Biography

Kyungwhoon (Paul) Cheun, is Corporate President & Chief Technology Officer (CTO) of Device eXperience (DX) Division and Head of Samsung Research at Samsung Electronics. As CTO, he is driving the company’s technological innovations and strategy to align them with Samsung’s corporate goals. He also heads up Samsung Research, the advanced Research & Development hub of the DX Division at Samsung Electronics.

Samsung Research leads the development of future technologies and identifies new growth opportunities for the company through efforts spanning 15 R&D and AI centers located globally. Samsung Research's main focus is Artificial Intelligence, Next Generation Communications, Robotics, Security & Privacy, etc. He leads Samsung Research in constantly driving innovation across AI technologies and developing new research-powered products to provide better and more meaningful experiences to our valued customers.

Biography

Joseph Gonzalez is a Professor of Computer Science at UC Berkeley, co-director of the RISELab and Sky Computing Lab, and a core member of the Berkeley AI Research (BAIR) group. His research focuses on large language models, agentic systems, and scalable machine learning infrastructure. He is known for his contributions to projects such as Chatbot Arena, Gorilla, MemGPT, and vLLM, which have significantly advanced the field of open LLM research. Joseph is also the co-founder and Head of AI at RunLLM.com. His broader work spans data systems, robotics, computer vision, graph analytics, and autonomous systems. Prior to Berkeley, he co-founded Turi Inc. (formerly GraphLab) based on his Ph.D. work. His research has been recognized with the Okawa Research Grant, NSF CAREER Award, VLDB Test-of-Time Award, and NSF Expeditions in Computing Award.

Advancing Agentic Intelligence

Large language models (LLMs) are evolving from passive completion engines into persistent agents capable of executing complex tasks over extended interactions. In this talk, I will present recent efforts from my lab to enhance the agentic capabilities of LLM-based systems. I will introduce sleep-time compute, a paradigm that allows agents to reason, learn, and plan between interactions—unlocking new modes of intelligence beyond reactive prompt completion. Next, I will share work on enabling agents to mirror human behavior and reason about human preferences, paving the way toward more socially aligned and intuitive AI. Finally, I will preview early efforts on advisor agents—meta-agents designed to guide and critique other agents as they work—laying the foundation for self-improving AI systems. Together, these capabilities move us closer to a future where LLMs operate as adaptive, collaborative, and trustworthy digital agents.

Biography

Pascal Hitzler is University Distinguished Professor and endowed Lloyd T. Smith Creativity in Engineering Chair at the Department of Computer Science at Kansas State University, Director of the Center for Artificial Intelligence and Data Science (CAIDS), and Director (Research) of the Institute for Digital Agriculture and Advanced Analytics (ID3A). He obtained his PhD in Mathematics from the University of Ireland, University College Cork, in 2001. His research record lists over 400 publications in diverse areas such as neurosymbolic artificial intelligence, semantic web, knowledge graphs, knowledge representation and reasoning, denotational semantics, and set-theoretic topology. He was founding Editor-in-chief of the Semantic Web journal, the leading journal in the field, and is founding Editor-in-chief of the new Neurosymbolic Artificial Intelligence journal, and of the IOS Press book series Studies on the Semantic Web. He is co-author of the W3C Recommendation OWL 2 Primer, and of the book Foundations of Semantic Web Technologies by CRC Press, 2010, which was named as one out of seven Outstanding Academic Titles 2010 in Information and Computer Science by the American Library Association's Choice Magazine, and has translations into German and Chinese. He is a founding steering committee member of the Neural-Symbolic Learning and Reasoning Association and the Association for Ontology Design and Patterns, and he frequently acts as conference chair in various functions. For more information about him, see http://www.pascal-hitzler.de.

Accelerating Knowledge Graph Based Data Management with Large Language Models

Semantic Web Technologies, including Knowledge Graphs and Ontologies, are not only powerful tools for data integration and management, but are also increasingly used as fact databases to improve Large Language Models. However, producing high-quality ontologies and knowledge graphs is an expensive process that has proven to be hard to automate.But the recent rise of Large Language Models with increasing capabilities such as solid coding assistance carries the promise that it may help to significantly accelerate knowledge graph and ontology engineering through at least semi-automatic processes.We will discuss the current opportunities arising from this development, and elucidate the state of the art and some promising directions.

Biography

Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science,  and Association for Computing machinery, and a recent recipient of the AAAI Patrick H. Winston Outstanding Educator award. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence,  the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI. Kambhampati’s research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets. He can be followed on Twitter @rao2z.

(How) Do LLMs Reason?

Large Language Models, auto-regressively trained on the digital footprints of humanity, have shown impressive abilities in generating coherent text completions for a vast variety of prompts. While they excelled from the beginning in producing completions in appropriate style, factuality and reasoning/planning abilities remained their Achilles heel (premature claims notwithstanding). More recently a breed of approaches dubbed “reasoning models” (LRMs). These approaches leverage  two broad and largely independent ideas: (i) test-time inference – which involves getting the base LLMs do more work than simply providing the most likely completion, including using them in generate and test approaches such as LLM-Modulo (that pair LLM generation with a bank of verifiers) and (ii) post-training methods–which  go beyond simple auto-regressive training on web corpora by collecting, filtering and training on derivational traces (that are often anthropomorphically referred to as “chains of thought” and “reasoning traces”), and modifying the base LLM with it using supervised finetuning or reinforcement learning methods. Their success on benchmarks notwithstanding, there are significant questions and misunderstandings about these methods–including whether they can provide correctness guarantees, whether they do adaptive computation, whether the intermediate tokens they generate can be viewed as reasoning traces in any meaningful sense, and whether they are costly Rube Goldberg reasoning machines that incrementally compile verifier signal into the generator or truly the start of a golden era of general purpose System 1+2 AI systems.  Drawing from our  ongoing work in planning, I will  present a broad perspective on these approaches and their promise and limitations.

Biography

Ben Zhao is Neubauer Professor of Computer Science at University of Chicago. He completed his Ph.D. at U.C. Berkeley (2004), and B.S. from Yale (1997). He is a Fellow of the ACM, TED/AI speaker, and a recipient of the NSF CAREER award, MIT Technology Review's TR-35 Award (Young Innovators Under 35), TIME100/AI (100 Influential people in AI), USENIX Internet Defense Prize, Concept Art Association's Community Impact Award, ComputerWorld Magazine's Top 40 Technology Innovators award, IEEE ITC Early Career Award, and Faculty awards from Google, Amazon, and Facebook. His work has been covered by many media outlets including New York Times, CNN, NBC, ABC, BBC, MIT Tech Review, Wall Street Journal, Forbes, and New Scientist. His current research focuses on mitigating the harms of misused AI. He has published over 190 articles in areas of security and privacy, machine learning, networking, and HCI. He served as TPC (co-)chair for the World Wide Web conference (WWW 2016) and ACM Internet Measurement Conference (IMC 2018), and on the Steering Committee of IEEE Conference on Secure and Trustworthy Machine Learning (SaTML).

Mitigating Harms of Generative AI using Adversarial Tools

Generative AI models promise tremendous benefits in the name of novel capabilities and innovation. Yet it also brings with it numerous harms often ignored by developers and proponents. Beyond issues of training on creative work product without consent, there are numerous issues of misuse and abuse, many of which produce concrete and tangible harms on some of the most vulnerable and creative segments of our population.  In this talk, I will discuss some of the technical solutions we have developed in our lab to combat and mitigate these harms, most notably Glaze and Nightshade, adversarial ML tools that disrupt model training and fine tuning. Since their releases in March 2023 and January 2024, Glaze and Nightshade have been downloaded more than 10 million times by artists world wide. Beyond the technical details of adversarial disruption of diffusion models, I will also discuss the evolution of our lab from an initial focus on AI security, to a mission focused on protecting and advocating for human creatives across industries, and our progress in recent years in the media, in legislature bodies, and in the courtroom. 

Biography

Stefano Ermon is an Associate Professor of Computer Science in the CS Department at Stanford University where he is affiliated with the Artificial Intelligence Laboratory. His research is centered on techniques for probabilistic modeling of data and generative models. He has won several awards, including multiple Best Paper Awards, a NSF Career Award, ONR and AFOSR Young Investigator Awards, Microsoft Research Fellowship, Sloan Fellowship, and the IJCAI Computers and Thought Award. Stefano earned his Ph.D. in Computer Science at Cornell University in 2015.

Diffusion Language Models

Diffusion models are at the core of many state-of-the-art generative AI systems for content such as images, videos, and audio. Most generative models for language, however, are instead based on autoregressive modeling (i.e. next token prediction). In this talk, we review recent advances in diffusion-based language modeling frameworks that achieve highly competitive performance with autoregressive models. These recent results challenge the longstanding dominance of autoregressive modeling and could pave the way for an alternative class of more efficient language models built from radically different principles.

Biography

Daehyun Kim is an executive vice president at Samsung research, leading the Global AI Center. Before joining Samsung, he has been at Google and Intel working on processor architecture and workload optimization for 15 years since he received the Ph.D. from Cornell University.

His research interests include on-device AI, machine learning accelerators, mobile SoC architecture, high performance computing, intelligent memory systems, and future applications. Main approach is workload-driven architecture design. He has done path-finding research for Samsung’s on-device AI consumer devices, Google’s Android & ChromeOS products, and Intel’s Xeon Phi coprocessors based on workload analysis and optimization of AI services, mobile applications, HPC scientific computing, medical imaging, and physical simulation.

Biography

Kyungwhoon (Paul) Cheun, is Corporate President & Chief Technology Officer (CTO) of Device eXperience (DX) Division and Head of Samsung Research at Samsung Electronics. As CTO, he is driving the company’s technological innovations and strategy to align them with Samsung’s corporate goals. He also heads up Samsung Research, the advanced Research & Development hub of the DX Division at Samsung Electronics.

Samsung Research leads the development of future technologies and identifies new growth opportunities for the company through efforts spanning 15 R&D and AI centers located globally. Samsung Research's main focus is Artificial Intelligence, Next Generation Communications, Robotics, Security & Privacy, etc. He leads Samsung Research in constantly driving innovation across AI technologies and developing new research-powered products to provide better and more meaningful experiences to our valued customers.

Biography

Joseph Gonzalez is a Professor of Computer Science at UC Berkeley, co-director of the RISELab and Sky Computing Lab, and a core member of the Berkeley AI Research (BAIR) group. His research focuses on large language models, agentic systems, and scalable machine learning infrastructure. He is known for his contributions to projects such as Chatbot Arena, Gorilla, MemGPT, and vLLM, which have significantly advanced the field of open LLM research. Joseph is also the co-founder and Head of AI at RunLLM.com. His broader work spans data systems, robotics, computer vision, graph analytics, and autonomous systems. Prior to Berkeley, he co-founded Turi Inc. (formerly GraphLab) based on his Ph.D. work. His research has been recognized with the Okawa Research Grant, NSF CAREER Award, VLDB Test-of-Time Award, and NSF Expeditions in Computing Award.

Agentic Intelligence의 발전

대규모 언어 모델(LLMs)은 수동적인 완성 엔진에서 벗어나 장기간의 상호작용을 통해 복잡한 작업을 수행할 수 있는 지속적인 에이전트로 진화하고 있습니다. 이 강연에서는 LLM 기반 시스템의 에이전트적 능력을 향상시키기 위한 제 연구실의 최근 노력을 소개하겠습니다. 저는 상호작용 사이에 에이전트가 추론, 학습, 계획을 할 수 있게 해주는 ‘수면 시간 계산(sleep-time compute)’이라는 패러다임을 소개할 것입니다. 이는 반응형 프롬프트 완성을 넘어서는 새로운 지능 모드를 가능하게 합니다. 다음으로, 에이전트가 인간의 행동을 반영하고 인간의 선호도를 고려할 수 있도록 하는 연구를 공유할 예정입니다. 이를 통해 더욱 사회적으로 정렬되고 직관적인 AI를 향한 길을 열 것입니다. 마지막으로, 다른 에이전트를 안내하고 비판하는 메타 에이전트인 ‘조언자 에이전트(advisor agents)’에 대한 초기 노력을 미리 보여드리겠습니다. 이는 자기 개선형 AI 시스템의 기반을 마련할 것입니다. 이러한 능력들은 LLM이 적응적이고 협업적이며 신뢰할 수 있는 디지털 에이전트로 작동하는 미래를 향해 우리를 더 가까이 데려갈 것입니다.

Biography

Pascal Hitzler is University Distinguished Professor and endowed Lloyd T. Smith Creativity in Engineering Chair at the Department of Computer Science at Kansas State University, Director of the Center for Artificial Intelligence and Data Science (CAIDS), and Director (Research) of the Institute for Digital Agriculture and Advanced Analytics (ID3A). He obtained his PhD in Mathematics from the University of Ireland, University College Cork, in 2001. His research record lists over 400 publications in diverse areas such as neurosymbolic artificial intelligence, semantic web, knowledge graphs, knowledge representation and reasoning, denotational semantics, and set-theoretic topology. He was founding Editor-in-chief of the Semantic Web journal, the leading journal in the field, and is founding Editor-in-chief of the new Neurosymbolic Artificial Intelligence journal, and of the IOS Press book series Studies on the Semantic Web. He is co-author of the W3C Recommendation OWL 2 Primer, and of the book Foundations of Semantic Web Technologies by CRC Press, 2010, which was named as one out of seven Outstanding Academic Titles 2010 in Information and Computer Science by the American Library Association's Choice Magazine, and has translations into German and Chinese. He is a founding steering committee member of the Neural-Symbolic Learning and Reasoning Association and the Association for Ontology Design and Patterns, and he frequently acts as conference chair in various functions. For more information about him, see http://www.pascal-hitzler.de.

대규모 언어 모델을 활용한 지식 그래프 기반 데이터 관리 가속화

지식 그래프와 온톨로지를 포함한 시맨틱 웹 기술은 데이터 통합 및 관리를 위한 강력한 도구일 뿐만 아니라 대규모 언어 모델의 성능을 향상시키기 위한 사실 데이터베이스로도 점점 더 많이 사용되고 있습니다. 그러나 고품질의 온톨로지와 지식 그래프를 생성하는 것은 자동화가 어려운 비용이 많이 드는 과정으로 입증되었습니다.그러나 최근 코딩 지원과 같은 다양한 기능이 강화된 대규모 언어 모델의 등장은 적어도 반자동 프로세스를 통해 지식 그래프와 온톨로지 엔지니어링을 크게 가속화할 수 있다는 가능성을 보여주고 있습니다. 우리는 이러한 발전에서 비롯된 현재의 기회를 논의하고, 최신 기술과 몇 가지 유망한 방향을 설명할 것입니다.

Biography

Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science,  and Association for Computing machinery, and a recent recipient of the AAAI Patrick H. Winston Outstanding Educator award. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence,  the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI. Kambhampati’s research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets. He can be followed on Twitter @rao2z.

(How) Do LLMs Reason?

인간성의 디지털 발자취를 기반으로 자동으로 훈련된 대규모 언어 모델은 다양한 프롬프트에 대해 일관된 텍스트 완성을 생성하는 놀라운 능력을 보여주었습니다. 초기부터 적절한 스타일과 사실성을 갖춘 완성을 생성하는 데 뛰어났지만, 논리적 추론 및 계획 능력은 그들의 아킬레스건이었습니다(조기 주장을 제외하고). 최근에는 "추론 모델(LRM)"이라고 불리는 접근 방식이 등장했습니다. 이러한 접근 방식은 두 가지 광범위하고 독립적인 아이디어를 활용합니다: (i) 테스트 시점 추론 – 이는 기본 LLM이 단순히 가장 가능성 높은 완성을 제공하는 것보다 더 많은 작업을 수행하도록 하는 것으로, LLM-Modulo와 같은 생성 및 테스트 접근 방식(예: LLM 생성과 검증기 은행을 결합)을 포함하며, (ii) 사후 훈련 방법 – 이는 웹 코퍼스에 대한 단순한 자동 회귀 훈련을 넘어, 연역적 흔적(종종 의인화되어 "사고의 연쇄" 또는 "추론 흔적"이라고 함)을 수집, 필터링 및 훈련하고 이를 사용하여 지도 학습 또는 강화 학습 방법을 통해 기본 LLM을 수정합니다. 이러한 방법들은 벤치마크에서의 성공을 거두었지만, 그들이 정확성 보장을 제공할 수 있는지, 적응형 계산을 수행하는지, 중간 토큰이 의미 있는 방식으로 추론 흔적으로 간주될 수 있는지, 그리고 이들이 검증기 신호를 생성기에 점진적으로 컴파일하는 복잡한 루브 골드버그 기계인지 아니면 진정으로 범용 System 1+2 AI 시스템의 황금 시대의 시작인지에 대한 중요한 질문과 오해가 있습니다. 우리의 계획 작업 진행 상황을 바탕으로, 이러한 접근 방식과 그 가능성 및 한계에 대한 광범위한 관점을 제시하겠습니다.

Biography

Ben Zhao is Neubauer Professor of Computer Science at University of Chicago. He completed his Ph.D. at U.C. Berkeley (2004), and B.S. from Yale (1997). He is a Fellow of the ACM, TED/AI speaker, and a recipient of the NSF CAREER award, MIT Technology Review's TR-35 Award (Young Innovators Under 35), TIME100/AI (100 Influential people in AI), USENIX Internet Defense Prize, Concept Art Association's Community Impact Award, ComputerWorld Magazine's Top 40 Technology Innovators award, IEEE ITC Early Career Award, and Faculty awards from Google, Amazon, and Facebook. His work has been covered by many media outlets including New York Times, CNN, NBC, ABC, BBC, MIT Tech Review, Wall Street Journal, Forbes, and New Scientist. His current research focuses on mitigating the harms of misused AI. He has published over 190 articles in areas of security and privacy, machine learning, networking, and HCI. He served as TPC (co-)chair for the World Wide Web conference (WWW 2016) and ACM Internet Measurement Conference (IMC 2018), and on the Steering Committee of IEEE Conference on Secure and Trustworthy Machine Learning (SaTML).

생성형 AI의 위해를 완화하는 적대적 도구 활용

생성형 AI 모델은 새로운 기능과 혁신을 통해 엄청난 이점을 약속합니다. 그러나 개발자와 지지자들이 종종 간과하는 수많은 위해도 함께 가져옵니다. 창작물의 동의 없는 학습 문제를 넘어서, 오용과 남용에 관한 다양한 문제들이 가장 취약하고 창의적인 인구 집단에게 구체적이고 실질적인 위해를 초래합니다. 이 강연에서는 우리가 개발한 기술적 솔루션 중 일부를 소개하며, 이러한 위해를 방지하고 완화하기 위한 노력을 다룰 것입니다. 특히, 모델 학습과 미세 조정을 방해하는 적대적 ML 도구인 Glaze와 Nightshade를 중점적으로 설명할 예정입니다. 2023년 3월과 2024년 1월에 출시된 Glaze와 Nightshade는 전 세계 예술가들에 의해 1,000만 회 이상 다운로드되었습니다. 확산 모델의 적대적 방해에 관한 기술적 세부 사항을 넘어, AI 보안에 초점을 맞춘 초기 연구실에서부터 다양한 산업 분야의 인간 창작자들을 보호하고 옹호하는 사명으로 진화한 과정, 그리고 최근 몇 년간의 미디어, 입법 기관, 법정에서의 진전 상황에 대해서도 논의할 예정입니다."

Biography

Stefano Ermon is an Associate Professor of Computer Science in the CS Department at Stanford University where he is affiliated with the Artificial Intelligence Laboratory. His research is centered on techniques for probabilistic modeling of data and generative models. He has won several awards, including multiple Best Paper Awards, a NSF Career Award, ONR and AFOSR Young Investigator Awards, Microsoft Research Fellowship, Sloan Fellowship, and the IJCAI Computers and Thought Award. Stefano earned his Ph.D. in Computer Science at Cornell University in 2015.

확산 언어 모델

확산(diffusion) 모델은 이미지, 비디오, 오디오와 같은 콘텐츠 생성을 위한 최첨단 생성형 AI 시스템의 핵심을 이루고 있습니다. 그러나 언어를 위한 대부분의 생성 모델은 자기회귀적 모델링(즉, 다음 토큰 예측)에 기반하고 있습니다. 이번 강연에서는 자기회귀적 모델과 경쟁할 만한 성능을 달성한 확산 기반 언어 모델링 프레임워크의 최신 발전을 살펴봅니다. 이러한 최근 성과들은 자기회귀적 모델링의 오랜 지배력을 흔들고, 근본적으로 다른 원리에 기반한 보다 효율적인 언어 모델의 새로운 방향을 제시할 수 있습니다.

Biography

Daehyun Kim is an executive vice president at Samsung research, leading the Global AI Center. Before joining Samsung, he has been at Google and Intel working on processor architecture and workload optimization for 15 years since he received the Ph.D. from Cornell University.

His research interests include on-device AI, machine learning accelerators, mobile SoC architecture, high performance computing, intelligent memory systems, and future applications. Main approach is workload-driven architecture design. He has done path-finding research for Samsung’s on-device AI consumer devices, Google’s Android & ChromeOS products, and Intel’s Xeon Phi coprocessors based on workload analysis and optimization of AI services, mobile applications, HPC scientific computing, medical imaging, and physical simulation.