AI Innovation Institute

Abstract: This talk is about the two ends of LLM training: pre-training and in-deployment learning. I will present an approach to disentangle knowledge from skill in model pre-training. This brings about a new class of LLMs that externalize knowledge, with dramatically different characteristics from common LLMs along dimensions of scale, factuality, and updateability. On the other end, I will discuss two in-deployment learning methods. I will describe how in-context learning abilities extend beyond supervised settings, showing that LLMs display in-context reinforcement learning from rewards. Finally, if time allows, I will describe continual learning from implicit interaction signals, demonstrating that LLMs can retrospectively decode latent interaction cues by observing how humans respond to their outputs.

Bio: Yoav Artzi is an Associate Professor in the Department of Computer Science and Cornell Tech at Cornell University, a visiting faculty researcher at Google DeepMind, and arXiv's associate faculty director. His research focuses on language modeling and learning in interactive and situated scenarios. His work was acknowledged by awards and honorable mentions at ACL, EMNLP, NAACL, and IROS, as well as a TACL test-of-time award. Yoav holds a B.Sc. from Tel Aviv University and a Ph.D. from the University of Washington.

Location: NCS 120

Towards Saving Lives with Natural Language Processing Andrew Schwartz Dept. of Computer Science Stony Brook Analyzing language use patterns is proving to be a valuable and unique approach to understanding the psychological, social, and health factors of people. On the individual level, Facebook and Twitter have been found predictive of mental health, personality, demographics, and occupational class (among others). At the community or county-level, Twitter has been found predictive of flu and allergy outbreaks, life satisfaction, atherosclerotic heart disease mortality, health behavioral risk factors, excessive drinking, and HIV prevalence. While these techniques have shown robust links over a plethora of important aspects of human life, it is not clear whether any lives have been saved, at least directly, by such work. At their core, some barriers to improving health care and saving lives are likely not NLP or even AI problems, but others are perhaps technical in nature and suggest changing the way we model data. This seminar will have two parts: a presentation and a discussion. I will start by going over recent and on-going work toward predicting mental health outcomes --- depression, addiction relapse, future psychological distress --- from human language use patterns. Then, I will present an imperfect vision of a future where NLP helps to save lives and open the floor for discussion of technical barriers and whether such a vision is practical. Biography: Andrew Schwartz received his PhD in Computer Science from the University of Central Florida in 2011 with research on acquiring lexical semantic knowledge from the Web. He then joined the University of Pennsylvania where he was a Postdoctoral Research Fellow and later Visiting Assistant Professor in Computer & Information Science. He is Lead Research Scientist for the World Well-Being Project, a multidisciplinary group of Computer Scientists and Psychologists studying physical and psychological well-being based on language in social media.

University Libraries Presents: The Library AI Club is a welcoming space for students, faculty, and staff to explore AI in a supportive, low-pressure environment. Meeting every two weeks, the club features discussions, collaborative projects, guest speakers, and hands-on experiments. Join us to learn, share ideas, and engage with AI responsibly and creatively. We'd love to see you at an upcoming meeting! Location: Melville Library, Scholarly Communication Seminar Room

Join Stony Brook University's Center for Excellence in Learning and Teaching (CELT) for a bootcamp on how to use AI to enhance your teaching and courses. This event will demonstrate how ChatGPT, Microsoft Copilot, and other generative AI platforms can support you in crafting learning objectives, writing exam questions, composing rubrics, and designing course content such as lesson plans, in-class activities, instructional videos, and more.

Register here.

Abstract: Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within adversarial prompts. While most existing defenses attempt to mitigate the effects of adversarial prompts, they often prove inadequate as adversarial prompts can take arbitrary, adaptive forms. This paper introduces RobustKV, a novel jailbreak defense that takes a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for an adversarial prompt to be effective, its tokens must achieve sufficient `importance' (measured by attention scores), which consequently lowers the importance of tokens in the concealed harmful query. Therefore, by carefully evicting the KVs of low-ranked tokens, RobustKV minimizes the harmful query's presence in the KV cache, thus preventing the LLM from generating informative responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's performance on benign queries. Notably, RobustKV creates an interesting effectiveness-evasiveness dilemma for the adversary, leading to its robustness against adaptive attacks.

Speaker: Tanqiu Jiang

Where: NCS 220 and Zoom (https://stonybrook.zoom.us/j/6406956411)

An interactive session to discover how to create ALT text tags from images and create high-impact visuals, from identification to communicating ideas with images.

Discover how to use AI to create ALT text from images as well as identify objects in your environment, and build relatable visuals for high-impact presentations. Images communicate ideas as a way to understand concepts. AI-generated images have helped allow anyone to create these.

In this session, you will

1. Creating image ALT Tags
2. Transform ideas into images that are visually appealing
3. Identify objects from visuals

Register here.

Abstract: DeepSeek-R1-Zero has shown that reinforcement learning (RL) at scale can directly enhance the reasoning capabilities of LLMs without supervised fine-tuning. In this work, we critically examine R1-Zero-like training by analyzing its two core components: base models and RL. We investigate a wide range of base models, including DeepSeek-V3-Base, to understand how pretraining characteristics influence RL performance. Our analysis reveals that DeepSeek-V3-Base already exhibit ''Aha moment'', while Qwen2.5 base models demonstrate strong reasoning capabilities even without prompt templates, suggesting potential pretraining biases. Additionally, we identify an optimization bias in Group Relative Policy Optimization (GRPO), which artificially increases response length (especially for incorrect outputs) during training. To address this, we introduce Dr. GRPO, an unbiased optimization method that improves token efficiency while maintaining reasoning performance. Leveraging these insights, we present a minimalist R1-Zero recipe that achieves 43.3% accuracy on AIME 2024 with a 7B base model, establishing a new state-of-the-art.

Speaker: Md. Saqib Hasan

Location: CS2311

Imagine machines that can see beyond human limitations--drones locating hidden survivors, cameras predicting structural failures, or medical devices detecting tumors beneath the skin. Traditional vision systems are constrained by the boundaries of human perception, missing vast information present in light interactions. This talk explores the development of advanced vision systems that capture underutilized dimensions of light, model intricate light-scene interactions, and extract hidden 3D information--around corners, beneath surfaces, and at high speeds. By jointly developing novel imaging hardware, efficient rendering models, and physics-based learning algorithms, we aim to transcend conventional vision capabilities--unlocking critical applications in autonomous navigation, structural monitoring, and non-invasive medical imaging.

Speaker Bio:

Akshat Dave is a Postdoctoral Associate at MIT Media Lab in the Camera Culture group working with Prof. Ramesh Raskar. He received his Ph.D. from Rice University ECE Department in 2023 where he was advised by Prof. Ashok Veeraraghavan. His research lies at the intersection of applied optics, computer graphics, and computer vision. His research focuses on developing vision systems that go beyond human perception. His work has been recognized by Rice University's Best Thesis Award, OSA Best Paper Prize, and fellowships by Texas Instruments and Qualcomm.