AI Innovation Institute

Please join us for the next CSE 600 Seminar this Friday, October 11th, at 2:30pm in New Computer Science 120 given by Assistant Professor Mohammad Javad Amiri. Abstract: Today's distributed transaction processing systems must deal with untrustworthy environments where multiple mutually distrustful entities communicate with each other, and maintain data on untrusted infrastructure. Byzantine Fault-Tolerant (BFT) protocols have recently been extensively used by distributed transaction processing systems to establish consensus on the order of transactions. However, the proliferation of different BFT protocols has made it difficult to navigate the BFT landscape, let alone determine the protocol that best meets application needs. Moreover, as novel smart contracts, modern hardware, and new cloud platforms arise, future-proof distributed transaction processing systems need to be designed with full-stack adaptivity in mind. This talk presents our vision for a reinforcement learning (RL)-based distributed transaction processing system that adjusts effectively in real-time to changing fault scenarios and workloads.

Abstract: Molecular learning has become an emerging field of AI, driving breakthroughs in drug discovery, protein design, and materials design. For high-stakes scientific tasks, however, predictive accuracy alone is not sufficient: models must also be interpretable and trustworthy. Our work aims to study molecular learning under a unified explainability perspective across two major model families: Graph Neural Networks (GNNs) and Large Language Models (LLMs).

GNNs are natural choices for molecular graphs and achieve strong performance on many molecular tasks. To enhance explainability, many GNN explanation methods have been proposed and work well for 2D GNNs. However, 3D GNNs introduce two key challenges: producing chemically meaningful substructures and reducing fidelity loss caused by dense geometric graphs. To address these challenges, I present two methods. 3DGraphX decomposes dense 3D graphs into chemically meaningful 3D motifs, enabling compact explanations that align with chemical intuition. EDMA introduces an energy-based discrete mask approximation approach to reduce the discrepancy between the soft mask optimized during training and the hard mask used for explanation, improving explanation fidelity.

LLMs present different characteristics and challenges compared with GNNs. LLMs can provide a certain level of explanation through step-by-step reasoning, and their natural-language outputs are easy for humans to understand and interpret. However, because LLMs are trained for general-purpose tasks, their performance on scientific tasks often lags behind specialized GNNs. To improve performance, existing methods guide LLMs by providing suggestions through brief feedback, retrieval-augmented generation (RAG), or planner agents. However, these approaches face several limitations, such as vague guidance, introduced bias problems, and high computational cost. To fill the gap, I propose RL-Guider, a lightweight reinforcement-learning agent that converts evaluation feedback into input-specific guidance for molecular optimization. RL-Guider improves over time by accumulating historical experience and transfers efficiently across different LLMs while preserving interpretability.

Together, these efforts aim to provide explanations that are scientifically meaningful and faithful, while also preserving or improving performance on molecular tasks to better meet real scientific needs.

Speaker: Xufeng Liu

Location: New Computer Science-1-Room 115

Title: Cultural Biases, World Languages, and User Privacy in Large Language Models
Abstract: In this talk, I will highlight three key aspects of large language models: (1) cultural bias in LLMs and pre-training data, (2) decoding algorithm for low-resource languages, and (3) human-centered design for real-world applications.

The first part focuses on systematically assessing LLMs' favoritism towards Western culture. We take an entity-centric approach to measure the cultural biases among LLMs (e.g., GPT-4, Aya, and mT5) through natural prompts, story generation, sentiment analysis, and named entity tasks. One interesting finding is that a potential cause of cultural biases in LLMs is the extensive use and upsampling of Wikipedia data during the pre-training of almost all LLMs. The second part will introduce a constrained decoding algorithm that can facilitate the generation of high-quality synthetic training data for fine-grained prediction tasks (e.g., named entity recognition, event extraction). This approach outperforms GPT-4 on many non-English languages, particularly low-resource African languages. Lastly, I will showcase an LLM-powered privacy preservation tool designed to safeguard users against the disclosure of personal information. I will share findings from an HCI user study that involves real Reddit users utilizing our tool, which in turn informs our ongoing efforts to improve the design of AI models.
Bio:

Wei Xu is an Associate Professor in the College of Computing and Machine Learning Center at the Georgia Institute of Technology, where she is the director of the NLP X Lab. Her research interests are in natural language processing and machine learning, with a focus on Generative AI, robustness and fairness of large language models, multilingual LLMs, as well as AI for science, education, accessibility, and privacy research. She is a recipient of the NSF CAREER Award, Google Academic Research Award, CrowdFlower AI for Everyone Award, Best Paper Awards and Honorable Mentions at COLING'18, ACL'23, ACL'24. She also received research funds from DARPA and IARPA. She is currently an executive board member of NAACL. Join Zoom Meeting https://stonybrook.zoom.us/j/98855994362?pwd=F2qnpwL85fhCBHAEW9ZBpXihfwGHsj.1 (ID: 98855994362, passcode: 172797) Join by phone (US) +1 646-876-9923 (passcode: 172797) Joining instructions: https://www.google.com/url?q=https://applications.zoom.us/addon/invitation/detail?meetingUuid%3DuDJcUTvyQueZkCaUSAwFlg%253D%253D%26signature%3Da3d49e0f7f2e74e7130f7308c74bd85ba7b99587b98ba2e34238bb657ca51a09%26v%3D1&sa=D&source=calendar&usg=AOvVaw2jTn5cjfRG8vXU8KHHlU2Y Meeting host: H.Andrew.Schwartz@stonybrook.edu

Join Zoom Meeting:
https://stonybrook.zoom.us/j/98855994362?pwd=F2qnpwL85fhCBHAEW9ZBpXihfwGHsj.1

Abstract: Pre-trained diffusion and flow matching models have made visual generation remarkably powerful, enabling high-fidelity synthesis of images and videos from natural language prompts. However, their behavior is still largely dictated by the pre-training data distribution and likelihood objective, which do not directly encode downstream desiderata such as fine-grained semantic alignment, controllability, or realism. This gap motivates post-training: starting from a base generator and further optimizing it with additional supervision signals derived from human or reward model preferences.This work presents post-training for visual generative models through two complementary case studies. First, Hummingbird addresses the problem of fine-grained contextual alignment in image-text-to-image generation. We introduce a multimodal context evaluator that scores the consistency between rich contextual descriptions and generated images, capturing fine-grained alignment beyond global CLIP similarity. By directly backpropagating these differentiable rewards through the diffusion sampler, Hummingbird substantially improves semantic faithfulness while preserving high visual quality.
Second, PISCES tackles post-training for text-to-video generation, where alignment is inherently semantic-spatio-temporal. We show that naive VLM-based rewards suffer from distributional mismatch and token-level misalignment, leading to reward hacking and suboptimal optimization. PISCES introduces a bi-objective, Optimal Transport (OT)-aligned reward module: distributional OT using Neural Optimal Transport to align text and video embedding distributions, and discrete, partial OT over a spatio-temporal cost matrix to capture semantic alignment at the token level. These rewards are integrated into both direct backpropagation and GRPO-style optimization to post-train state-of-the-art text-to-video generators. Together, Hummingbird and PISCES provide a unified view of how carefully designed visual reward models, coupled with OT-based representation alignment, can reliably improve the downstream behavior of pre-trained image and video generators.

Speaker: Minh Quan Le

Location: NCS 220

Zoom: https://stonybrook.zoom.us/j/94798224254?pwd=CFraer25qnpORbJ14aAVHRwaSJOjJM.1

Learn how to use Stony Brook Google Gemini for uploading files. Using the AI to change tone and how long the text is with a few clicks.

In this session, you will

1. Understand Stony Brook Google Gemini
2. See the New Tone and Length features

Register here.

CSE 600 Seminar Series | Fall 2025



Abstract:

We often talk about AI as if it begins with a dataset and ends with an application. But behind every model lie two sets of actors who are rarely acknowledged in technical documentation: the workers who train AI systems and the researchers who try to make sense of them. This talk brings both groups into view.
Dr. Ben Zhang will offer an on-the-ground examination of the prevailing values and invisible labor that underpin commercial AI production and data production. Drawing on ethnographic research inside AI data annotation centers in China, he introduces the concept of precision labor to unpack the labor dimension of constructing, managing, and performing technical accuracy. This concept highlights the hidden and excessive labor required to reconcile the ambiguity and uncertainty involved in AI training. A precision labor lens challenges the legitimacy and sustainability of the relentless pursuit of technical accuracy, raising new questions about its consequences and implications.
On the other end of the pipeline, as LLMs become embedded in society, social scientists like Dr. Jieshu Wang is scrutinizing their potential biases while employing them as research tools. She will present her recent work auditing LLM responses across different contexts, revealing that LLMs exhibit varying levels of environmental awareness and disproportionately reward institutional prestige in peer-review simulations. She also demonstrates how LLMs can serve as useful tools in social-science pipelines, e.g., extracting location information, inferring demographics, parsing citations, mapping social networks, and analyzing occupational data.
By placing these two worlds side by side - the labor of training AI and the scholarly efforts to study it - we show why responsible AI should go beyond the deployment phase - emphasizing fairness audits, and model explainability. It requires reimaging the values, labor regimes, and social science practices that shape AI systems from annotation to analysis.


Bios:

Dr. Jieshu Wang is an interdisciplinary researcher studying the human and social dimensions of artificial intelligence (AI) and how people can thrive in an AI-integrated future. She combines computational methods with qualitative insights to trace technology trends and understand their broader societal impact. She earned her Ph.D. in Human and Social Dimensions of Science and Technology from Arizona State University, after earlier degrees in Civil Engineering, Economics, and Science and Technology Studies. She has also worked as a patent examiner, an editor at a popular science magazine, and co-founded Synced (机器之心), an AI-focused media company in China. Her research looks both backward and forward. Backward-looking, she examines how AI are created, who creates them, and who is missing from the process. Forward-looking, she studies how AI is transforming the way we live, connect, invent, work, and adapt, as well as how AI might help address challenges such as climate change and workforce transitions.
Dr. Ben Zhang is an Assistant Professor in the Department of Technology. His research explores the production and sociotechnical impacts of AI systems in critical areas such as work, health, and sustainability. Drawing from his background in Human-Computer Interaction (HCI), Human-Centered AI, and Science and Technology Studies (STS), he employs a life-cycle-centered approach to holistically examine the promises and harms of these systems and to inform the design of responsible AI infrastructures across their development, deployment, and governance. Ben received his Ph.D. in Information Science from the University of Michigan. Ben's work has been supported by competitive awards and fellowships, including the University of Michigan Rackham Predoctoral Fellowship and the Weizenbaum Fellowship. His research has appeared in premier computing venues, including ACM CHI, ACM CSCW, and AAAI ICWSM.

Location: NCS 120

What Does Learning Mean? presented by Jeffrey Heinz

ABSTRACT
When we develop learning algorithms, what computational problems are we solving? In this talk, I discuss different answers that have been proposed for this question, and discuss some of the consequences for machine learning and artificial intelligence. The main lessons I offer are that (1) feasible solutions to learning problems require careful consideration of a target class C of functions, (2) that such a class C cannot include all functions, or even all computable functions, and so many logically possible functions must be outside of C and (3) class C must have significant structure which the solutions take advantage of. These main ideas are motivated and illustrated from modeling language acquisition and the related problem of grammatical inference from example sequences belonging to formal languages.