Abstract: Reward hacking, where a reasoning model exploits loopholes in a reward function to achieve high rewards without solving the intended task, poses a significant threat. This behavior may be explicit, i.e. verbalized in the model's chain-of-thought (CoT), or implicit, where the CoT appears benign thus bypasses CoT monitors. To detect implicit reward hacking, we propose TRACE (Truncated Reasoning AUC Evaluation). Our key observation is that hacking occurs when exploiting the loophole is easier than solving the actual task. This means that the model is using less effort than required to achieve high reward. TRACE quantifies effort by measuring how early a model's reasoning becomes sufficient to obtain the reward. We progressively truncate a model's CoT at various lengths, force the model to answer, and estimate the expected reward at each cut-off. A hacking model, which takes a shortcut, will achieve a high expected reward with only a small fraction of its CoT, yielding a large area under the reward-vs-length curve. TRACE achieves over 65% gains over our strongest 72B CoT monitor in math reasoning, and over 30% gains over a 32B monitor in coding. We further show that TRACE can discover unknown loopholes during training. Overall, TRACE offers a scalable unsupervised approach for oversight where current monitoring methods prove ineffective.

Speaker: Dikshya

Location: Old Computer Science Building - CS2311

You are cordially invited to attend the biweekly Brookhaven AI Mixer (BAM). BAM includes one short talk on AI research happening at BNL, followed by an open mixer over coffee and snacks for everyone to network and discuss all things AI. The first half hour will consist of presentations that will be available via ZOOM, and the second half hour will be for in person only networking.

Abstract: Two-dimensional (2D) materials such as graphene, hBN, and TMDs offer atomically sharp interfaces and unprecedented tunability when vertically assembled into van der Waals heterostructures. These stacks have enabled discoveries ranging from moiré superconductivity and correlated insulators to quantum emitters and next-generation nanoelectronic devices. Yet constructing high-quality heterostructures remains largely artisanal: researchers manually identify exfoliated flakes, align a polymer stamp by eye, and finely adjust temperature and contact geometry through tacit skill. This manual workflow is difficult to reproduce, scales poorly, and prevents systematic exploration of the enormous combinatorial space of materials, twist angles, and interfacial conditions. AutoLab is an autonomous platform that translates this tacit human expertise into programmable, feedback-driven control. Instead of pressing flakes with predefined trajectories, AutoLab uses machine vision to detect polymer-wafer contact, dynamically regulates contact evolution through closed-loop actuation and temperature control, and captures high-quality flakes with the cleanliness and precision of expert manual fabrication. The system integrates perception, decision making, and motion planning into a single robotic framework, enabling reproducible stacking, wafer-level coverage, and accelerated discovery. Beyond 2D materials, AutoLab illustrates a broader paradigm for AI-native scientific automation: codifying human experimental reasoning into algorithms that interrogate data in real time, adaptively adjust instrumentation, and generate scalable, high-fidelity datasets. Such platforms could generalize to diverse research domains--quantum device fabrication, optical alignment, surface science, autonomous microscopy, and other workflows where expert intuition currently limits throughput and reproducibility. By bridging artisanal manipulation and robotic autonomy, AutoLab points toward a future where scientific discovery is accelerated by machines that not only execute instructions, but learn, respond, and collaborate with human scientists.

Biography: Dr. Yutao Li is a research associate from Department of Condensed Matter Physics and Material Science, Brookhaven National Laboratory. He has 8 years of experience in 2D material sample fabrication, and investigation in their electronic transport, optical and mechanical properties.

Join us every other Tuesday at noon in CDSD's Training Room (building 725, 2nd floor) to learn about interesting AI methods and applications, engage with potential collaborators, prepare for pending FASST funding calls, and build a community of AI for Science at BNL.

Location: CDS, Bldg. 725, Training Room

Join ZoomGov Meeting: https://bnl.zoomgov.com/j/1604383624?pwd=ffQ5cUPNxTI7nzClKQO6cnsNbhF9Vf.1

Meeting ID: 160 438 3624
Passcode: 558449

Abstract:

Recent advances in deep learning have significantly enhanced the capabilities of Natural Language Processing (NLP) and Vision-Language Models (VLMs). However, these advancements come with increased vulnerabilities, notably through backdoor attacks that pose severe security threats. This thesis addresses two critical dimensions of Trustworthy AI and Efficient Multimodal Representation Learning: (1) security through analyzing, detecting, and designing backdoor attacks in NLP and VLMs, and (2) efficiency through advanced multimodal representation methods tailored for clinical and medical imaging applications.

In the first dimension, we explore the internal mechanisms exploited by backdoor attacks, identifying the distinctive phenomenon of attention focus drifting in compromised transformer models, where trigger tokens consistently hijack attention. Leveraging these insights, we propose robust detection frameworks, including the attention-based Trojan detector (AttenTD) and a task-agnostic logit-based detection method (TABDet), achieving effective identification of backdoored NLP models across diverse tasks. We further introduce novel backdoor attack methodologies: the Trojan Attention Loss (TAL), enhancing attack efficiency and stealth through direct attention manipulation, and BadCLM, demonstrating critical vulnerabilities in clinical decision-support systems by effectively compromising clinical language models.

Extending our security exploration to multimodal settings, we investigate backdoor attacks on Vision-Language Models (VLMs), particularly in complex image-to-text generation tasks, proposing innovative techniques (TrojVLM, VLOOD) capable of embedding backdoors without direct access to original training data, thus showcasing practical risks in real-world scenarios.

In the second dimension, we address efficiency and interpretability challenges in clinical and pathology applications. We introduce TCP-LLaVA, the first multimodal large language model (MLLM) designed explicitly for Whole Slide Image (WSI) Visual Question Answering (VQA). Utilizing a novel token compression mechanism inspired by transformer-based models, TCP-LLaVA substantially reduces computational resource consumption while maintaining superior VQA performance across multiple tumor subtypes. Additionally, we present a multimodal transformer model integrating structured Electronic Health Records (EHR) with clinical notes, demonstrating enhanced predictive accuracy and interpretability for in-hospital mortality prediction through integrated gradient-based interpretability methods.

Together, these contributions present a comprehensive approach to ensuring AI models are not only secure against malicious manipulation but also efficient and interpretable for critical clinical applications, underscoring the essential need for trustworthy and effective AI systems.

Speaker: Weimin Lyu

Zoom: https://stonybrook.zoom.us/j/2392326575?pwd=SVQ2VkFXTnZZYmJUMXgvTXBuZWM3UT09

Meeting ID: 239 232 6575
Passcode: 436192
Come learn of the exciting research being done across so many fields using AI! The recipients of AI3's seed awards will present their work in our showcase on November 17, 2025 and we would love to see you there!

The schedule is listed below.

Location: New Computer Science Room 120

Session 1 - 10:30 AM to 11:45

Kevin Reed, PI, Introducing the AI Techniques in Assessing the Future Changes of Extreme Precipitation and Associated Flood Risks
Co-PIs: Tangnyu Song, Ishrat Dollan
Consultant: Jayesh Rathi

Ruwen Qin, PI, AI-Assisted Analysis of Materials in Recycling Streams
Consultant: Vismay Vora

Giuseppe Gazzola, PI, Using AI to Investigate National Literatures: Italy, France, Spain 1733- 1794
Consultant: Jayesh Rathi

Joseph Lemelin, PI, IAE2^3: AI Ecologies
Co-PIs: Katherine Johnston, Aruna Balasubramanian, Matthew Salzano

Niranjan Balasubramanian, Co-PI, Molecular Foundations for Sustainability: Data Analytics for Sustainable Cellulose Scaffolding Modifications to Remediate Diverse Water Contamination Challenges
PI: Benjamin Hsiao, Co-PI: I. V. Ramakrishnan

Owen Rambow, PI,Achieving Common Ground Through Language and Vision in Mixed-Initiative Human-Machine Communication Via zoom
Co-PI Susan Brennan

Session 2 - 12:30 PM to 1:45

Jack McSweeney, PI, Developing Machine Learning Approaches to Classify Internal Waves
Consultant: Vismay Vora

Eric Josephs, PI, Learning Design Rules to Personalize Precision CRISPR Gene Therapies with Interpretable AI
Consultant: Deboparna Banerjee

Shyam Sharma, PI, Fostering Writing-to-Learn Skills through Critical AI Literacy: A Faculty Development and Student Support Program
Co-PIs: Rose Tirotta-Esposito, Christine Fena

Ritwik Banerjee, PI, A Pragmatic Approach to AI for Digital Media Integrity: Combating Complex Misinformation Through Fallacies and Propaganda
Co-PI: Ruobing Li

Ziyu Shu, Co-PI, Novel Clinical Applications of Deep Image Prior-based CT Image Reconstruction
PI: Xin Qian, Co-PIs: Tiezhi Zhang, Zhaozheng Yin

Prateek Prasanna, Co-PI, An Artificial Intelligence-Driven Clinical Decision Support Tool for the Management of Abdominal Aortic Aneurysm
PI: Apostolos Tassiopoulos, Co-PI's: Mary Saltz, Janos Hajagos, Tahsin Kurc



Presenters will give a 5-minute talk with 2 minutes for Q & A.
What AI tools are available to help with the scholarly research process? Are they helpful? What do they do and is it worth the time and energy to try them out? Join librarian Christine Fena to explore and compare established and emerging AI research tools such as Elicit, Scite, Consensus, and Undermind. The online workshop will provide a starting point to understanding what these tools are, the basics of how they work, and how AI research assistants might bring changes to your search process in the future. All are welcome!



Register here via Zoom.

Learn how to summarize docs with AI, output a PowerPoint from AI, & Create professional visuals

Unlock greater efficiency and impact in your university role with AI productivity tools. This workshop is your introduction to a few ways that I have found to make our daily tasks more efficient. Discover how easily you can create presentations (that outputs to a PowerPoint format), summarize content using AI, and get information from images. These AI tool tips are invaluable resources designed to streamline your work processes. Start working smarter today!

In this session, you will

  1. Summarize docs with AI
  2. Output a PowerPoint from AI
  3. Gather information from visuals

Register here.
Abstract: Sea ice is crucial to Earth's climate, Arctic communities, and ecosystems, yet climate change is driving significant losses, threatening polar stability. Quantifying the long-term impacts of a declining sea ice cover requires tools which improve climate-timescale prediction and bring new understanding of climate interactions. In this talk, I discuss how meeting this challenge requires a multi-disciplinary approach. Climate models, while essential, suffer from systematic biases due to missing or inaccurate physics, leading to uncertainty in future projections. I show how data assimilation (DA) offers a statistical framework for integrating satellite observations with climate models to quantify systematic sea ice model errors. Using convolutional neural networks (CNNs), we can learn these errors based on the model's atmospheric, oceanic, and sea ice conditions--what I term a state-dependent representation of the error. This approach enables real-time corrections to subsequent model simulations, which systematically reduces global sea ice biases. I highlight key successes and challenges in developing this hybrid ML+climate modeling framework, including transfer learning to enhance online generalization of ML models, and new methods for integrating Python-based ML frameworks with Fortran climate model code. Finally, I introduce GPSat, a scalable Gaussian process-based tool for reconstructing complete sea ice fields from sparse satellite altimetry data. Together, the DA+ML framework and GPSat offer future opportunities for improving targeted model physics errors for more robust climate simulation.


IACS Seminar Speaker: William Gregory, Princeton University

Location: IACS Seminar Room

Abstract: Traditional questionnaires remain the primary method for assessing psychological outcomes and beliefs, capturing individuals' and populations' inner states. This dissertation presents an alternative computational method that overcomes key limitations in current mental health monitoring, particularly in spatiotemporal resolution, responses to major events, and automatic belief identification. By analyzing ∼1 billion Tweets from 2 million geo-located users, we created a big data pipeline for estimating depression and anxiety at the county-week level. These Language-Based Mental Health Assessments (LBMHA) demonstrated higher reliability and validity than traditional survey measures. Our approach effectively captured mental health trends and highlighted significant increases in mental illness following major events. Using the LBMHA pipeline, we conducted quasi-experiments, research designs that simulate randomized control trials, to generate explanations for mental health changes due to COVID-19 incidence/death. Utilizing these time-series analyses, we conducted discontinuity forecasting for community-specific anxiety shifts using statistical learning via ensemble and contextual models. To likewise investigate individual internal states, we created a novel task and annotated dataset for self belief language identification. Our fine-tuned language model for self-belief classification, despite its relatively small scale, outperformed GPT-4o. The self belief topics identified by our model successfully predicted depression, anxiety, and stress, offering insights into the relationship between self-conceptualization and mental health. The adoption of scalable language-based assessments with modern distributed computation presents a promising avenue for advancing community and individual mental health research.

Speaker: Siddharth Mangalik

https://stonybrook.zoom.us/j/91251321639?pwd=faggV5jZ7ByFDCFmnLXD3HiYxjQ1Eb.1&jst=2
Abstract: Autonomous systems, whether on Earth or in space, rely on 3D perception to understand and interact with the world around them. Yet traditional techniques for 3D understanding often depend on human designed features, fixed sensors, and conventional imaging modalities. This constrained approach can limit every stage of perception, from sensing to interpretation to decision making.
In this talk, we'll explore an alternative paradigm for imaging: physically based neural representations for 3D scenes and 3D sensing systems. We will discuss how recent advances in large scale learned representations can be used to jointly optimize both 3D scene models and the design of sensing systems for 3D capture, with the goal of enabling task specific perception systems.
Unlike modern AI models trained on internet scale datasets, these specialized 3D representations typically operate in data sparse regimes and therefore require a different kind of prior. We'll examine how grounding these learned representations in the physics of light transport can improve our understanding of scene structure, and inform imaging system design even with limited data. By connecting physical insights with learned representations, we'll highlight new possibilities for robust, efficient, and adaptive perception in challenging environments.

Speaker: Nikhil Behari is a graduate student in the Camera Culture group at the MIT Media Lab, advised by Professor Ramesh Raskar. His research interests include computational imaging, 3D scene understanding, and multi-agent decision-making under uncertainty, with a focus on automating imaging system design for 3D perception in human and planetary health. His research is supported by the NASA Space Technology Graduate Research Fellowship. He received his bachelor's in Computer Science and Statistics from Harvard University in 2022.
Communication-Efficient Heterogeneity-Aware Machine Learning System and Architecture by Xuehai Qian

ABSTRACT: The key success of deep learning is the increasing size of models that can achieve high accuracy. At the same time, it is difficult to train the complex models with large data sets. Therefore, it is crucial to accelerate training with distributed systems and architectures, where communication and heterogeneity are two key challenges. In this talk, I will present two heterogeneity-aware decentralized training protocols without communication bottleneck. Specifically, Hop supports arbitrary iteration gap between workers by novel queue-based synchronization which can tolerate heterogeneity with system techniques. Prague uses randomized communication to tolerate heterogeneity with a new training algorithm based on partial reduce -- an efficient communication primitive. If time permits, I will present the systematic tensor partitioning for training on heterogeneous accelerator arrays (e.g., GPU/TPU). We believe that our principled approaches are crucial for achieving high-performance and efficient distributed training.

BIO: Xuehai Qian is an assistant professor at University of Southern California. His research interests include domain-specific systems and architectures, performance tuning and resource management of cloud systems and parallel computer architectures. He received his PhD from the University of Illinois Urbana Champaign and was a postdoc at UC Berkeley. He is the recipient of W.J Poppelbaum Memorial Award at UIUC, NSF CRII and CAREER Award, and the inaugural ACSIC (American Chinese Scholar In Computing) Rising Star Award.