Research in Responsible AI Lab

The Responsible AI Lab designs, analyzes, and deploys trustworthy AI systems that are scalable, privacy-preserving, fair, robust, and useful in high-impact domains. Our recent work emphasizes the system-level foundations of responsible AI: efficient LLM serving and agent runtimes, privacy-preserving AI pipelines, trustworthy generative AI for health, fairness and robustness evaluation of deployed models, and AI systems for edge, IoT, manufacturing, agriculture, scientific computing, and biomedical discovery.

The selected highlights below focus on representative papers rather than an exhaustive publication list. Together, they show how the lab integrates model design, deployment infrastructure, evaluation protocols, human oversight, and domain constraints into dependable end-to-end AI systems.

LLM Systems

Scalable LLM Serving and Agent Systems

System focus: We study how large language models behave as deployed systems, including token flow, latency, context management, agent orchestration, and efficient long-context inference.

ProMCP: Profiling Token Flows and Latency Costs in Model Context Protocol-Based LLM Agents

Findings of the Association for Computational Linguistics (ACL'26)

ProMCP analyzes Model Context Protocol-based LLM agents as end-to-end runtime systems. It profiles token movement, context construction, tool invocation, model calls, and latency costs across agent workflows, making performance bottlenecks visible at a level that application developers can act on. The work supports more transparent debugging, resource-aware orchestration, and scalable deployment of tool-using LLM agents.

LLM AgentsModel Context ProtocolToken FlowLatency Profiling

DAM: Dynamic Attention Mask for Long-Context LLM Inference Acceleration

DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration

Findings of the Association for Computational Linguistics (ACL'25)

DAM accelerates long-context LLM inference by learning dynamic sparse attention masks from frozen pretrained models. Unlike fixed sparse patterns, it captures heterogeneous attention behavior across layers and heads while reducing memory and computation requirements. The framework makes long-context inference more practical across model sizes and sequence lengths without requiring full model fine-tuning.

Long-Context LLMsSparse AttentionInference AccelerationSystem Efficiency

Privacy Systems

Privacy-Preserving AI Systems

System focus: We build privacy layers for AI services and evaluation workflows, protecting sensitive prompts, datasets, model interfaces, and performance assessments without sacrificing practical usefulness.

DP-GTR: Differentially Private Prompt Protection via Group Text Rewriting

Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP'25)

DP-GTR protects sensitive prompts before they are submitted to online LLM services. The framework combines local differential privacy with group text rewriting and uses both document-level and word-level signals to preserve semantic utility while transforming private content. As a deployable privacy layer, it supports safer LLM use in clinical, educational, personal, and proprietary settings.

Differential PrivacyPrompt ProtectionText RewritingOnline LLM Services

DAAP: Privacy-Preserving Model Accuracy Estimation on Unlabeled Datasets Through Distribution-Aware Adversarial Perturbation

USENIX Security'24

DAAP estimates model performance on private, unlabeled datasets without requiring direct exposure of the full test set. It uses a public intermediary dataset and distribution-aware adversarial perturbations to bridge distribution gaps in white-box and black-box settings. The resulting evaluation pipeline protects both data and model assets while enabling reliable assessment of deployed AI systems.

Private EvaluationUnlabeled DataAdversarial LearningModel Assessment

Health AI

Trustworthy Health AI Systems

System focus: We integrate generative models with expert review, retrieval grounding, hallucination analysis, fidelity auditing, and actionable feedback for high-stakes health and mental-health applications.

Expert-Verified Synthetic Mental Health Data

Generating High-Fidelity Synthetic Mental Health Records with Expert-in-the-Loop Verification

ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE'26)

This work addresses the scarcity and sensitivity of mental-health data through a generation-and-verification system for synthetic clinical records. Domain experts assess realism, consistency, safety, and downstream usefulness, creating a human-in-the-loop quality-control process rather than relying on automatic generation alone. The framework supports responsible model development and data sharing while reducing dependence on directly identifiable clinical records.

Synthetic Health RecordsMental Health AIExpert VerificationPrivacy-Preserving Data

HALO: Grounded Clinical Decision Support

HALO: Hallucination Analysis and Learning Optimization to Empower LLMs with Retrieval-Augmented Context for Guided Clinical Decision Making

ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE'25)

HALO treats hallucination as an end-to-end reliability problem in LLM-assisted clinical decision support. The system combines hallucination analysis, learning optimization, and retrieval-augmented context so generated guidance remains connected to more verifiable evidence. This work supports clinical AI tools that are grounded, auditable, and better aligned with high-stakes decision-making needs.

Clinical Decision SupportRAGHallucination AnalysisTrustworthy LLMs

HealthLit: A Large Language Model Driven Health Literacy Fidelity Audit and Feedback System

ACM/IEEE International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE'25)

HealthLit uses LLMs to audit and provide feedback on the health-literacy fidelity of patient-facing information. It evaluates whether communication is understandable, accessible, and faithful to intended medical guidance, then produces structured feedback that can be incorporated into content-development workflows. The project embeds human-centered evaluation directly into a practical health communication system.

Health LiteracyAI AuditingFeedback SystemsAccessible Communication

Fairness and Evaluation

Fairness, Bias, Robustness, and Evaluation

System focus: We design auditing and mitigation methods for deployed, compressed, generative, and search systems. This research treats fairness and robustness as operational properties that must be measured and improved across the full AI lifecycle.

Task-Free Fairness-Aware Bias Mitigation for Black-Box Deployed Models

IEEE Transactions on Dependable and Secure Computing

This project mitigates bias in black-box deployed models when auditors may not have access to training data, architecture, or internal parameters. It supports practical post-deployment improvement through observable model behavior rather than assuming full retraining or model modification. The work frames fairness as an ongoing systems requirement for operational AI services.

Black-Box ModelsBias MitigationPost-Deployment AuditingResponsible AI

Towards Fairness-aware Adversarial Network Pruning

International Conference on Computer Vision (ICCV'23)

This work develops an end-to-end framework that jointly optimizes neural network compression, predictive accuracy, and demographic fairness. Rather than applying pruning and debiasing as separate stages or relying on handcrafted pruning rules, the method uses adversarial learning to optimize pruning behavior against utility and fairness criteria, including disparate impact and equalized odds. The framework adapts across network architectures and demonstrates that compressed models can preserve strong pruning performance while substantially reducing demographic disparities. From a systems perspective, the project makes fairness a first-class constraint in model compression and resource-efficient deployment.

Fairness-Aware PruningModel CompressionAdversarial TrainingEfficient Deployment

PreciseDebias: Automatic Prompt Engineering to Mitigate Image Demographic Biases

PreciseDebias: An Automatic Prompt Engineering Approach for Generative AI to Mitigate Image Demographic Biases

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV'24)

PreciseDebias reduces demographic bias in text-to-image generation by transforming generic prompts into demographically informed prompts. An instruction-following LLM reasons about target demographic distributions and guides image generation toward more balanced outcomes. The work contributes an end-to-end mitigation pipeline that connects bias measurement, prompt engineering, and image generation.

Generative AIPrompt EngineeringDemographic BiasImage Generation

Has CEO Gender Bias Really Been Fixed? Adversarial Attacking and Improving Gender Fairness in Image Search

AAAI'22

This work uses adversarial query construction to audit gender bias in deployed image search engines. Profession- and country-based queries reveal representation failures that remain hidden under ordinary evaluation, and reranking methods are introduced to improve gender fairness in returned results. The project frames search fairness as an end-to-end systems problem involving data, ranking, user-facing outputs, adversarial testing, and mitigation.

Image SearchGender FairnessAdversarial AuditingReranking

Edge, IoT, and Scientific AI

Edge/IoT and Scientific AI Systems

System focus: We connect AI models with sensors, devices, domain workflows, uncertainty reasoning, and hardware acceleration so intelligent systems can operate reliably in agriculture, manufacturing, biology, and scientific computing.

SocialCattle: IoT-Based Mastitis Detection and Control Through Social Cattle Behavior Sensing in Smart Farms

IEEE Internet of Things Journal

SocialCattle is an IoT smart-farm system that uses wearable GPS sensing to capture movement and contact patterns, constructs directed and weighted cattle social graphs, and estimates mastitis infection risk through probabilistic disease-transmission modeling. It integrates sensing, graph analytics, risk inference, and field validation into a deployable precision-agriculture decision-support pipeline.

Smart AgricultureIoT SensingDisease Risk ModelingDecision Support

DeepWelding: A Deep Learning Enhanced Approach to GTAW Using Multisource Sensing Images

IEEE Transactions on Industrial Informatics

DeepWelding integrates multisource optical sensing with deep learning for gas tungsten arc welding process monitoring and penetration detection. The system includes image denoising, sensing-image selection, classification, and ensemble strategies designed to improve robustness across changing experimental conditions. It highlights the importance of sensor quality, domain shift, and deployment reliability in smart manufacturing.

Smart ManufacturingMultisource SensingProcess MonitoringRobust Deployment

OrgaCast: A Trustworthy Spatiotemporal Diffusion Model for Fluorescence Organoid Forecasting

AAAI'26

OrgaCast forecasts cardiac organoid fluorescence by combining historical image sequences, experimental metadata, biological descriptions, and spatiotemporal diffusion modeling. Its uncertainty-aware inference component communicates confidence alongside predictions, supporting more trustworthy biomedical forecasting and in-silico experimentation. The project connects generative AI with scientific discovery infrastructure rather than treating image generation as an isolated task.

AI for BiologySpatiotemporal DiffusionUncertainty AwarenessScientific Discovery

These representative projects share a common goal: to make AI systems not only capable, but also efficient, private, fair, auditable, and dependable in the environments where they are actually used.