ✓ ✗ ✓ ✗ ✓ ✓ LEGOBench Singh et al. (2024) - ✓ ✓ ✓ ✓ ✓ ✓ ✗ ✓ ✗ ✓ ✓ SciLead ¸ Sahinüç et al. (2024) - ✓ ✓ ... | https://arxiv.org/abs/2505.17465v1 |
less important as tools like Grobid (Lopez, 2009) grow in matu- rity to transform PDF files into a logical structure format, such as XML. 4 Overview of ALG Evaluation Metrics One key issue in the field has been the use of var- ious metrics for ALG evaluation, hindering result comparisons. Appendix E lists all metrics u... | https://arxiv.org/abs/2505.17465v1 |
when averaging per paper, whereas the recall would have been 34.92 if averaged globally across all tuples. Research PaperDocument Representation Tuple Candidate Generation Tuple V erification and Entity Alignment Updating Leaderboard DatabaseLeaderboard DatabaseFigure 2: ALG Unified Conceptual Framework. neglects facto... | https://arxiv.org/abs/2505.17465v1 |
is accept- able. For example, systems can identify the key scientific concepts (e.g., extracting experiment at- tributes from relevant tables (Kardas et al., 2020)) to compose the tuples. For PTT approaches, the aim is to match the predefined tuple with the source document, in order to check for an improvement in perfo... | https://arxiv.org/abs/2505.17465v1 |
updating the leaderboard database. Kardas et al. (2020) link experimental conditions to existing Papers with Code entries. Data may be normalised prior to this step (¸ Sahinüç et al., 2024), and filtered to exclude, for example, ablation studies (Kardas et al., 2020). Most prior work does notdetail this step, as the fo... | https://arxiv.org/abs/2505.17465v1 |
Leaderboard recall reflects whether leaderboards are correctly identi- fied. Paper coverage measures whether all relevant papers are linked. Result coverage assesses the proportion of extracted results, and average over- lap quantifies agreement between generated and ground truth leaderboards. When possible, results sh... | https://arxiv.org/abs/2505.17465v1 |
on current ML terminology.7.4 Feasibility of Extraction at Scale Most research papers benchmark ALG systems on dozens or hundreds of papers. However, production-grade leaderboards such as Papers with Code integrate tens of thousands of papers. Extract- ing tuples at this scale introduces feasibility chal- lenges in com... | https://arxiv.org/abs/2505.17465v1 |
Shyamal Anadkat, et al. 2023. GPT-4 technical re- port. arXiv preprint . Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Al- shamsi, Alessandro Cappelli, Ruxandra Cojocaru, Mérouane Debbah, Étienne Goffinet, Daniel Hess- low, Julien Launay, Quentin Malartic, et al. 2023. The falcon series of open language models. arXiv ... | https://arxiv.org/abs/2505.17465v1 |
Tresa M Pollock, Samantha Daly, and Samuel Carton. 2024. Toward reliable ad-hoc sci- entific information extraction: A case study on two materials datasets. arXiv preprint:2406.05348v1 . Alex Graves and Jürgen Schmidhuber. 2005. Framewise phoneme classification with bidirectional LSTM and other neural network architect... | https://arxiv.org/abs/2505.17465v1 |
sequence-to-sequence pre-training for natural language generation, translation, and com- prehension. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics , pages 7871–7880, Online. Association for Computa- tional Linguistics. Patrice Lopez. 2009. GROBID: Combining automatic bibli... | https://arxiv.org/abs/2505.17465v1 |
Amjad Almahairi, Yasmine Babaei, Niko- lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint:2307.09288v1 . Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von... | https://arxiv.org/abs/2505.17465v1 |
extracting leaderboard tuples rely on distinct document representations. The document representation defines which sec- tions of a research paper are used before ex- tracting leaderboard-related information. Doc- TAET contains text from a Document’s Title, Abstract, Experimental Setup, and Table informa- tion. DocREC c... | https://arxiv.org/abs/2505.17465v1 |
and "ROUGE-2" refer to the same metric. Similarly, the SC-DM infers entailment relation- ships between the SC document representations and dataset-metric tuples. Both models use the BERT model limited to 512 tokens (Devlin et al., 2019), although newer models with larger token capacities may improve performance. PI Gra... | https://arxiv.org/abs/2505.17465v1 |
model for TDMS-IE, Kabongo et al. 2021, in implementing the ORKG-TDM methodol- ogy, also experimented with the pre-trained SciB- ERT model (Beltagy et al., 2019), designed for scientific text, and XLNet (Yang et al., 2019), an autoregressive transformer capable of handling con- texts longer than BERT’s 512-token maximu... | https://arxiv.org/abs/2505.17465v1 |
(Lopez, 2009) for PDF files and PANDOC (MacFarlane, 2006–) to convert LaTeX files into XML TEI markup. Then, ORKG-LB ex- tracts DocTAET (Hou et al., 2019), focusing on sections likely to contain task–dataset–metric men- tions, reducing noise and enhancing generalisation. For training the inference, for each paper, posi... | https://arxiv.org/abs/2505.17465v1 |
com- plicates direct comparisons between experiments. E Definitions of Metrics In this section, we define the micro and macro ver- sions of the Precision, Recall, and F1 metrics for the ALG task. Based on our best guess, most of the existing works typically compute micro preci- sion, micro recall, and micro F1 by first... | https://arxiv.org/abs/2505.17465v1 |
- - Kabongo et al. (2021) v2 ✓ ✓ ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✗ ✓ 3,753† 11,724† 1,608† 5,060† - - Kabongo et al. (2023b) v3 ✓ ✓ ✓ ✗ ✗ ✗ ✓ ✗ ✓ ✗ ✗ 587† 9,614† 270† 4,096† - - Kabongo et al. (2023b) v4 ✓ ✓ ✓ ✗ ✗ ✗ ✓ ✗ ✓ ✗ ✓ 2,946† 9,614† 1,262† 4,096† - - Kabongo et al. (2023b) v5 ✓ ✓ ✓ ✗ ✗ ✓ ✗ ✗ ✓ ✗ ✗ 587† 9,614† 270† 4,096† - - Kabo... | https://arxiv.org/abs/2505.17465v1 |
the number of papers is unspecified.§v1–v3 are few-shot experiment datasets with document representations: v1 (DocFULL), v2 (DocREC), and v3 (DocTAET). v4–v6 are zero-shot experiment datasets with the same representations: v4 (DocFULL), v5 (DocREC), and v6 (DocTAET).#the same data source as v2, but with updated timesta... | https://arxiv.org/abs/2505.17465v1 |
19.6 25.4 31.8 30.6 31.0 NLP-TDMS-v2 SM† Hou et al. (2019) 68.6 40.3 50.8 29.6 29.1 28.1 NLP-TDMS-v2 TDMS-IE TAE# Hou et al. (2019) 50.0 23.7 32.2 20.8 20.1 19.4 NLP-TDMS-v2 TDMS-IE TAT# Hou et al. (2019) 47.9 14.2 21.9 11.3 11.3 10.7 NLP-TDMS-v2 TDMS-IE TA# Kardas et al. (2020) 65.8 58.5 61.9 56.0 55.8 54.1 NLP-TDMS-v... | https://arxiv.org/abs/2505.17465v1 |
Kabongo et al. (2023a) 10.1 76.8 17.8 14.9 86.4 22.7 ORKG-PwC-v7*ORKG-TDM XLNet ¸ Sahinüç et al. (2024) 55.1 25.8 35.1 SciLead AxCell ¸ Sahinüç et al. (2024) 40.7 39.5 40.1 SciLead TDMR-PR Llama 2+CS ¸ Sahinüç et al. (2024) 35.9 34.9 35.4 SciLead TDMR-PR Llama 2 ¸ Sahinüç et al. (2024) 58.4 52.1 55.1 SciLead TDMR-PR Mi... | https://arxiv.org/abs/2505.17465v1 |
0.48 0.89 0.62 SciREX TDMS-IE Results of Extracting ⟨Task, Dataset, Metric, Method ⟩for Open Domain Problem Framing Jain et al. (2020) 0.53 0.72 0.61 SciREX SciREX-IE Results of Extracting ⟨Task⟩for Closed Domain Problem Framing Kardas et al. (2020) 70.6 57.3 63.3 60.7 62.6 59.7 PwC-LB-v1 AxCell Kabongo et al. (2021) 9... | https://arxiv.org/abs/2505.17465v1 |
Sahinüç et al. (2024) 63.66 33.87 44.22 SciLead AxCell ¸ Sahinüç et al. (2024) 68.93 58.81 63.47 SciLead TDMR-PR Llama 2+CS ¸ Sahinüç et al. (2024) 62.60 55.03 58.57 SciLead TDMR-PR Llama 2 ¸ Sahinüç et al. (2024) 85.03 73.20 78.67 SciLead TDMR-PR Mixtral+CS ¸ Sahinüç et al. (2024) 81.68 71.26 76.12 SciLead TDMR-PR Mix... | https://arxiv.org/abs/2505.17465v1 |
TDMS-PR Llama2 7B FS TAET# Kabongo et al. (2024) 20.41 0.94 36.73 1.68 TDMS-Ctx-v1 TDMS-PR Llama2 7B FS Full# Kabongo et al. (2024) 26.38 18.70 40.18 28.49 TDMS-Ctx-v2 TDMS-PR Mistral 7B FS REC# Kabongo et al. (2024) 28.66 24.23 40.41 34.16 TDMS-Ctx-v3 TDMS-PR Mistral 7B FS TAET# Kabongo et al. (2024) 32.43 0.57 45.95 ... | https://arxiv.org/abs/2505.17465v1 |
al. (2024) 1.81 LEGOBench MS-PR‡Zephyr Beta 7B Singh et al. (2024) 13.87 LEGOBench MS-PR‡Gemini Pro Singh et al. (2024) 13.06 LEGOBench MS-PR‡GPT-4 Results of Extracting ⟨Score⟩for Hybrid Domain Problem Framing ¸ Sahinüç et al. (2024) 23.75 31.61 27.12 SciLead TDMR-PR Llama 2 ¸ Sahinüç et al. (2024) 44.62 41.75 43.13 S... | https://arxiv.org/abs/2505.17465v1 |
SLearnLLM: A Self-Learning Framework for Efficient Domain-Specific Adaptation of Large Language Models Xiang Liu †1,2, Zhaoxiang Liu †*1,2, Peng Wang3, Kohou Wang1,2, Huan Hu1,2, Kai Wang1,2, and Shiguo Lian*1,2 1Unicom Data Intelligence, China Unicom 2Data Science & Artificial Intelligence Research Institute, China Un... | https://arxiv.org/abs/2505.17470v1 |
cycle, students undergo several key steps: they receive an exercise set and answer the questions independently; they then check their answers against the reference answers; they carefully filter out the questions they answered incorrectly or did not fully understand; and finally, they focus on analyzing and learning th... | https://arxiv.org/abs/2505.17470v1 |
approach is designed to surpass the limitations of human-labeled data by generating high-quality, di- verse problem-solving data through the model itself. The authors demonstrate that self-training can enhance model performance across various reasoning tasks, 4 J. Huang et al. pushing forward the scalability and adapta... | https://arxiv.org/abs/2505.17470v1 |
answered QA pairs, much like a student reviewing their own work. By leveraging the strong logical reasoning and deep contextual under- standing of large language models (LLMs), we design a Chain of Thought (CoT) prompt to facilitate a self-check process. As illustrated in Figure 3, the self-check CoT workflow is struct... | https://arxiv.org/abs/2505.17470v1 |
domains. For the medical domain, we engaged multiple medical experts to manually craft 35,000 QA pairs, which are highly professional and domain-specific. In contrast, for the agricultural domain, we employed GPT-4o to assist in the generation process. The GPT-assisted methodology for the agricultural domain is as foll... | https://arxiv.org/abs/2505.17470v1 |
set to 3e-4, weight decay set to 0.01, max length set to 512, lora_r set to 64 and lora_alpha set to 16. Additionally, to further assess the effectiveness of fine-tuning, we conducted control experiments by fine- tuning the models with the same configuration on the entire SFT datasets. All our experiments were conducte... | https://arxiv.org/abs/2505.17470v1 |
already perform well in many specialized domains, they still fall short of the high precision needed for production use. Our SLearnLLM offers an innovative approach to enhancing model performance in specific domains Abbreviated paper title 11 while minimizing resource expenditure. Our framework mimics the efficient hu-... | https://arxiv.org/abs/2505.17470v1 |
12 J. Huang et al. 14. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalk- wyk,J.,Dai,A.M.,Hauth,A.,etal.:Gemini:afamilyofhighlycapablemultimodal models. arXiv preprint arXiv:2312.11805 (2023) 15. Tian, Y., Peng, B., Song, L., Jin, L., Yu, D., Mi, H., Yu, D.: Toward self- improvement o... | https://arxiv.org/abs/2505.17470v1 |
arXiv:2505.17471v1 [cs.CL] 23 May 2025FinRAGBench-V: A Benchmark for Multimodal RAG with Visual Citation in the Financial Domain Suifeng Zhao1, Zhuoran Jin2, Sujian Li3*, Jun Gao1* 1Key Laboratory of High Confidence Software Technologies, School of Computer Sciences, Peking University 2School of Artificial Intelligence... | https://arxiv.org/abs/2505.17471v1 |
featuring grounded visual citation. This benchmark effectively integrates multimodal data and provides visual citations to ensure traceability, as shown in Figure 1 (b). Specifically, we construct a large- scale retrieval corpus from diverse real-world fi- nancial sources, comprising 60,780 Chinese pages from 1,104 doc... | https://arxiv.org/abs/2505.17471v1 |
evaluation method for vi- sual citation . Specifically, we define the evalua- tion metrics, precision and recall, at both the page- level and block-level, and introduce two evaluation strategies: box-bounding and image-cropping. We conduct extensive experiments and eval- uations on FinRAGBench-V . For retrieval, we con... | https://arxiv.org/abs/2505.17471v1 |
(Huang et al., 2025) focuses solely on chart-based questions, lack- ing integration with textual and tabular information. Cho et al. (2024) and Suri et al. (2024) utilize real- world PDFs but support only limited question types. MME-Finance (Gan et al., 2024) provides diverse financial questions, yet its reliance on is... | https://arxiv.org/abs/2505.17471v1 |
did the amount of mitigation -related financing using this instrument increase from 2021 to 2022 ? [1][2] QueryGeneration & Citation Answer : Based on the given information, the most -used instrument for both mitigation and adaptation for projects is Non-concessional loans . For the mitigation category, the amount in U... | https://arxiv.org/abs/2505.17471v1 |
examples provided as prompts (provided in Appendix A). In terms of data scope, the dataset includes both single-page and multi-page questions; Regarding data format, it covers text, charts, and tables; As for answers, it contains both short and long ones; Considering the characteristics of financial domain, we further ... | https://arxiv.org/abs/2505.17471v1 |
metrics. 5.2 Generation with Visual Citation During the generation stage, based on the retrieval result, the generator model Mis tasked with pro- ducing textual answer aaccompanied by visual ci- tations C, given the query q. To enable the simulta- neous generation of both answers and citations, we follow the visual cit... | https://arxiv.org/abs/2505.17471v1 |
automatic citation evaluation methods is demonstrated in Section 7.3. Citation Metrics. Inspired by Gao et al. (2023a), we evaluate both page-level and block-level cita- tions using the following two metrics: Recall evaluates whether the cited images are sufficient to support the answer. If the union of the citation se... | https://arxiv.org/abs/2505.17471v1 |
The best results are highlighted in bold . ModelChinese English ROUGE Acc P_Rec P_Prec B_Rec B_Prec ROUGE Acc P_Rec P_Prec B_Rec B_Prec Proprietary MLLMs o4-mini 38.55 58.13 78.01 75.77 54.74 48.20 40.21 69.20 75.32 75.32 60.11 55.75 GPT-4o 26.82 33.26 92.15 87.27 61.01 52.80 24.66 43.41 89.98 81.81 54.17 44.66 GPT-4V ... | https://arxiv.org/abs/2505.17471v1 |
varying strengths on Chi- nese and English datasets. Concretely, models such as GPT-4o, GPT-4V , Gemini-2.0-Flash, and 7 Eval Approach Eval ModelConsistency with IoU Consistency with Human Eval Pearson Spearman Kendall Pearson Spearman Kendall image-croppingGPT-4o 65.06 63.08 54.58 68.01 64.03 57.37 GPT-4v 63.27 61.49 ... | https://arxiv.org/abs/2505.17471v1 |
underperforms due to noise introduced by redundant visual content. Accordingly, we adopt GPT-4o with image-cropping in our experiments. 8 Conclusion In this paper, we introduce FinRAGBench-V , a benchmark designed for multimodal RAG with vi- sual citations in the financial domain, covering a 3https://github.com/HumanSi... | https://arxiv.org/abs/2505.17471v1 |
model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning Research , pages 3929–3938. PMLR. Rujun Han, Yuhao Zhang, Peng Qi, Yumo Xu, Jenyuan Wang, Lan Liu, William Yang Wang, Bonan Min, and Vitt... | https://arxiv.org/abs/2505.17471v1 |
338– 354. Association for Computational Linguistics. Aviv Slobodkin, Eran Hirsch, Arie Cattan, Tal Schuster, and Ido Dagan. 2024. Attribute first, then gener- ate: Locally-attributable grounded text generation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap... | https://arxiv.org/abs/2505.17471v1 |
by specific about which document you are referring to and describe the required text, table, or chart content without explicitly mentioning the figure or table number. ###Your questions can target the text, tables, charts, or any other elements in the image. ###Design three different queries for each document, ensuring... | https://arxiv.org/abs/2505.17471v1 |
it doesn’t cover. Here are my images: (Image 1 in Base64) (Image 2 in Base64) . . . Table 11: Prompt for Block-Level Citation Evaluation using Box-Bounding 12 Answer: {answer} Below are some extracts from the images, please decide if they cover the answers given, your answer can only be ’yes’ if it covers or ’no’ if it... | https://arxiv.org/abs/2505.17471v1 |
Text Inference Question Figure 9: An example of finance-related academic paper 15 Query: From the document ’Independent auditors’ report to the members of Craneware plc’, what is the significance of revenue recognition as a key audit matter in the context of the Group’s financial state- ment? Category: Text Inference A... | https://arxiv.org/abs/2505.17471v1 |
increase in Salaries and short-term employee benefits for Executive Directors from 2022 to 2023? Category: Table-Numerical Calculations Answer: An increase of approximately 84.94%. Reference Image: Table 21: QA Dataset Example 9: An Example of Table-Numerical Calculations Question 23 Query: According to the Q3 2024 Fin... | https://arxiv.org/abs/2505.17471v1 |
demonstrates two citation evaluation methods: box-bounding and image-cropping. E Case Study In this section, we provide several error cases based on both the different stages in the RGenCite base- line and the typical task types in finance. E.1 Error Case Study Based on Different Stages in RGenCite To illustrate the po... | https://arxiv.org/abs/2505.17471v1 |
aggregate the information correctly. As a result, it misreport- ing the growth rate of Brazil and the U.S., leading to an inaccurate comparison. This case highlights the difficulty MLLMs face in maintaining contex- tual continuity across paginated tables, a common format in financial documents. F Resource Usage Through... | https://arxiv.org/abs/2505.17471v1 |
arXiv:2505.17473v2 [cs.CV] 27 May 2025OrionBench: A Benchmark for Chart and Human-Recognizable Object Detection in Infographics Jiangning Zhu1Yuxing Zhou1Zheng Wang1Juntao Yao1Yima Gu1 Yuhui Yuan2Shixia Liu1∗ 1School of Software, BNRist, Tsinghua University2Microsoft Research Asia {zjn23,zhouyx23,wangz24,yaojt24,gu-ym2... | https://arxiv.org/abs/2505.17473v2 |
in diverse styles, spanning from realistic depictions to abstract representations of real-world objects. Second, the visual interplay between charts and HROs often results in ambiguous boundaries, making it difficult to distinguish one element from another in context. To effectively handle the highly varied infographic ... | https://arxiv.org/abs/2505.17473v2 |
/C73/C8 /C51/C9/C10 /C55/C54/C10 /C51/C9 /C67 /C64/C15/C7 /C64/C15/C12 /C55/C15/C80 /C85/C73/C54 /C88/C87/C63 /C91/C76 /C56 /C94/C73/C87 /C99/C45/C109 /C42/C41/C31 /C30/C106/C37/C29/C100/C39/C107/C42/C35/C35/C33/C100/C43/C29/C42/C45 /C99/C45/C109 /C42/C41/C31 /C30/C106/C37/C29/C100/C39/C26/C45/C45/C42/C43/C30/C43/C29/C... | https://arxiv.org/abs/2505.17473v2 |
training and evaluation. Informed by this finding, we collect infographics from two sources to balance authenticity, diversity, and scalability: 1) real infographics from online platforms, and 2) synthetic infographics programmatically created from design templates. Real infographic collection Keyword-based searching is... | https://arxiv.org/abs/2505.17473v2 |
for real ones. Programmatic synthetic infographic annotation Synthetic infographic annotations are program- matically generated with a parser integrated into the infographic generation process. This parser extracts bounding boxes for texts, charts, and HROs from the corresponding SVG file, which encodes the visual and s... | https://arxiv.org/abs/2505.17473v2 |
can gain a better understanding of the processed images over the original ones before responding to the complex user request. Considering that chart understanding essentially requires more complex, fine-grained visual reasoning over the elements within infographic images, we con- struct a Thinking-with-Boxes scheme to e... | https://arxiv.org/abs/2505.17473v2 |
and o4-mini [ 35]. For each VLM, we compare our method against three widely used baseline prompting methods: 1) Direct prompting with the chart image and the question, 2) Chain-of-Thought [37] (CoT), which prompts the model to reason step-by-step for the provided image and question, and 3) Program-of-Thought [38] (PoT)... | https://arxiv.org/abs/2505.17473v2 |
/C128/C111/C130 /C117 /C108/C101/C130 /C99/C121/C88/C101/C130 /C88/C128/C114/C117 /C108/C130 /C85/C98/C125/C123/C97/C130 /C96/C94/C92/C91/C130 /C119/C119/C119 /C130 /C93/C119 /C130 /C124 /C111 /C128/C88/C130 /C117 /C108/C101/C130 /C123/C121/C106/C101/C123/C130 /C119/C119/C119 /C130 /C114/C101/C121/C111/C130 /C117 /C108... | https://arxiv.org/abs/2505.17473v2 |
into two categories: foundation models that support zero-/few-shot detection and traditional deep learning models that require fine-tuning before detecting novel classes. We select 7 Table 3: Evaluation results of the foundation and the traditional models . The best one is bold . (a) Zero-shot prompting ModelAverage Pre... | https://arxiv.org/abs/2505.17473v2 |
OrionBench with Co-DETR. Bounding boxes in colors are the predictions for charts and HROs . 8 models like DINO-X fail to interpret these concepts through textual prompts, often missing key components. Contrary to prior findings [ 47], providing annotated example infographics does not lead to notable performance improvem... | https://arxiv.org/abs/2505.17473v2 |
Analysis As shown in Table 4, pre-training on OrionBench improves model performance when fine-tuned on Rico and DocGenome, demonstrating the effectiveness of OrionBench in enhancing graphic layout detection. With the growing interest in integrating multiple datasets for training foundation models [ 54], OrionBench serve... | https://arxiv.org/abs/2505.17473v2 |
Y . Wu, X. Shu, J. Wu, S. Fu, W. Cui, and Y . Wu, “Visimages: A fine-grained expert- annotated visualization dataset,” IEEE Transactions on Visualization and Computer Graphics , vol. 29, no. 7, pp. 3298–3311, 2023. [11] S. E. Kahou, A. Atkinson, V . Michalski, Ákos Kádár, A. Trischler, and Y . Bengio, “FigureQA: An anno... | https://arxiv.org/abs/2505.17473v2 |
pp. 917–926, 2020. [26] C. Wohlin, “Guidelines for snowballing in systematic literature studies and a replication in soft- ware engineering,” in Proceedings of the international conference on evaluation and assessment in software engineering , 2014, pp. 1–10. [27] “Statista,” 2025, accessed: 2025-02-04. [Online]. Avail... | https://arxiv.org/abs/2505.17473v2 |
Joulin, P. Krähenbühl, and I. Misra, “Detecting twenty-thousand classes using image-level supervision,” in Proceedings of European conference on computer vision , 2022, pp. 350–368. [42] S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su et al. , “Grounding dino: Marrying dino with groun... | https://arxiv.org/abs/2505.17473v2 |
S. Liu, “Foundation models meet visualizations: Challenges and opportunities,” Computational Visual Media , vol. 10, no. 3, pp. 399–424, 2024. 13 Supplemental Material for OrionBench: A Benchmark for Chart and Human-Recognizable Object Detection in Infographics A Online Platforms for Real Infographic Collection We coll... | https://arxiv.org/abs/2505.17473v2 |
Gauge chart, 30) Layered area chart, 31) Horizontal grouped bar chart, 32) Circular bar chart, 33) Multiple semi-donut chart, 34) Multiple pie chart, 35) Grouped scatterplot, 36) Multiple spline graph, 37) Multiple gauge chart, 38) Bubble chart, 39) Multiple step line graph, 40) Small multiple area chart, 41) Scatter p... | https://arxiv.org/abs/2505.17473v2 |
of box ID 1, ID 2, ..., ID N, (your finding based on the boxes). 4 Below is the image of original infographic chart, followed by your task: You are given a factoid question that you need to answer based on the provided image. You need to think step-by-step, but your final answer should be a single word, number, or phrase... | https://arxiv.org/abs/2505.17473v2 |
B, or K are required, use the exact notation shown in the chart. If there are multiple final answers, put them in brackets using this format [’Answer1’, ’Answer2’]. Remember to return a python code only without any additional text. Question: What proportion of Chinese public respondents have a neutral impression of Japa... | https://arxiv.org/abs/2505.17473v2 |
to the official settings, as well as the computational costs, in terms of GPU hours using NVIDIA GeForce RTX 4090 D. For few-shot fine-tuning, we adjust the number of training epochs inversely with the number of random infographics, ensuring consistent computational costs. All other fine-tuning hyperparameters remain unch... | https://arxiv.org/abs/2505.17473v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.