Jingyu Yao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jingyu Yao has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Litong Song
Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX 78712, USA
2 shared publications - Weishen Chu
Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL 60208, USA
2 shared publications - Chao Yao
Ira A. Fulton Schools of Engineering, Arizona State University, Tempe, AZ 85281, USA
1 shared publication - Chenchen Yu
Trine University, USA
1 shared publication - Danni Huang
College of Graduate and Professional Studies, Trine University, Angola, IN 46703, USA
1 shared publication - Hui Wang
Department of Computer Science, Yale University, New Haven, CT 06520, USA
1 shared publication
Latest Publications
- Open Access
- 8 October 2026
- Published by MDPI AG in Preprints.org
Beyond Accuracy: Explainable Analysis of Weather and Lighting Recognition for Autonomous Driving
Autonomous vehicles (AVs) must recognize weather and lighting conditions that affect perception and driving decisions. Classification accuracy alone, however, does not reveal which visual evidence a model uses. We apply explainable artificial intelligence (XAI) to condition recognition on 31,420 CARLA images spanning 21 weather–time classes. ResNet18 and ResNet50 are compared with global color baselines, and Grad-CAM and Integrated Gradients (IG) explain predictions from a designated ResNet18 model. Across three training seeds, the neural models achieve internal joint accuracies of 99.547 ± 0.123% and 99.710 ± 0.117%, respectively (mean ± sample standard deviation), with detected structural-similarity groups kept within partitions. Analysis of 210 correct and all 18 incorrect predictions reveals masking sensitivity to replacement choices and prediction correctness. After full parameter randomization, signed IG correlations approach zero while positive-grid correlations remain around 0.4, showing that different attribution representations can yield different assessments. These findings demonstrate how explanatory analysis complements accuracy in evaluating AV perception models, while independent-route generalization remains unverified. Future work will extend the analysis from simulation to real driving images using the Adverse Conditions Dataset with Correspondences (ACDC), with compatible tasks for comparing explanations across domains. This simulation-to-real extension aims to inform the development of explanation-aware AV benchmarks that evaluate predictive performance alongside explanation robustness.
- 8 October 2026
- Published by MDPI AG in Preprints.org
Open AccessBeyond Accuracy: Explainable Analysis of Weather and Lighting Recognition for Autonomous Driving
Autonomous vehicles (AVs) must recognize weather and lighting conditions that affect perception and driving decisions. Classification accuracy alone, however, does not reveal which visual evidence a model uses. We apply explainable artificial intelligence (XAI) to condition recognition on 31,420 CARLA images spanning 21 weather–time classes. ResNet18 and ResNet50 are compared with global color baselines, and Grad-CAM and Integrated Gradients (IG) explain predictions from a designated ResNet18 model. Across three training seeds, the neural models achieve internal joint accuracies of 99.547 ± 0.123% and 99.710 ± 0.117%, respectively (mean ± sample standard deviation), with detected structural-similarity groups kept within partitions. Analysis of 210 correct and all 18 incorrect predictions reveals masking sensitivity to replacement choices and prediction correctness. After full parameter randomization, signed IG correlations approach zero while positive-grid correlations remain around 0.4, showing that different attribution representations can yield different assessments. These findings demonstrate how explanatory analysis complements accuracy in evaluating AV perception models, while independent-route generalization remains unverified. Future work will extend the analysis from simulation to real driving images using the Adverse Conditions Dataset with Correspondences (ACDC), with compatible tasks for comparing explanations across domains. This simulation-to-real extension aims to inform the development of explanation-aware AV benchmarks that evaluate predictive performance alongside explanation robustness.
- Open Access
- 22 January 2026
- Published by MDPI AG in Preprints.org
TrustLLM-Fin: A Privacy-Centric and Auditable Impact Assessment Framework for Large Language Models in Automated Financial Reporting
The powerful use of Large Language Models to automate financial reporting is currently hindered by afundamental disconnect: popular models generate stochastic outputs, but the industry expects absolute precision and privacy. General-purpose LLMs are outstanding at reiterating guidelines in natural language, but are still prone to “knowledge overriding,” hallucinated financial figures, as well as adversarial prompt injections, shortcomings that are existential threats to the institutional fiduciary duties and regulatory compliance required by the EU AI Act. We propose TrustLLM-Fin, a domain specific, privacy-preserving and auditable impact assessment framework to systematically model the real-world financial disclosure workflow. We identify orthogonal dimensions in the “Trust Surface” of LLM-powered FinTech Agents: Prompt Safety & Privacy, Factuality & Robustness, Auditability, and wefurther quantize abstract governance principles into a numerical “Trust Score” using the Analytic Hierarchy Process. We test the efficacy of the TrustLLM-Fin framework in a high-fidelity financial dataset constructed from real-world SEC 10-K/10-Q filings. Strikingly, a standard general-purpose LLMscored a TS of only 0.234, due to a 93.33% success rate for context-embedded adversarial attacks. The Guardrail-Enhanced TrustLLM-Fin scored a TS of 0.710, revealing the critical decoupling between generative fluency and instruction following. Our results show that in practice, trustworthiness is not an emergent property of scale, but requires systemic alignment by design. The proposed framework provides auditors and financial institutions with a pragmatic, extensible framework to ensure the deployment of generative agents is transparent, robust and auditable in a changing regulatory landscape.
- 22 January 2026
- Published by MDPI AG in Preprints.org
Open AccessTrustLLM-Fin: A Privacy-Centric and Auditable Impact Assessment Framework for Large Language Models in Automated Financial Reporting
The powerful use of Large Language Models to automate financial reporting is currently hindered by afundamental disconnect: popular models generate stochastic outputs, but the industry expects absolute precision and privacy. General-purpose LLMs are outstanding at reiterating guidelines in natural language, but are still prone to “knowledge overriding,” hallucinated financial figures, as well as adversarial prompt injections, shortcomings that are existential threats to the institutional fiduciary duties and regulatory compliance required by the EU AI Act. We propose TrustLLM-Fin, a domain specific, privacy-preserving and auditable impact assessment framework to systematically model the real-world financial disclosure workflow. We identify orthogonal dimensions in the “Trust Surface” of LLM-powered FinTech Agents: Prompt Safety & Privacy, Factuality & Robustness, Auditability, and wefurther quantize abstract governance principles into a numerical “Trust Score” using the Analytic Hierarchy Process. We test the efficacy of the TrustLLM-Fin framework in a high-fidelity financial dataset constructed from real-world SEC 10-K/10-Q filings. Strikingly, a standard general-purpose LLMscored a TS of only 0.234, due to a 93.33% success rate for context-embedded adversarial attacks. The Guardrail-Enhanced TrustLLM-Fin scored a TS of 0.710, revealing the critical decoupling between generative fluency and instruction following. Our results show that in practice, trustworthiness is not an emergent property of scale, but requires systemic alignment by design. The proposed framework provides auditors and financial institutions with a pragmatic, extensible framework to ensure the deployment of generative agents is transparent, robust and auditable in a changing regulatory landscape.
- Top Cited PapersOpen Access
From Black Box to Glass Box: A Practical Review of Explainable Artificial Intelligence (XAI)
Explainable Artificial Intelligence (XAI) has become essential as machine learning systems are deployed in high-stakes domains such as security, finance, and healthcare. Traditional models often act as “black boxes”, limiting trust and accountability. Traditional models often act as “black boxes”, limiting trust and accountability. However, most existing reviews treat explainability either as a technical problem or a philosophical issue, without connecting interpretability techniques to their real-world implications for security, privacy, and governance. This review fills that gap by integrating theoretical foundations with practical applications and societal perspectives. define transparency and interpretability as core concepts and introduce new economics-inspired notions of marginal transparency and marginal interpretability to highlight diminishing returns in disclosure and explanation. Methodologically, we examine model-agnostic approaches such as LIME and SHAP, alongside model-specific methods including decision trees and interpretable neural networks. We also address ante-hoc vs. post hoc strategies, local vs. global explanations, and emerging privacy-preserving techniques. To contextualize XAI’s growth, we integrate capital investment and publication trends, showing that research momentum has remained resilient despite market fluctuations. Finally, we propose a roadmap for 2025–2030, emphasizing evaluation standards, adaptive explanations, integration with Zero Trust architectures, and the development of self-explaining agents supported by global standards. By combining technical insights with societal implications, this article provides both a scholarly contribution and a practical reference for advancing trustworthy AI.
Top Cited PapersOpen AccessFrom Black Box to Glass Box: A Practical Review of Explainable Artificial Intelligence (XAI)
Explainable Artificial Intelligence (XAI) has become essential as machine learning systems are deployed in high-stakes domains such as security, finance, and healthcare. Traditional models often act as “black boxes”, limiting trust and accountability. Traditional models often act as “black boxes”, limiting trust and accountability. However, most existing reviews treat explainability either as a technical problem or a philosophical issue, without connecting interpretability techniques to their real-world implications for security, privacy, and governance. This review fills that gap by integrating theoretical foundations with practical applications and societal perspectives. define transparency and interpretability as core concepts and introduce new economics-inspired notions of marginal transparency and marginal interpretability to highlight diminishing returns in disclosure and explanation. Methodologically, we examine model-agnostic approaches such as LIME and SHAP, alongside model-specific methods including decision trees and interpretable neural networks. We also address ante-hoc vs. post hoc strategies, local vs. global explanations, and emerging privacy-preserving techniques. To contextualize XAI’s growth, we integrate capital investment and publication trends, showing that research momentum has remained resilient despite market fluctuations. Finally, we propose a roadmap for 2025–2030, emphasizing evaluation standards, adaptive explanations, integration with Zero Trust architectures, and the development of self-explaining agents supported by global standards. By combining technical insights with societal implications, this article provides both a scholarly contribution and a practical reference for advancing trustworthy AI.