Dr. Prithviraj (Raj) Dasgupta, Section Head (Acting), Distributed Intelligent Systems Section, Naval Research Laboratory, Washington, DC, USA.

Abstract: Reinforcement learning (RL) is a widely used learning paradigm that has shown significant successes in solving many hard AI problems including mastering real-time strategy games, autonomous car driving, and LLM model alignment. The formal mathematical framework underlying many of the problems solved by RL is game theory. However, these two areas are taught, and usually researched, independently of each other. In this tutorial, I will attempt to bridge this gap by introducing the fundamental concepts in RL and game theory and draw parallels between RL algorithm concepts like value updates, credit assignment, advantage and policy convergence, and their counterparts in game theory like backward induction, Nash equilibrium and regret. We will use a 2-player game in an AI Gymnasium environment as a hands-on, working example to illustrate how these concepts are analyzed and solved in RL and game theory respectively.

Bio: Dr. Prithviraj (Raj) Dasgupta is a senior research scientist and Section Head (Acting) of the Distributed Intelligent Systems Section at the Naval Research Laboratory, Washington, DC. His research is in the area of artificial intelligence focusing on reinforcement learning, game theory and multi-agent systems. He has led several large, federally-funded research projects in these areas and has over 150 research publications in premier journals and conferences on these topics. From 2001-2019, Dr. Dasgupta was a tenured, full professor at the University of Nebraska, Omaha where he had established and led the robotics lab called CMANTIC, and, developed and taught several courses on game theory, multi-robot systems and machine learning. He has received several awards including the best researcher award called ADROCA at the University of Nebraska, and multiple best paper awards; he is a senior member of IEEE. He received his Ph.D. in Computer Engineering in 2001 from the University of California, Santa Barbara.

Pavan Reddy, The George Washington University

Abstract: As LLM systems move from prototypes into real products and research stacks, security and robustness are often underexamined relative to capability gains. This hands-on tutorial presents a code-first, Attack→Defense workflow for prompt injection in retrieval-augmented generation (RAG) and toolusing LLMs. Using a small Car Dealership web application (run in Google Colab or locally via Docker) and prepared notebooks, we reproduce three escalating scenarios and implement focused mitigations: (1) LLM→Database integration with direct and indirect prompt injection that manipulates database state; (2) EchoLeak-style indirect prompt injection for sensitive data exfiltration; and (3) injection-driven remote code execution via an LLM-controlled tool chain. Each module instruments minimal measurements (e.g., context recall, answer faithfulness, tool-call traces) and introduces lightweight defenses suitable for research and teaching (schema-validated tool calls, retrieval/source isolation, prompt/routing hardening). Attendees leave with runnable notebooks, a containerized demo app, drop-in attack and defense modules, and a repeatable evaluation workflow. This tutorial benefits scientists, applied researchers, and students who need rigorous, reproducible methods to analyze and improve LLM pipeline behavior.

Bio: Pavan Reddy is a software engineer at Automata, where he leads efforts in vulnerability management, FedRAMP ATO preparation, FIPS compliance, and the roadmap for extending security practices to AI systems. His work spans adversarial machine learning, LLM robustness, and RAG evaluation. He has delivered talks and hands-on tutorials at venues including SquadCon, ACM SIGCITE, FedCertWeek, CAPWIC, and BSidesNoVA, primarily on adversarial ML and AI security topics. He is also set to deliver a Lab session at AAAI 2026. Pavan maintains an active social-media presence sharing short-form ML/LLM content and live demos, and regularly publishes teaching materials. He combines applied engineering with research and academic experience, making him well-suited to deliver an entirely hands-on tutorial that emphasizes reproducibility, measurement, and practical experimentation for scientists and applied researchers.

Dr. Prithviraj (Raj) Dasgupta, Section Head (Acting), Distributed Intelligent Systems Section, Naval Research Laboratory, Washington, DC, USA.

Abstract: Reinforcement learning (RL) is a widely used learning paradigm that has shown significant successes in solving many hard AI problems including mastering real-time strategy games, autonomous car driving, and LLM model alignment. The formal mathematical framework underlying many of the problems solved by RL is game theory. However, these two areas are taught, and usually researched, independently of each other. In this tutorial, I will attempt to bridge this gap by introducing the fundamental concepts in RL and game theory and draw parallels between RL algorithm concepts like value updates, credit assignment, advantage and policy convergence, and their counterparts in game theory like backward induction, Nash equilibrium and regret. We will use a 2-player game in an AI Gymnasium environment as a hands-on, working example to illustrate how these concepts are analyzed and solved in RL and game theory respectively.

Bio: Dr. Prithviraj (Raj) Dasgupta is a senior research scientist and Section Head (Acting) of the Distributed Intelligent Systems Section at the Naval Research Laboratory, Washington, DC. His research is in the area of artificial intelligence focusing on reinforcement learning, game theory and multi-agent systems. He has led several large, federally-funded research projects in these areas and has over 150 research publications in premier journals and conferences on these topics. From 2001-2019, Dr. Dasgupta was a tenured, full professor at the University of Nebraska, Omaha where he had established and led the robotics lab called CMANTIC, and, developed and taught several courses on game theory, multi-robot systems and machine learning. He has received several awards including the best researcher award called ADROCA at the University of Nebraska, and multiple best paper awards; he is a senior member of IEEE. He received his Ph.D. in Computer Engineering in 2001 from the University of California, Santa Barbara.

Pavan Reddy, The George Washington University

Abstract: As LLM systems move from prototypes into real products and research stacks, security and robustness are often underexamined relative to capability gains. This hands-on tutorial presents a code-first, Attack→Defense workflow for prompt injection in retrieval-augmented generation (RAG) and toolusing LLMs. Using a small Car Dealership web application (run in Google Colab or locally via Docker) and prepared notebooks, we reproduce three escalating scenarios and implement focused mitigations: (1) LLM→Database integration with direct and indirect prompt injection that manipulates database state; (2) EchoLeak-style indirect prompt injection for sensitive data exfiltration; and (3) injection-driven remote code execution via an LLM-controlled tool chain. Each module instruments minimal measurements (e.g., context recall, answer faithfulness, tool-call traces) and introduces lightweight defenses suitable for research and teaching (schema-validated tool calls, retrieval/source isolation, prompt/routing hardening). Attendees leave with runnable notebooks, a containerized demo app, drop-in attack and defense modules, and a repeatable evaluation workflow. This tutorial benefits scientists, applied researchers, and students who need rigorous, reproducible methods to analyze and improve LLM pipeline behavior.

Bio: Pavan Reddy is a software engineer at Automata, where he leads efforts in vulnerability management, FedRAMP ATO preparation, FIPS compliance, and the roadmap for extending security practices to AI systems. His work spans adversarial machine learning, LLM robustness, and RAG evaluation. He has delivered talks and hands-on tutorials at venues including SquadCon, ACM SIGCITE, FedCertWeek, CAPWIC, and BSidesNoVA, primarily on adversarial ML and AI security topics. He is also set to deliver a Lab session at AAAI 2026. Pavan maintains an active social-media presence sharing short-form ML/LLM content and live demos, and regularly publishes teaching materials. He combines applied engineering with research and academic experience, making him well-suited to deliver an entirely hands-on tutorial that emphasizes reproducibility, measurement, and practical experimentation for scientists and applied researchers.

David Bisant, Central Security Service

Abstract: Deep learning neural models require hardware acceleration. The current thirst for this acceleration is exceeding current capabilities and reality.  At current trends, by 2045, one half of the world's electricity will be consumed by training deep learning models.  This tutorial will cover background and a history of the field, the acceleration which is currently available, and what is expected in the future. 

Bio: Dr. David Bisant has over 30 years of experience in neural networks, machine learning, and the application of these algorithms to problems in engineering and natural sciences. He has received training at Colorado State University, the University of Maryland, George Washington University, and Stanford University. He has held past positions at Medtronic and Stanford University. He is currently a member of the Central Security Service, where he works in the fields of high performance computing, physical science research, and defense.  He has been both a contributor and organizer of the FLAIRS Conference.  He has cochaired a number of special tracks, primarily the Neural Network and Data Mining Special Track, which he has cochaired for the last 16 years.

Re-thinking Education in the Face of AI

William Swartout, Chief Science Officer, USC Institute for Creative Technologies

Abstract: When ChatGPT was released in the fall of 2022, the education community panicked.  Suddenly there existed a highly capable AI that was facile with language.  Teachers feared students would use Gen AI to cheat and write their essays for them, and the press and internet were rife with articles proclaiming the Death of the Term Paper.  Banning AI was problematic since preventing students from using AI while in school would not prepare them for the world into which they would graduate, and the detectors that purported to tell if text was written by an AI or a human had significant false positive and false negative error rates.  Working with faculty from the USC undergraduate writing program we developed a writing tool called ABE that takes a different approach.  In ABE, we use generative AI to help students brainstorm about a topic, and then when they are finished with their essays (which they write themselves) we use generative AI again, not as a writer, but as a reader, to read their essays and offer critiques, answering questions such as: Does the essay have a good hook?  Is there adequate support for the claims?  Are there other points of view that should be considered, but weren't?   Surveys have shown that students have received ABE very positively and have found it helpful in their writing.Stepping back a bit, I believe Gen AI is going to force us to reconsider how we teach across a very broad spectrum of intellectual domains.  While each domain presents its own challenges, I believe that our experience with ABE is an exemplar of how Gen AI can be integrated into instructional design to actually improve students' critical thinking skills rather than detract from them.

Bio: William Swartout is chief science officer at the USC Institute for Creative Technologies, providing overall direction to the institute’s research programs. He is also co-Director of the Center for Generative AI and Society and a research professor in the Computer Science Department at the USC Viterbi School of Engineering.Swartout has been involved in cutting edge research and development of artificial intelligence systems throughout his career. In 2009, Swartout received the Robert Engelmore Award from the Association for the Advancement of Artificial Intelligence (AAAI) for seminal contributions to knowledge-based systems and explanation, groundbreaking research on virtual human technologies and their applications, and outstanding service to the artificial intelligence community. Swartout is a Fellow of the AAAI, has served on their Board of Councilors and is past chair of the Special Interest Group on Artificial Intelligence (SIGART) of the Association for Computing Machinery (ACM).He has served as a member of the Air Force Scientific Advisory Board, the Board on Army Science and Technology of the National Academies and the JFCOM Transformation Advisory Group. Prior to helping found the ICT in 1999, Swartout was the Director of the Intelligent Systems Division at the USC Information Sciences Institute. His particular research interests include virtual humans, natural language processing, particularly explanation and text generation, knowledge acquisition, knowledge representation, and intelligent computer based education. He received his Ph.D. and M.S. in computer science from MIT and his bachelor’s degree from Stanford University.

Explainable Neural Text Classification

Diana Inkpen, Professor, School of Electrical Engineering and Computer Science,University of Ottawa, Canada

Abstract:  Advances in Large Language Models (LLMs) allow us to develop highly accurate neural text classifiers. One of their major disadvantages is their lack of explainability, due to their black box nature. I am looking into neural text classifiers that are explainable, in order to open their black box architecture, at least partially. Explainability can come at the level of the classification model or at the level of the decision made for each new test data. The explanations need to look into what was learnt from the training data (unless there is no training or minimal training) and also into the pre-trained model (LLM) that was used as a basis for the classifier. To explain the individual decisions for each test data, one step is to calculate feature importance with methods such as LIME, SHAP, or Integrated Gradients. More useful full-text explanations can be generated via customized prompting, or via joint leaning of classes and explanations during training. I will show results for two case studies: applications to legal text mining and to mental health text mining. The evaluation of the generated explanations is done via automatic measures, as well as with human judges, in order to see if they find the explanations relevant and useful.

Bio:  Diana Inkpen is a Professor at the University of Ottawa, in the School of Electrical Engineering and Computer Science. She received her Ph.D. in Computer Science from the University of Toronto, Canada, and her M.Sc. and B.Eng. in Computer Science and Engineering from the Technical University of Cluj-Napoca, Romania. Her research is in applications of Natural Language Processing and Deep Learning. She is the editor-in-chief of the Computational Intelligence journal and the associate editor for the Natural Language Engineering journal. She published a book on Natural Language Processing for Social Media (Morgan and Claypool Publishers, Synthesis Lectures on Human Language Technologies, the third edition appeared in 2020), 11 book chapters, more than 45 journal articles, and more than 150 conference papers. She has received many research grants, the majority of which include intensive industrial collaborations.

From Hallucinations to Hybrid Interpretability: Responsible NLP for High-Stakes Social Media Signals

Bonnie J. Dorr, Professor, University of Florida

Abstract: Generative AI is powerful but often unreliable in high-stakes settings, where hallucinations and overconfidence can cause real harm. This talk presents a responsible NLP framework for analyzing sensitive social media signals through hybrid interpretability, ambiguity-aware inference, and privacy-first pipelines—treating mental-health-adjacent detection as a stress test for trustworthy language AI.

Bio: Dr. Bonnie Dorr is a Professor in the Department of Computer & Information Science & Engineering at the University of Florida and Director of the NLP & Culture Laboratory. Her expertise spans trustworthy and interpretable NLP, ambiguity-aware inference, and responsible language technologies for high-stakes applications, including social media analysis. At FLAIRS, she would contribute technical depth in responsible AI and hybrid interpretability, helping broaden FLAIRS’ coverage of cutting-edge NLP while strengthening its focus on real-world deployment, evaluation rigor, and societal impact.

AI Systems That Think, Team, and Fight: A New Paradigm for Defense 

Svitlana Volkova, Chief of AI, Office of Science and TechnologyAptima, Inc.

Abstract:  As AI systems become increasingly capable, the Department of War faces a critical challenge: how do we develop, rigorously evaluate, and safely deploy multi-agent AI frontier systems across domains ranging from multimodal knowledge discovery to cognitive warfare? This talk presents lessons learned from building compound AI architectures that orchestrate large language models, vision-language models, and specialized agents through retrieval-augmented generation and agentic AI workflows. I will demonstrate how these systems enable cross-disciplinary knowledge synthesis for biosecurity, cognitive warfare planning and execution, and operator-AI team optimization in wargaming and readiness applications. Finally, I will present our emerging capabilities in multi-domain wargaming, where cognitively inspired AI agents execute doctrine-based maneuvers across air, space, cyber, and information domains.Evaluating these systems requires moving beyond traditional AI benchmarks. I will present our multi-dimensional ecosystem combining quantitative measures, qualitative SME assessments scaled through simulated domain expert agents, and causal investigations using structure learning algorithms to understand "why" behaviors emerge and "how" interventions affect mission outcomes. For safety evaluation, we examine human-agent-environment interactions holistically addressing alignment failures, emergent capabilities under distributional shift, and systemic risks from multi-agent coordination through counterfactual "what-if" analysis and continuous monitoring. The era of scientifically grounded operationally validated human-AI team optimization has begun, and this talk charts the path forward for defense applications.

Bio:  Dr. Svitlana Volkova is Chief of AI at Aptima, Inc., where she sets the company's AI vision and leads a portfolio of advanced research programs in compound frontier AI systems, human-AI teaming, and AI Test and Evaluation for national defense. A recognized thought leader in AI for national security, she has shaped the technical direction of multi-million-dollar federal research initiatives with a focus on transitioning AI technologies to operational use. Her pioneering work spans multimodal frontier models, agentic AI architectures, human digital twins, and causal AI/ML—with a focus on decision advantage, readiness, and cognitive warfare applications. Dr. Volkova has authored 100+ publications with 4,900+ citations, delivered keynotes and invited talks at premier venues spanning AI research (AAAI, ACL, EMNLP), defense (I/ITSEC, MODSIM, INFOPAC), academia (Stanford, CMU), and industry (Google Research, Amazon), and served as a trusted advisor to government leadership on AI strategy. Prior to Aptima, she led AI research initiatives at Pacific Northwest National Laboratory and conducted research at Microsoft Research. She holds a PhD in Computer Science from Johns Hopkins University.

Business Meeting