publications

2026

1. ICLR

   Clinical Hallucination Detection with Coverage-Gated Ontology Checks

   Hao Liu and Venkata Sai Ram Dasari

   2026

   Under review at the International Conference on Learning Representations (ICLR 2027)

   Bib

   @unpublished{liu2026clinical,
     title = {Clinical Hallucination Detection with Coverage-Gated Ontology Checks},
     author = {Liu, Hao and Dasari, Venkata Sai Ram},
     note = {Under review at the International Conference on Learning Representations (ICLR 2027)},
     year = {2026},
   }

2. HealthCom

   MAUQ-CLIP: Missingness-Aware Uncertainty Quantification for Clinical LLM Prediction

   Venkata Sai Ram Dasari, Vaibhavi Tiwari, B. Mamidala, and Vaibhav Anu

   In IEEE International Conference on E-health Networking, Application and Services (HealthCom), 2026

   Abs Bib

   Large language models (LLMs) can produce consistent clinical predictions even in the absence of key patient information, leading to the misinterpretation of output consistency as reliability. We propose MAUQ-CLIP, Missingness-Aware Uncertainty Quantification for Clinical LLM Prediction, a framework for black-box LLM-based uterine cancer prediction. MAUQ-CLIP features a weighted input-quality score, explicit missingness markers, a prompt-level data-quality advisory, sparsity-weighted ensemble disagreement, cross-task clinical consistency, model self-assessment and uncertainty-based referral for oncologist review to treat input completeness as an explicit uncertainty signal. Structured records from the Surveillance, Epidemiology, and End Results registry are converted into natural-language clinical timelines and analyzed on recurrence, survival, staging and treatment-related tasks. MAUQ-CLIP goes beyond the traditional output-centric uncertainty quantification by ensuring that the lack of clinical evidence is still captured in the final uncertainty score even when repeated model answers are consistent. Our system allows calibrated, interpretable, and clinically actionable use of LLMs for cancer decision support without access to the core model parameters or logits.

   @inproceedings{dasari2026mauqclip,
     title = {MAUQ-CLIP: Missingness-Aware Uncertainty Quantification for Clinical LLM Prediction},
     author = {Dasari, Venkata Sai Ram and Tiwari, Vaibhavi and Mamidala, B. and Anu, Vaibhav},
     booktitle = {IEEE International Conference on E-health Networking, Application and Services (HealthCom)},
     address = {New York, NY, USA},
     year = {2026},
   }

3. CCWC

   GANterpolate+: A Unified Framework for Data Reconstruction in Sparse and Heterogeneous Domains

   Vaibhavi Tiwari, Krishanu Agrawal, and Venkata Sai Ram Dasari

   In 2026 IEEE 16th Annual Computing and Communication Workshop and Conference (CCWC), Jan 2026

   Abs DOI Bib

   Reconstructing incomplete or irregular time series remains a persistent challenge across scientific, industrial, and operational domains, where missing observations can hinder accurate modeling, forecasting, and decision-making. Building on our earlier GANterpolate framework—which integrates Generative Adversarial Networks with interpolation refinement—this work introduces GANterpolate+, a generalized and domain-adaptive architecture designed for high-fidelity reconstruction in heterogeneous spatiotemporal datasets. The framework incorporates a domain-aware consistency objective to stabilize adversarial-learning under varying data regimes, along with spatiotemporal embedding layers that capture structural, temporal, and contextual dependencies without relying on domain-specific assumptions. To evaluate its cross-domain adaptability, GANterpolate+ is applied to multiple real-world settings characterized by data sparsity, distributional shifts, and irregular sampling patterns. Experimental results demonstrate that GANterpolate+ consistently surpasses interpolation-only and GAN-only baselines across both random and clustered missingness, achieving improved reconstruction accuracy, stability, and multivariate coherence. Beyond accuracy gains, the unified and modular design enables straightforward extension to new data modalities with minimal reconfiguration. These findings establish GANterpolate+ as a scalable, robust, and versatile hybrid solution for recovering high-quality information from incomplete spatiotemporal observations across diverse application domains.

   @inproceedings{tiwari2026ganterpolateplus,
     title = {GANterpolate+: A Unified Framework for Data Reconstruction in Sparse and Heterogeneous Domains},
     author = {Tiwari, Vaibhavi and Agrawal, Krishanu and Dasari, Venkata Sai Ram},
     booktitle = {2026 IEEE 16th Annual Computing and Communication Workshop and Conference (CCWC)},
     address = {Las Vegas, NV, USA},
     pages = {281--290},
     year = {2026},
     month = jan,
     publisher = {IEEE},
     doi = {10.1109/CCWC67433.2026.11393785},
   }

4. ACDSA

   Cognitive Vestigiality in AI-Assisted Learning

   Vaibhavi Tiwari, Venkata Sai Ram Dasari, and Krishanu Agrawal

   In 2026 International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA), Feb 2026

   Abs DOI Bib

   Artificial Intelligence (AI) is increasingly integrated into education, professional practice, and daily thought processes, providing efficiency and personalization. However, this trend raises concerns about the potential deterioration of human capabilities resulting from diminished engagement. Utilizing insights from neuroplasticity and the use–disuse principle, this survey presents the notion of cognitive vestigiality to elucidate how ongoing dependence on AI for memory, reasoning, and metacognitive regulation could progressively reduce independent competence. This study integrates empirical findings related to cognitive offloading, working memory, logical reasoning, and calibration accuracy, placing them in the context of biological and cultural analogies of vestigiality. In order to facilitate empirical exploration, we present measurable constructs, featuring a composite Vestigiality Index and a longitudinal methodological framework designed to uncover early signs of dependency. The implications for education highlight the importance of promoting active involvement and providing chances for independent thinking, reflection, and retrieval in environments enhanced by AI. This work systematically synthesizes evidence and integrates theory to clarify potential risks, outline testable pathways for long-term assessment, and provide recommendations informed by research for sustaining cognitive resilience in environments enriched by AI.

   @inproceedings{tiwari2026cognitive,
     title = {Cognitive Vestigiality in AI-Assisted Learning},
     author = {Tiwari, Vaibhavi and Dasari, Venkata Sai Ram and Agrawal, Krishanu},
     booktitle = {2026 International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA)},
     address = {Boracay Island, Philippines},
     pages = {1--7},
     year = {2026},
     month = feb,
     publisher = {IEEE},
     doi = {10.1109/ACDSA67686.2026.11468156},
   }

5. IEMTRONICS

   The Future of Artificial Intelligence in Forensics: Advancements, Challenges, and Ethical Considerations

   Vaibhavi Tiwari, Venkata Sai Ram Dasari, and Jiayin Wang

   In Proceedings of IEMTRONICS 2025, 2026

   Abs DOI Bib

   The integration of artificial intelligence (AI) in forensic science is transforming investigative methodologies by enhancing efficiency, accuracy, and predictive capabilities. AI-driven forensic techniques, including machine learning (ML), deep learning (DL), and natural language processing (NLP), facilitate the rapid analysis of evidence, detection of patterns, and minimization of human error. This paper explores the current applications of AI in forensic science, covering digital forensics, biometric analysis, predictive policing, and forensic DNA examination. Additionally, it highlights emerging advancements such as deepfake detection, behavioral analysis, and blockchain-based evidence authentication. While AI presents unparalleled opportunities for forensic investigations, its implementation raises critical challenges, including algorithmic bias, ethical concerns, data security risks, and legal admissibility. This study emphasizes the need for robust technical frameworks, regulatory standards, and ethical considerations to ensure AI’s responsible and effective use in forensic science. The findings contribute to the ongoing discourse on AI’s role in forensic advancements, underscoring its potential to redefine crime-solving strategies while maintaining justice and integrity.

   @incollection{tiwari2026forensics,
     title = {The Future of Artificial Intelligence in Forensics: Advancements, Challenges, and Ethical Considerations},
     author = {Tiwari, Vaibhavi and Dasari, Venkata Sai Ram and Wang, Jiayin},
     booktitle = {Proceedings of IEMTRONICS 2025},
     series = {Lecture Notes in Electrical Engineering},
     pages = {351--374},
     year = {2026},
     publisher = {Springer Nature Singapore},
     doi = {10.1007/978-981-95-0429-9_24},
   }

2025

1. UEMCON

   Extending GANterpolate: Multi-Variable Synthetic Data Generation for Climate Modeling

   Vaibhavi Tiwari, Venkata Sai Ram Dasari, and Krishanu Agrawal

   In 2025 IEEE 16th Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON), Oct 2025

   Best Paper Award Abs DOI Bib

   Best Paper Award, IEEE UEMCON 2025.

   The precise reconstruction of atmospheric variables plays a crucial role in the progression of climate modeling; however, the presence of observational gaps in critical areas undermines the dependability of reanalysis and forecasting systems. In response to this challenge, we enhance the GANterpolate framework by developing a multivariate architecture that simultaneously reconstructs temperature, pressure, wind speed, and precipitation. This is achieved by merging the generative capabilities of Generative Adversarial Networks (GANs) with the local accuracy of interpolation methods. The comprehensive model undergoes assessment in both random and clustered masking situations to replicate real-world data scarcity and is juxtaposed with linear interpolation, cubic interpolation, and GAN-only benchmarks. The findings indicate that the hybrid method leads to lower reconstruction errors and enhanced correlations with actual data, while successfully capturing inter-variable dependencies and maintaining fine-scale details. The results indicate that GANterpolate serves as a strong and adaptable approach for improving the quality of climate reanalysis, facilitating forecasting applications, and allowing for the generation of highfidelity synthetic data in various environmental fields.

   @inproceedings{tiwari2025extending,
     title = {Extending GANterpolate: Multi-Variable Synthetic Data Generation for Climate Modeling},
     author = {Tiwari, Vaibhavi and Dasari, Venkata Sai Ram and Agrawal, Krishanu},
     booktitle = {2025 IEEE 16th Annual Ubiquitous Computing, Electronics \& Mobile Communication Conference (UEMCON)},
     address = {Yorktown Heights, NY, USA},
     pages = {21--28},
     year = {2025},
     month = oct,
     publisher = {IEEE},
     doi = {10.1109/UEMCON67449.2025.11267609},
   }

2. IEMCON

   Synthetic Data as a Catalyst: Applications in Healthcare, Environment, and Beyond

   Vaibhavi Tiwari, Venkata Sai Ram Dasari, and Krishanu Agrawal

   In 2025 IEEE 16th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Oct 2025

   Abs DOI Bib

   The availability of high-quality datasets is essential for advancing artificial intelligence, yet sensitive domains such as healthcare, finance, crime investigation, and legal systems face acute data scarcity due to privacy regulations and ethical constraints. Synthetic data, generated to preserve the statistical properties of real datasets while mitigating disclosure risks, has emerged as a viable solution to these challenges. This paper reviews state-of-the-art generation methods—including generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion models, and transformer-based approaches—alongside applications in high-stakes sectors, comparisons with traditional anonymization, and the role of hybrid and privacy-preserving synthesis. We further analyze key challenges such as the realism–privacy trade-off, evaluation gaps, bias amplification, and regulatory uncertainty, while outlining future directions in domain-specific pipelines, governance frameworks, and synthetic data marketplaces. Our findings highlight synthetic data as a cornerstone for ethical, reproducible, and scalable AI deployment in sensitive and data-restricted environments.

   @inproceedings{tiwari2025synthetic,
     title = {Synthetic Data as a Catalyst: Applications in Healthcare, Environment, and Beyond},
     author = {Tiwari, Vaibhavi and Dasari, Venkata Sai Ram and Agrawal, Krishanu},
     booktitle = {2025 IEEE 16th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)},
     address = {Berkeley, CA, USA},
     pages = {417--426},
     year = {2025},
     month = oct,
     publisher = {IEEE},
     doi = {10.1109/IEMCON67450.2025.11381044},
   }

3. IEMCON

   Mitigating Cryptocurrency Misuse: A Survey with a Cross-Domain Adaptive Framework

   Vaibhavi Tiwari, Venkata Sai Ram Dasari, and Krishanu Agrawal

   In 2025 IEEE 16th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Oct 2025

   Abs DOI Bib

   Blockchain enables decentralized, transparent, and tamper-resistant value transfer, yet the same properties also create avenues for misuse, including money laundering, ransomware payments, darknet trade, DeFi exploits, NFT wash trading, and sanctions evasion via stablecoins. This paper offers a comprehensive analysis that links blockchain’s foundational attributes—pseudonymity, immutability, and global accessibility—to observable misuse patterns, organizes the principal threat modes with empirical evidence, and evaluates mitigation across regulatory controls, technological advances, decentralized countermeasures, and industry practices. Building on comparative insights from federated learning, decentralized identity, and distributed storage, the paper then introduces the Cross-Domain Adaptive Mitigation Framework (CDAMF), a layered, privacy-preserving approach that combines reputation-backed accountability, anomaly-driven adaptive guardrails, and parametric backstops with community adjudication, with a feedback loop that updates reputations, thresholds, and models over time. An integrated architecture and event-time algorithm are presented to operationalize CDAMF while preserving decentralization. Together, the survey and framework provide a practical pathway to balance innovation, privacy, and security, strengthening trust, interoperability, and resilience in cryptocurrency ecosystems.

   @inproceedings{tiwari2025cryptocurrency,
     title = {Mitigating Cryptocurrency Misuse: A Survey with a Cross-Domain Adaptive Framework},
     author = {Tiwari, Vaibhavi and Dasari, Venkata Sai Ram and Agrawal, Krishanu},
     booktitle = {2025 IEEE 16th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON)},
     address = {Berkeley, CA, USA},
     pages = {427--436},
     year = {2025},
     month = oct,
     publisher = {IEEE},
     doi = {10.1109/IEMCON67450.2025.11381178},
   }