text string | source string |
|---|---|
a standard treatment for allergic conjunctivitis. Maybe for symptom relief, but not a primary treatment. **D) Fluorometholone eye drops** Fluorometholone is a corticosteroid. Corticosteroids are anti-inflammatory and can be used for severe allergic reactions, but they’re usually not the first-line treatment for allergi... | https://arxiv.org/abs/2505.19300v1 |
about their orbital periods and the masses/radii of their host stars. The probability of a transit is related to the inclination of the planet’s orbit, and planets with shorter orbital periods around smaller stars have a higher probability of transiting. 2. **Key Factors**: - **Orbital Period**: Planet_1 has an orbital... | https://arxiv.org/abs/2505.19300v1 |
to find his place of birth is ’peo- ple.person.place_of_birth’. 2. Next, I need to retrieve the entity associated with the relation ’people.person.place_of_birth’ for JaMarcus Russell. Query: <entity>JaMarcus Russell, people.person.place_of_birth</entity> <result> [’Mobile’] </result> The result indicates that the plac... | https://arxiv.org/abs/2505.19300v1 |
is likely the key here. But to get the count of how many were played on a hard surface, I might need to look at the entire "Surface" column and count the occurrences of "hard surface." Wait, but the question is a bit vague. It’s asking "how many were played on a hard surface?" So, I need to determine what "were played"... | https://arxiv.org/abs/2505.19300v1 |
east**: After retrieving the staple, I need to move east. The command for this is ‘go east‘. 4. **Rest the staple on the shelf inside the cellar**: Finally, I need to go to the cellar, find the shelf, and place the staple there. The command to put an item on something is usually ‘put [item] on [location]‘. Let’s start ... | https://arxiv.org/abs/2505.19300v1 |
objective, after getting the staple, I need to head east. 4. **Command**: ‘go east‘ Now, I’m back in the cellar with the staple. The next instruction is to rest the staple on the shelf inside the cellar. I’ll check the description to locate the shelf. <description>14 | go west, go west, get staple, go east</description... | https://arxiv.org/abs/2505.19300v1 |
arXiv:2505.19302v1 [cs.DB] 25 May 2025Odin : A NL2SQL Recommender to Handle Schema Ambiguity Kapil Vaidya1∗, Abishek Sankararaman2, Jialin Ding2, Chuan Lei2, Xiao Qin2, Balakrishnan Narayanaswamy2, Tim Kraska2 1Parallel Web Systems2Amazon Web Services ABSTRACT NL2SQL (natural language to SQL) systems translate natural ... | https://arxiv.org/abs/2505.19302v1 |
For example, consider the question, “ What is the revenue per customer? ”. In this case, the database schema might in- clude both a customers table and an orders table, where the orders table contains a column named revenue , while the customers table may include an aggregated column, total_revenue . Depending on which... | https://arxiv.org/abs/2505.19302v1 |
selectively modifying the infor- mation provided to the LLM. Specifically, it removes certain schema elements from previously generated SQL queries, encouraging the LLM to explore different schema components and generate diverse queries. The generator may produce incorrect SQL queries by omitting too many important sch... | https://arxiv.org/abs/2505.19302v1 |
model’s context size. These methods focus on generating more accurate SQL queries for a given NL question. However, they neither explicitly handle schema ambiguity nor learn the preferences from users. Consequently, these state-of-the-art methods are not able to learn from user feedback and improve their SQL generation... | https://arxiv.org/abs/2505.19302v1 |
too many options. Maximizing accuracy alone does not guarantee a good user experience. When users are presented with an excessive number of choices, identifying the correct query becomes challenging. Therefore, it is essential to balance accuracy and the size of the result set to enhance user experience. The optimizati... | https://arxiv.org/abs/2505.19302v1 |
LLM to produce incorrect SQL queries. Additionally, some generated queries might not align with user preferences. The primary objective ofOdin is to maintain a compact set of generated SQL queries while ensuring accuracy. Removing these incorrect and misaligned queries can help Odin reduce the set size. To address this... | https://arxiv.org/abs/2505.19302v1 |
while lower temperatures produce more focused, predictable results. One might assume that high temperatures would generate diverse SQL queries, but in practice, it leads to only superficial differences, such as varying table/column aliases, using equivalent functions, or altering the SQL structure with subqueries. Whil... | https://arxiv.org/abs/2505.19302v1 |
then use the limitations of this simple algorithm to motivate the schema masking algorithm used by Odin ’s Generator. 5.1 Naive Schema Masking Algorithm AnaiveversionoftheschemamaskingalgorithmisshowninFig.1(B). It generates SQL queries by progressively restricting a given schema and exploring different masked schemas ... | https://arxiv.org/abs/2505.19302v1 |
The algo- rithm takes the user question, database schema and a limit of LLM calls as input (Line 12). If the LLM fails to find relevant entities in the masked schema and is unable to generate a valid SQL query, the exploration may terminate early. The limit on LLM calls rep- resents only the maximum number of attempts,... | https://arxiv.org/abs/2505.19302v1 |
J. Ding, C. Lei, X. Qin, B. Narayanaswamy, T. Kraska Algorithm 1 Greedy Tree Search with Resource Constraints 1:Input: 2:𝑓_𝑠𝑐ℎ- Full schema 3:𝑞- Initial query 4:𝑚𝑎𝑥_𝑐𝑎𝑙𝑙𝑠 - Maximum number of LLM calls 5:Output: 6:𝑓𝑖𝑛𝑎𝑙 _𝑞𝑢𝑒𝑟𝑖𝑒𝑠 - List of SQL queries 7:Algorithm: 8:𝑁𝐿2𝑆𝑄𝐿 - Generates SQL for... | https://arxiv.org/abs/2505.19302v1 |
subset of labels such that the true label is included with high probability. The key idea behind conformal prediction is straightforward: a scoring function is used to assess how likely a label is incorrect, and all labels below a carefully chosen threshold are selected. This threshold is determined using a calibration... | https://arxiv.org/abs/2505.19302v1 |
establishing this connection, we leverage the statistical guarantees provided by conformal pre- diction to ensure that the correct SQL query is retained with high probability after the selection process. Conformal prediction con- structs prediction sets that contain the true outcome with a specified confidence level 1−... | https://arxiv.org/abs/2505.19302v1 |
35], while another is to use the logit probabilities of specific tokens. We adopt the logit probabilities approach for finer-grained scoring. We use a prompt that asks the LLM if the SQL query answers the question, providing two options: A.Yes, and B.No (as shown in Fig.2). We use the logit probability of option Bas th... | https://arxiv.org/abs/2505.19302v1 |
is then transformed into a format that influ- ences both the Generator and Selector components of the system. Specifically, we provide these preferences in a textual format that biases the LLM’s output. The key information conveyed in the tex- tual hints is that when a user refers to a particular entity, a specific sch... | https://arxiv.org/abs/2505.19302v1 |
enhancing the precision of the generated hints. Finally, the Format_Hint function integrates the entity, the cor- rect schema mapping, and the list of incorrect mappings into a tem- plate for generating hints. After processing all entities, the function returns the list of textual hints. 7.2 Using Personalization Hints... | https://arxiv.org/abs/2505.19302v1 |
will be correct in the calibration set. While X may receive a higher score during the selector stage, the cut-off will be set such that both X and Y are included in the final selection, as either could be correct. 8 EVALUATION We begin by detailing the experimental setup (Section 8.1) and then present the results of an... | https://arxiv.org/abs/2505.19302v1 |
•Forced Diversity : In this baseline, we provide the LLM with all previously generated SQL queries and instruct it to produce a new SQL query that differs from the prior ones. The generated SQL queries are then shown to the user. •Odin : The Odin baseline comprises three modules: Generator , Selector , and Personalizer... | https://arxiv.org/abs/2505.19302v1 |
( 𝐴𝑣𝑔𝐴𝑐𝑐 ) of the generated results versus the average number of SQL results shown to the user ( 𝐴𝑣𝑔𝑅𝑒𝑠𝑢𝑙𝑡𝑆𝑖𝑧𝑒 ) for three distinct ambiguity types: join ambiguity, table ambiguity, and column ambiguity. The figure compares Odin with two baselines: Sampling and Forced Diversity. For all methods, we g... | https://arxiv.org/abs/2505.19302v1 |
accuracy to the baselines for small result sets, it rapidly improves, achieving up to a 25% gain in accuracy as the set size increases. Impact of the Selector: We assess the Selector component, which is designed to enhance precision by filtering out incorrect SQL state- ments while keeping the correct ones, thereby imp... | https://arxiv.org/abs/2505.19302v1 |
the scoring function and the alpha value. In Fig. 7, we illustrate the trade-off points achieved K. Vaidya, A. Sankararaman, J. Ding, C. Lei, X. Qin, B. Narayanaswamy, T. Kraska Figure 6: Odin ’s Generator results alongside various SQL generation baselines for ambiguity types such as Join, Table Synonyms, and Column Sy... | https://arxiv.org/abs/2505.19302v1 |
without the Selector, reflecting the trade-off that the Selector is designed to manage. 9 CONCLUSION In this paper, we introduced Odin , an NL2SQL system that handles schema ambiguity. Odin further personalizes SQL results based on user feedback. Odin demonstrates consistent accuracy improve- ments across different amb... | https://arxiv.org/abs/2505.19302v1 |
[16] Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023. Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 13067–13075. [17] Haoyang Li, Jing Zhang, Hanbing Liu, Ju Fan, Xiaokang Zhang, Jun Zhu, Renjie Wei, Hongyan P... | https://arxiv.org/abs/2505.19302v1 |
Collaborative Framework for Text-to-SQL. CoRR abs/2312.11242 (2023). [32] Bing Wang, Changyu Ren, Jian Yang, Xinnian Liang, Jiaqi Bai, Qian-Wen Zhang, Zhao Yan, and Zhoujun Li. 2023. Mac-sql: Multi-agent collaboration for text-to-sql. arXiv preprint arXiv:2312.11242 (2023). [33] Bailin Wang, Richard Shin, Xiaodong Liu,... | https://arxiv.org/abs/2505.19302v1 |
arXiv:2505.19345v1 [cs.CL] 25 May 2025PatentScore: Multi-dimensional Evaluation of LLM-Generated Patent Claims Yongmin Yoo Macquarie University yooyongmin91@gmail.comQiongkai Xu Macquarie University qiongkai.xu@mq.edu.auLongbing Cao Macquarie University longbing.cao@mq.edu.au Abstract Natural language generation (NLG) ... | https://arxiv.org/abs/2505.19345v1 |
1 defines comprehensive but precise legal and technical scopes of IP value and protection and determines critical legal boundaries of patent rights (Merges, 1997). It is essential to en- sure legal and technical accuracy when evaluating AI-generated patent claims. Errors or ambiguities in Claim 1 can lead to litigation... | https://arxiv.org/abs/2505.19345v1 |
As the legal foundation of patent protec- tion, claims demand structural, technical, and legal precision. To this end, we propose PatentScore, a multi-dimensional evaluation framework, and con- duct large-scale experiments on claims generated by GPT-4o-mini. The results show a strong correla- tion with expert judgments... | https://arxiv.org/abs/2505.19345v1 |
incorporating patent-specific evaluation dimensions into its framework. 2.3 NLG Metrics for Patent Evaluation NLG metrics dominate patent generation evalua- tion, we review them specifically. First, traditional n-gram based metrics like BLEU and ROUGE mea- sure lexical overlap between generated and refer- ence texts. H... | https://arxiv.org/abs/2505.19345v1 |
LLMs. To mitigate the inher- ent variability of GPT model outputs, we perform ten iterations of evaluation for each metric and use the mean score as the final value. To further enhance output consistency, we fix the tempera- ture parameter at 0.3. This configuration ensures more reliable and stable evaluation results. ... | https://arxiv.org/abs/2505.19345v1 |
both consistency and compre- hensive analysis. TASK: Patent Claim 1 Assessment FOCUS: Component-specific evaluation OUTPUT FORMAT: Detailed analysis and score (1-5) I. Evaluation Criteria Primary Requirements: - [Component-specific requirement 1] - [Component-specific requirement 2] Secondary Requirements: - [Additiona... | https://arxiv.org/abs/2505.19345v1 |
offers a robust and comprehensive method for evaluating patent claim quality. 4 Data and Verification Measures 4.1 Claim Data Generation and Composition We use a subset of the HUPD dataset (Suzgun et al., 2022), consisting of patents filed in 2016 and 2017. This subset includes the first claim of patents clas- sified u... | https://arxiv.org/abs/2505.19345v1 |
their limitations in patent claim evaluation. The effectiveness of the framework is further val- idated by its strong performance across multiple correlation measures, including Spearman’s rank correlation ( ρ= 0.813) and Kendall’s rank correla- tion (τ= 0.665). Beyond correlation coefficients, our approach demonstrate... | https://arxiv.org/abs/2505.19345v1 |
0.002·MBS (3)These findings highlight the minimal impact of BERTScore on overall framework performance. Removing BERTScore results in only a 0.1% de- crease, indicating that semantic metrics play a com- plementary rather than critical role in patent claim evaluation. In contrast, structural and legal met- rics demonstr... | https://arxiv.org/abs/2505.19345v1 |
participated in the evalua- tion process provided voluntary and informed con- sent before contributing to the study. Their identi- ties remain anonymous to protect their privacy and confidentiality. The study does not involve any personally iden- tifiable information or sensitive data, and it strictly adheres to ethica... | https://arxiv.org/abs/2505.19345v1 |
Property Statis- tics Database. https://www.oecd.org/sti/ intellectual-property-statistics.htm . Ac- cessed: 2024-12-09. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for ... | https://arxiv.org/abs/2505.19345v1 |
Q:Does it comply with format requirements? Tasks: • Check format compliance: –Indentation and spacing –Punctuation usage (colons, semi- colons) –Conjunction placement • Identify rule violations • Remember formatting errors 4.Standard-Based Comparative Analysis Input: Compliance assessment results Output: Comparative an... | https://arxiv.org/abs/2505.19345v1 |
Q:Do the essential punctuation marks meet requirements? Tasks: • Confirm essential punctuation: –Appropriate colon usage –Appropriate semicolon usage –Appropriate "and" usage • Verify punctuation locations: –Exact position of each mark •Remember missing or inappropriate items 3.Format Compliance Assessment Input: Punct... | https://arxiv.org/abs/2505.19345v1 |
Input: Reference verification results Output: Compliance evaluation report Q:Does reference usage comply with format requirements? Tasks: • Check reference format compliance: –Reference term consistency –Reference terminology clarity • Remember formatting errors 4.Standard-Based Comparative Analysis Input: Compliance a... | https://arxiv.org/abs/2505.19345v1 |
verification reportQ:Do the article usages meet requirements? Tasks: • Confirm proper introduction: –Proper first mention articles –Proper subsequent articles • Verify reference accuracy: –Match between references • Remember improper article usage 3.Format Compliance Assessment Input: Article verification results Outpu... | https://arxiv.org/abs/2505.19345v1 |
Output: Validity verification report Q:Do the elements contribute unique value? Tasks: • Confirm technical contribution: –Novel features –Specific improvements • Verify element differentiation: –Technical distinctions • Remember contradictions or repetitions 3.Format Compliance Assessment Input: Verification results Ou... | https://arxiv.org/abs/2505.19345v1 |
Output: Scope verification report Q:Do the elements clearly define boundaries? Tasks: • Confirm boundary clarity: –Clear technical limits –Specific definitions • Verify scope precision: –Term specificity • Remember ambiguous or broad terms 3.Format Compliance Assessment Input: Scope verification results Output: Complia... | https://arxiv.org/abs/2505.19345v1 |
Architectures of Error: A Philosophical Inquiry into AI and Human Code Generation Camilo Chac´ on Sartori∗ Artificial Intelligence Research Institute (IIIA-CSIC), Bellaterra, 08193, Barcelona, Spain. Abstract With the rise of generative AI (GenAI), Large Language Models are increas- ingly employed for code generation, ... | https://arxiv.org/abs/2505.19353v1 |
. . . . . . . . . . . . . . . . . . 26 3.2 The Attribution Problem: Rethinking “Responsibility” . . . . . . . . . 29 3.3 Trust Calibration Based on Mechanism Awareness . . . . . . . . . . . 31 3.4 Human Cognitive Adaptation and System Effects . . . . . . . . . . . . 33 4 Discussion 35 4.1 Integrating Architectural Insi... | https://arxiv.org/abs/2505.19353v1 |
evolved cognitive capacities; the other artificial, grounded in stochastic-statistical modeling. Although their outputs (functional code) may seem similar, this resemblance is largely illusory. The underlying processes differ across sev- eral key dimensions, with significant implications for how we understand, verify, ... | https://arxiv.org/abs/2505.19353v1 |
understanding these divergent “Architectures of Error” can inform both software engineering practice and philosophical inquiry into technology and AI. While many technical papers on LLM-driven code generation emerge daily, few address the philosophical and epistemic relationship between AI-generated and human-generated... | https://arxiv.org/abs/2505.19353v1 |
functional organization for processing information determines the system’s capa- bilities, irrespective of the physical substrate or appeals to subjective awareness or intrinsic semantic properties. With GenAI in code generation, we encounter a similar occurrence. Given a prompt, a model may return functioning code wit... | https://arxiv.org/abs/2505.19353v1 |
corpus of syntactically and semantically valid examples. This enables it to probabilistically inter- nalize coherent coding patterns. However, this process is fragile. For instance, if a preprocessing script replaces every instance of cout in C++ with Python’s print , the model may wrongly learn that print is standard ... | https://arxiv.org/abs/2505.19353v1 |
“success” criterion. A human can also change their notion of “success,” but it affects their well-being. What, by contrast, affects an AI model? Recent LLM-based agent systems show promise in self-correcting coding tasks through distributed cross-verification, approximating epistemic cooperation. Robeyns, Szum- mer, an... | https://arxiv.org/abs/2505.19353v1 |
benchmark models could serve as a form of distributed epistemic validation , akin to a network of cooperat- ing epistemic agents (though engineered and overseen by humans), applying practical checks for anomalous or harmful outputs (Simon, 2015). Therefore, vulnerabilities from AI and human programmers differ fundament... | https://arxiv.org/abs/2505.19353v1 |
components or legacy code. Human cognitive architecture, while capable of integrating broad, multi-modal information streams (text, diagrams, conversations, past experiences) and forming rich situational models, is nonetheless fallible due to constraints on attention, working memory, and the reconstructive nature of lo... | https://arxiv.org/abs/2505.19353v1 |
data and architectural constraints (e.g., reliance on statistical correlations, limitations in extrapolating beyond learned patterns). These limits become evident when the AI is tasked with handling genuinely novel abstractions or problem types not well- represented in its training data. Human boundaries, while also re... | https://arxiv.org/abs/2505.19353v1 |
analysis tools become essential for managing systems with non-deterministic elements (Beschastnikh, Wang, Brun, & Ernst, 2016). However, debugging GenAI outputs faces unique hurdles stemming from the inter- action between its variability and its specific limitations in semantic coherence (Dimension 1) and context handl... | https://arxiv.org/abs/2505.19353v1 |
to ensure the traceability of results in software development involv- ing GenAI, a human programmer should adopt a stance informed by Rescher’s conception of rationality (1992). Rescher argues: Rationally is to make use of one’s intelligence to figure out the best thing to do in the circumstances. Rationality is a matt... | https://arxiv.org/abs/2505.19353v1 |
the interaction between D1andD3in a GenAI system (see first row in Table 1), analyzed through distinct LoAs. A higher LoA captures global semantic 22 coherence—meaning, logic, and project-wide constraints—while a lower, architectural LoA reflects internal processing limits, such as context window size. A failure at the... | https://arxiv.org/abs/2505.19353v1 |
the distinct architectural origin and the failure modes previously analyzed (Section 2) 25 in relation to code generation has direct and practical implications for development, testing, and management of software involving GenAI. Consider the traditional approach to learning programming, described by Carlos Baquero in ... | https://arxiv.org/abs/2505.19353v1 |
re-evaluation of security assurance. This involves considering the epistemic grounds for trusting code whose vulnerabilities may stem from statistical learning rather than intentional design flaws, suggesting that standard vulnerability checks may be insufficient without targeted, architecture-aware auditing. 27 V&V mu... | https://arxiv.org/abs/2505.19353v1 |
we trust that code produced by a stochastic architecture—with its well-known limitations and propensity for error—can lead to genuinely novel and practically useful solutions? 3.2 The Attribution Problem: Rethinking “Responsibility” Discussions of “responsibility” concerning AI often focus on the ethical and social imp... | https://arxiv.org/abs/2505.19353v1 |
used, outputs obtained, model versions selected, and parameters chosen, providing at least a partial audit trail for any inte- gration of cognitive and artificial architectures. If this is not addressed, could the human–AI symbiosis in code generation entail that programmers bear a new “duty of care” when using AI tool... | https://arxiv.org/abs/2505.19353v1 |
the one hand, GenAI tools can offer positive effects, such as enabling programmers to work more efficiently, potentially accelerating learning processes, and increasing motivation by alleviating tedious and repetitive tasks. There are even indications of shifts in teamwork dynamics, with software engineers turning to G... | https://arxiv.org/abs/2505.19353v1 |
defining the criticality and potential impact of the AI-generated code within the larger software system. Thus, responsibility shifts from the non-comprehending AI to the human-led governance and integration framework overseeing its output. 4 Discussion This section situates the “Architecture of Error” framework within... | https://arxiv.org/abs/2505.19353v1 |
understanding of these intrinsic characteristics is a necessary step before developing effective explainability or V&V strategies (Section 3.1). Understanding the “Architecture of Error” in inherently stochastic systems is, in my view, an epistemic prerequisite for assessing which types of explanations are feasible or ... | https://arxiv.org/abs/2505.19353v1 |
Floridi’s. This allows technological advancements (which can be seen as shifts in LoAs or capabilities within them) to be positioned and analyzed within these four fundamental dimensions, as briefly explored in the discussion on dimensional interactions (Section 2.5). 4.Are four dimensions sufficient for this analysis?... | https://arxiv.org/abs/2505.19353v1 |
its analyses are not definitive but rather a spur to further thought, grounded in the code itself and the human and artificial processes that create it. Ultimately, understanding these distinct “Architectures of Error” is a pragmatic necessity for a more reliable, responsible, and effective future in software engineeri... | https://arxiv.org/abs/2505.19353v1 |
initial set of all tasks and their base properties. * ‘k_target_size‘, ‘k_flexibility‘: Define the desired size and flexibility of WorkPackages. * ‘global_constraints‘: Constraints that each WorkPackage *must* satisfy (sum of IDs, sum of costs, sum of values). * ‘category_affinity_rules‘: Complex rules about how tasks ... | https://arxiv.org/abs/2505.19353v1 |
are retired from the "active" pool. * If ‘success‘ is ‘False‘, the ‘task_ids‘ must return to the "available for assignment" pool (possibly with an urgency penalty or a cooldown to prevent rapid, futile retries). Their "dependency fulfilled" status does not change for other tasks. update_task_urgency_model(self, task_id... | https://arxiv.org/abs/2505.19353v1 |
Failures, and the Fallacy of Generative AI. Harvard Data Science Review (Special Issue 5), , https://doi.org/10.1162/99608f92.ad8ebbd4 (https://hdsr.mitpress.mit.edu/pub/1yo82mqa) 48 Beschastnikh, I., Wang, P., Brun, Y., Ernst, M.D. (2016, March). Debugging dis- tributed systems: Challenges and options for validation a... | https://arxiv.org/abs/2505.19353v1 |
Rescher, N. (2003). Epistemology: An introduction to the theory of knowledge . State University of New York Press. Rescher, N. (2017). Value reasoning: On the pragmatic rationality of evaluation . Springer International Publishing. Robeyns, M., Szummer, M., Aitchison, L. (2025). A self-improving coding agent. Retrieved... | https://arxiv.org/abs/2505.19353v1 |
arXiv:2505.19354v1 [cs.CL] 25 May 2025GC-KBVQA: A New Four-Stage Framework for Enhancing Knowledge Based Visual Question Answering Performance Mohammad Mahdi Moradi1, Sudhir Mudur1, 1Concordia University, Correspondence: mudur@cs.concordia.ca Abstract Knowledge-Based Visual Question Answering (KB-VQA) methods focus on ... | https://arxiv.org/abs/2505.19354v1 |
pre-training on di- verse text datasets, inherently acquire a substantial amount of factual knowledge (Wang et al., 2023), enabling them to answer a wide variety of gen- eral knowledge questions without the need for task- specific retraining. In the context of VQA, LLMs can process textual information extracted from bo... | https://arxiv.org/abs/2505.19354v1 |
be broadly categorized into three groups(Kafle and Kanan, 2017): (i) joint embed- ding, (ii) attention mechanisms, and (iii) external knowledge integration, also known as KB-VQA. Joint embedding methods (Antol et al., 2015; Fukui et al., 2016; Ma et al., 2016) map visual and textual inputs into a common latent space fo... | https://arxiv.org/abs/2505.19354v1 |
regions. Ad- ditionally, the patch-based division may oversim- plify complex relationships by focusing on local areas and ignoring global context, which is cru- cial for questions involving object interactions and overall scene understanding. Providing LLMs with relevant, context-rich prompts from images and external k... | https://arxiv.org/abs/2505.19354v1 |
treesQuestion Content Extraction Question Relevant Region ExtractionC1,C2,C3,C4,C5,... all-MiniLM-L6-v2 LLMQuestion Based Caption RankingC1:The wooden fence serves as a barrier between the giraf fe and the trees. C2:In the image, there is a giraf fe standing in a grassy field next to a wooden fence. C3:The type of fenc... | https://arxiv.org/abs/2505.19354v1 |
the question. Ques- tions often include verbose or extraneous content that obscures their core intent, potentially mislead- ing the filtering process. This can result in retaining less relevant captions or discarding highly perti- nent ones, ultimately compromising the accuracy of the answer prediction. To address this... | https://arxiv.org/abs/2505.19354v1 |
into a unified, comprehensive prompt along with instructions. This prompt is then provided to the Llama-3-8B- Instruct model. The prompt follows a structured format, carefully designed to ensure clarity and provide the model with all necessary context and relevant auxiliary information: " Infer an answer for the follow... | https://arxiv.org/abs/2505.19354v1 |
outperforms notable few-shot baselines such as PICa and achieves performance comparable to state-of-the-art methods like PromptCap. This demonstrates the efficiency and strong generaliza- tion capability of our modular framework, which leverages pre-trained components for knowledge- intensive VQA tasks while avoiding t... | https://arxiv.org/abs/2505.19354v1 |
al., 2023) 175B 46.4 43.9 - DecomVQA (Khan et al., 2024) 13B 39.79 53.36 - DIETCOKE (Li et al., 2024) 7B 49.2 48.6 - ours 47B 54.57 53.87 67.96 Zero-shot evaluation with extra end-to-end training VLKD (Dai et al., 2022) 408B 13.3 - 44.5 Flamingo (Alayrac et al., 2022) 80B 50.6 - 56.3 BLIP-2 (Li et al., 2023) 12B 45.9 -... | https://arxiv.org/abs/2505.19354v1 |
miss out on sufficient image-related content. Cap- tions help the model associate specific regions and objects in the image with textual descriptions, mak- ing it easier to locate relevant entities when answer- ing. However, incorporating structured reasoning examples can further enhance accuracy. The in- tegration of ... | https://arxiv.org/abs/2505.19354v1 |
Barezi and Parisa Kordjamshidi. 2024. Find the gap: Knowledge base reasoning for visual question answering. arXiv preprint arXiv:2404.10226 . Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 . Remi Cadene, Hedi Ben-Younes, Matthieu Cord, and Nicolas Thome. 2019. Murel: Multimoda... | https://arxiv.org/abs/2505.19354v1 |
, 36. Yunshi Lan, Xiang Li, Xin Liu, Yang Li, Wei Qin, and Weining Qian. 2023. Improving zero-shot visual question answering via large language models with reasoning question prompts. In Proceedings of the 31st ACM International Conference on Multimedia , pages 4389–4400. Junnan Li, Dongxu Li, Silvio Savarese, and Stev... | https://arxiv.org/abs/2505.19354v1 |
in large language models: Knowledge, retrieval and domain-specificity. arXiv preprint arXiv:2310.07521 . Jialin Wu, Jiasen Lu, Ashish Sabharwal, and Roozbeh Mottaghi. 2022. Multi-modal answer validation for knowledge-based vqa. In Proceedings of the AAAI conference on artificial intelligence , volume 36, pages 2712–272... | https://arxiv.org/abs/2505.19354v1 |
have horns. QA1: What do the two animals in the image have in common? Horns QA2: What feature of the two animals’ heads is similar in the image? Horns Predicted Answer: Horns Question: What is the person on the left doing with their body? GT: surfing/crouching/bending/ KeyWords: body, person, left Question Concept: act... | https://arxiv.org/abs/2505.19354v1 |
What type of structure is being used? GT: cement/concrete/barrier/concrete barrier KeyWords: structure Question Concept: type of structure being used Caption1: Skateboarding on a makeshift ramp made of urban elements. Caption2: In the image, a person is riding a skateboard on top of a concrete barrier. Caption3: This p... | https://arxiv.org/abs/2505.19354v1 |
are no leaves on the trees. QA1: What is the state of the trees in Autumn’s touch? Bare QA2: What can be attributed to the lack of leaves on the trees? Autumn Predicted Answer: Autumn Question: What is the man in green’s job? GT: construction worker/road construction/street maintenance KeyWords: green Question Concept:... | https://arxiv.org/abs/2505.19354v1 |
style with her lip piercing, adding a touch of rebellion to her smile. Caption3: The girl is wearing a lip ring that adds to her unique style. QA1: What type of body modification is featured in the image? Lip ring QA2: What does the lip piercing add to the girl’s smile? Rebellion Predicted Answer: Vertical labret Quest... | https://arxiv.org/abs/2505.19354v1 |
Caption3: In the image, there are several dishes that appear to be part of a meal. QA1: What are they doing in the image? Posing QA2: What is visible on the table? Dishes Predicted Answer: Asian Question: When was this airline founded? GT: 1996/nineties/february 29 1996 KeyWords: airline Question Concept: founding date... | https://arxiv.org/abs/2505.19354v1 |
in 1955, this McDonald’s location has been serving up delicious meals to the community for over six decades. Caption3: The background restaurant in the image is McDon- ald’s, which was founded in 1955 by Ray Kroc. QA1: When was the background restaurant founded? 1955 QA2: What restaurant is depicted in the image? McDon... | https://arxiv.org/abs/2505.19354v1 |
older homes? Radiator Predicted Answer: Radiator Question: What piece of ski equipment do all of these people have? GT: skis/poles/ski stick KeyWords: ski equipment Question Concept: a common ski equipment item everyone has. Caption1: People in the image are equipped with ski poles. Caption2: Skiing together in the sno... | https://arxiv.org/abs/2505.19354v1 |
white circle clock on a concrete wall, serving as a silent sentinel of the passing hours. Caption3: A large, round clock face is affixed to a light grayish- tan colored concrete wall of the building. QA1: What is the white circle in the image? Clock QA2: What is the white circle clock serving as? Time-telling device Pr... | https://arxiv.org/abs/2505.19354v1 |
eating. QA1: What utensil is necessary for eating a particular meal in the image? Fork QA2: What utensils would be necessary for eating a meal with food that is served on a plate and cut into smaller pieces? Fork and knife Predicted Answer: Fork Question: What is the colorful object on the man’s head used for? GT: shad... | https://arxiv.org/abs/2505.19354v1 |
image? Glove QA2: What is a crucial tool in the game of baseball? Glove Predicted Answer: Glove 26 Examples involving counting number of objects. Question: How many animal trunks are visible here? GT: 2 KeyWords: animal trunks Predicted Answer: 2 Question: How many people have their arms around the woman? GT: 1 KeyWord... | https://arxiv.org/abs/2505.19354v1 |
offers an unexpected savory twist to the classic dish. Caption3: The green leafy vegetable is not included in this dish, as it is not typically used as a topping for pizza or flatbread. QA1: What is the unexpected flavor ingredient? Bacon QA2: What is the green leafy vegetable? Kale Predicted Answer: Kale Question: Who... | https://arxiv.org/abs/2505.19354v1 |
to the overall decor of the living room. Caption2: A wall that serves as a canvas for artistic expression, creating a harmonious and inviting atmosphere in the living space. Caption3: The wall is used to make the living room look nice. QA1: What is the purpose of the wall in the image? Decorative backdrop QA2: What doe... | https://arxiv.org/abs/2505.19354v1 |
older individuals, with a range of ages from teenagers to adults. Caption2: A diverse group of individuals, spanning from young adults to older teens, gathered for a casual pizza party. Caption3: The age range of the group is not precisely defined, but it is clear that they are young adults. QA1: What is the age range ... | https://arxiv.org/abs/2505.19354v1 |
the meal on the right would contain around 15-20 grams of protein per serving. QA1: What is the source of protein in the meal? Beans QA2: What is the estimated amount of protein in the meal on the left? 20-30 grams Predicted Answer: 20-30 grams Question: How much do these boats weigh? GT: 5 tons/4.57/1000 pounds/ton Ke... | https://arxiv.org/abs/2505.19354v1 |
standing near the bus and others walking along the sidewalk. QA1: What is parked on the side of the road? bus QA2: Where is the X22 bus ready to transport passengers? Ashington Generated Answer: 50 Q:How many people can this bus carry? GT: 50/40/39 Cap1: a passenger bus traveling on a street side. Cap2: blue commuter b... | https://arxiv.org/abs/2505.19354v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.