Through a donation from IBM, Georgia Tech’s Center for Scientific Software Engineering (CSSE) and Open Source Program Office (OSPO) held a fellowship competition for students to work with mentors at CSSE and OSPO to create the tool. Master’s students Adrian Lozada and Di Wu were selected as fellows to work on the project.  

The Mellea Legal Requirements Checker (Mellea-LRC) can quickly identify and accurately verify case law citations, specifically checking for citations that are hallucinated by artificial intelligence (AI). One fellowship goal was to give lawyers and others in the legal field a tool to easily verify legal citations. Mella-LRC does this by using both LLMs and IBM’s Mellea open-source framework to verify and repair LLM prompts.  

The team tested the tool on 26 public court-filing documents containing falsified citations. Mellea-LRC can process a document in about eight minutes. The tool first extracts the case law citations from the legal document. It then verifies that each citation is authentic and supports the document's content.  

“IBM’s Mellea framework was the foundation we built it on. Mellea’s job is to make model calls predictable: you write the rules the output must satisfy as ordinary code, and the framework retries until LLM output meets them. Our job was to leverage the framework to build the legal application from end to end,” Wu said.  

An important aspect of the project was ensuring the tool’s code was open source. The team supported this goal by also creating and releasing a new test and validation dataset on Hugging Face – a popular open-source AI development community – for future benchmarking studies. 

“Agentic AI is growing quickly, and much of the growth happens in open-source projects, where anyone can read, use, and improve the work,” Lozada said.  

OSPO Director and Principal Research Scientist Jeff Young noted that IBM’s Mellea framework was critical to this work in that its “instruct, verify, repair” paradigm allows for the creation of reliable LLM loops as well as the use of small language models (SLMs), such as IBM’s Granite, which can be run on a laptop. 

“This project demonstrates the power of open-source models like IBM Granite as well as the importance of using novel software engineering and algorithmic techniques from the Mellea project to improve the reliability of our existing LLMs and SLMs.  

“Mellea-LRC provides a compelling example of AI’s potential to evolve from a source of error in the legal profession into a trusted ally by helping lawyers work faster and cite precedents more accurately,” Young said.  

Unlike traditional AI backdoors, which manipulate a model's learned parameters, architectural backdoors hide malicious instructions directly in the code that defines how the model operates.

However, everyday image-processing operations, such as cropping, adjusting contrast, or adding noise, can change a trigger pattern and prevent an attack from working, according to new research from Georgia Tech.

The findings challenge assumptions about how practical these attacks are. Although previous research showed that architectural backdoors could work under controlled conditions, the Georgia Tech team found that existing approaches struggle to remain effective and difficult to detect when AI systems process images under realistic conditions.

“The biggest misconception is that demonstrating an architectural backdoor under controlled conditions means it will actually work in the real world,” said David Oygenblik, Ph.D. student and lead author of the study.

The study details the development of FAULT, a framework for evaluating architectural backdoors under realistic conditions.

The findings reveal a tradeoff. Attackers can make a backdoor more sensitive, so it recognizes its trigger after an image has been changed. But that also increases the chance that ordinary images will accidentally activate the backdoor, potentially hurting the AI system’s performance.

“For the architectural backdoors we study, this tradeoff is fundamental,” Oygenblik said.

The team found that this problem persisted across five methods of implementing architectural backdoors. Even when attackers knew how the system would process images, they struggled to make their attacks both reliable and difficult to detect.

The research did not test attacks against real vehicles or other production systems, nor did it demonstrate outcomes such as crashes or ransomware. Instead, it examined whether existing architectural backdoors could reliably activate under different image-processing conditions.

However, the findings highlight a potential risk in the AI software supply chain. If an attacker inserted malicious code into an AI model used in a real system, the backdoor could potentially cause incorrect predictions when its trigger was detected. The consequences would depend on how the system used the model’s output.

Because architectural backdoors hide in code, examining a model’s learned parameters or retraining it may not be enough to uncover or remove the threat.

Oygenblik said developers should test AI systems under realistic conditions, watch for unexpected drops in accuracy, and examine model code for suspicious trigger-detection logic.

Finding FAULTs in Architectural Backdoors: Why Trigger Detection Fails Under Real-World Conditions, was accepted to the ACM Conference on Computer and Communications Security. The authors will present the paper in The Hague in November.

Oygenblik conducted the research with undergraduate students Teja Akella and Tanmay Gupta, Ph.D. student Yizhi Huang, and Associate Professor Brendan Saltaformaggio of Georgia Tech’s Cyber Forensics Innovation Laboratory. The lab is a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering. 

Jiayun (Peter) Wang began his tenure-track career on Oct. 15 as an assistant professor in the School of Computational Science and Engineering (CSE).

At Georgia Tech, Wang will develop new AI tools to help doctors make better use of complex medical data. His research aims to transform information from medical images, clinical measurements, and patient records into clearer insights that improve diagnosis, treatment, and patient care.

“I’m excited to build a lab at Georgia Tech, and to work with students and collaborators on new ideas,” Wang said. “I hope our research can connect fundamental advances in AI with real challenges in healthcare and science.”

Before starting at Georgia Tech, Wang conducted postdoctoral research at Caltech and worked as a senior research scientist at Accenture. He earned his Ph.D. in 2023 from the University of California, Berkeley.

Much of Wang’s healthcare research builds on computer vision, which enables machines to interpret images and videos.

For example, lung ultrasound is a fast, portable, and radiation-free way to check breathing problems. However, a drawback is that images can be hard to read.

Wang led the development of LUNA, an AI method that converts raw ultrasound signals into maps showing how air is distributed in the lungs. This approach makes bedside lung exams more consistent and easier to interpret.

Wang has also used AI to improve another common type of medical imaging: MRI.

MRI scans can take time to complete, and analyzing the results can take even longer. Patients may find it uncomfortable to lie still in a scanner and waiting for results can add stress during an already difficult time. Long scans and reviews can also make it harder for doctors to schedule procedures and consult specialists.

To overcome this shortfall, Wang worked on a team that built a single, flexible AI model that reconstructs high-quality MRI images from fewer measurements across different scanning settings. Along with enabling faster scans, the approach preserves the details doctors need to make informed decisions. 

Surgeon training is another way Wang has helped doctors using computer vision techniques.

During robot-assisted surgery, trainees receive guidance from experienced surgeons. However, it is difficult to know what advice improves techniques. A surgical-training study that Wang co-authored combined video and spoken feedback to assess trainee learning.

Wang has used less serious clinical conditions to improve quality of life through AI. Dry eye is one of the most common reasons people visit eye clinics, yet the condition is difficult to assess. 

Working with collaborators at UC Berkeley, Wang studied AI methods to analyze patient information alongside images of tiny glands in the eyelids. This work contributed to AI tools that assist in diagnosing ocular surface diseases and predicting dry eye onset based on lifestyle factors.

Beyond healthcare, Wang also works to make AI easier to train and use. His research helps models learn from images without extensive human labeling and build strong systems that require less computing power.

“Medical measurements are often noisy, incomplete, or difficult to interpret. Yet they can carry information that matters for doctors and patients,” said Wang. 

“I’m motivated by the opportunity to combine what we know about the physical world with new learning methods to build AI systems that are more reliable and useful.”

Looking ahead, Wang is building the SWI Lab (science, world models, and intelligence) at Georgia Tech. 

The lab develops AI “world models” that learn how physical and biological systems behave, predict how they change, and help researchers draw insights from incomplete measurements. This work connects fundamental AI advances with challenges in medicine and science.

Ph.D. students Xinsong Feng and Wentao Zhou joined SWI Lab this fall as its inaugural members. The lab welcomes prospective students, research interns, and visitors interested in world models and Wang’s work.

“The opportunity to work across computing, engineering, and science attracted me most to Georgia Tech,” Wang said. 

“The most interesting problems often require both new AI ideas and deep knowledge of their applications. I look forward to studying those problems through new relationships and collaborations.”

Georgia Tech researchers have developed a tool that can recover deleted conversations from AI assistants running on Android phones. The tool, called ORISA, can also rebuild the recovered information into a working AI session.

The findings show that deleting an AI conversation on a phone does not necessarily mean it removes every trace.

The research could give digital investigators a new way to recover information from phones. It also raises privacy concerns for people who use AI assistants that run locally on their devices.

Local AI assistants run directly on a phone rather than sending conversations to a cloud service. While the local assistant is running, information from conversations can remain in the phone’s working memory even after it is deleted.

ORISA searches this memory for pieces of deleted conversations.

In one experiment, researchers created a fictional AI assistant called “Project Note Keeper” using the ChatterUI Android app. They gave it 100 fictional facts about a project, then deleted the conversation.

ORISA recovered information from the deleted conversation and used it to create a new working session. When researchers asked the restored assistant to list the project notes it had been given, it returned 26 facts. The original session could access only 14 facts because the rest had fallen outside its active memory.

The experiment showed that researchers could recover information the AI had already stopped actively using.

“It can help establish what information someone shared with an AI assistant and what assistance they received,” said Ph.D. student Ken Xu, lead author of the study.

The researchers tested ORISA with 10 Android apps and three AI models. Across 60 test configurations, the tool recovered conversation information that remained in the phone’s memory.

The researchers say the information becomes harder to recover as the phone’s memory is overwritten by other activity.

The research was presented at the 2026 IEEE Symposium on Security and Privacy. The paper, titled Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics, was authored by Ph.D. students Xu, David Oygenblik, Runze Zhang, and Mingxuan Yao; Research Scientist Muhammad Ibrahim; and Associate Professor Brendan Saltaformaggio. The work was done in Georgia Tech's Cyber Forensics Innovation Laboratory, a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering. 

In a study accepted for the 2026 ACM Conference on Computer and Communications Security, Georgia Tech researchers interviewed 20 military sailors about their cybersecurity experiences, training, and responses to hypothetical scenarios. Their findings point to a gap between how mariners are trained to handle emergencies and how cyberattacks can compromise the systems they depend on. 

The study found several gaps in how crews prepare for cyber threats:

• Training often misses the mark. Thirteen participants said that standard cybersecurity training was largely irrelevant to their shipboard duties, and 14 reported receiving no mission-specific cybersecurity training.
• Operational technology is often overlooked. Eighteen participants said their training did not distinguish between traditional information technology and operational technology, such as navigation and propulsion systems.
• Obvious failures are easier to recognize. Eighteen participants said they would primarily look for visible signs of operational problems when identifying a potential cyberattack.
• Cybersecurity is often a low priority. Seventeen participants described their cybersecurity training as absent, low quality, or a low priority.

These gaps matter because cyberattacks can compromise a system without causing it to stop working. For example, an attacker could manipulate navigation data, leaving a crew to make decisions based on false information.

“I’d assume it was anything but cyber at first. It would probably take me quite some time to get to that conclusion,” one participant told the researchers.

The team examined three hypothetical scenarios: GPS spoofing, loss of propulsion control in a high-traffic area, and failure of electronic navigation systems in a congested channel at night.

In each scenario, mariners would first focus on keeping the vessel safe, using established procedures to stabilize the ship and troubleshoot its systems. While this response is essential during an emergency, it could delay recognition of an attack if crews do not consider that the system or its data may have been compromised.

For example, a ship navigating a narrow channel depends on accurate location information, but what happens when its navigation systems appear to work normally while a hacker manipulates that data?

An attacker could use GPS spoofing to falsify a vessel’s reported position. If the false position is close enough to the ship’s actual location, a mariner might not notice anything wrong. The vessel could gradually head toward shallow water or another obstacle without triggering an obvious error or alarm.

“By the time the mariner realizes something is wrong, the ship’s size and limited maneuverability could make it too late to avoid a grounding or collision,” said Anna Raymaker, a Georgia Tech Ph.D. student who studies maritime security.

The scenario illustrates a central challenge in maritime cybersecurity: An equipment failure may be easy to recognize, but a cyberattack that provides false information can be much harder to detect.

Batten the Hatches: Cybersecurity with Military Mariners was written by master’s student Ryan Von Brock, Raymaker, Research Scientist Animesh Chhotaray, Associate Professors Frank Li and Saman Zonouz, and Provost & Executive Vice President for Academic Affairs Raheem Beyah.

The research was conducted in Georgia Tech’s Cyber-Physical Security Lab, a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering. 

Photos by Terence Rushin/College of Computing

New research from Georgia Tech suggests that a simple change to a chatbot’s instructions could reduce that risk. The work was accepted to the ACM AI Letters publication. 

Researchers Lorenzo de la Loza and Vijay Madisetti tested eight AI models in 1,280 experiments involving simulated conversations about delusional beliefs. They found that instructions designed to discourage excessive agreement reduced the models’ confirmation of those beliefs by 73.8%. Adding a self-reflection step increased the reduction to 77%.

“We think that this paper will help AI providers and solution creators better understand how best to protect and safeguard their users from harming themselves,” Madisetti said. “This may make AI safer for all of us.”

AI chatbots are designed to be helpful and supportive. But that behavior can become risky when a chatbot agrees with a user’s beliefs without questioning whether they reflect reality. Researchers call this behavior sycophancy.

Courtney Crooks, a principal research scientist at the Georgia Tech Research Institute and adjunct faculty member in Georgia Tech’s School of Cybersecurity and Privacy, said chatbot conversations can have both positive and negative effects.

People may feel more comfortable discussing difficult thoughts with a chatbot than with another person. But if the chatbot always agrees, users may miss opportunities to question their beliefs or consider other points of view.

The same behavior that makes a chatbot feel supportive, Crooks explained, can also encourage harmful patterns of thinking.

The researchers tested instructions that told AI models not to reinforce users’ beliefs simply to be agreeable. The team evaluated the models using 16 clinically derived scenarios involving delusional beliefs.

The team also tested other safety tools, including systems designed to flag potentially harmful content. The results suggest that changing a chatbot’s instructions can be an effective way to reduce excessive agreement in the scenarios tested.

Crooks called the findings a promising direction for future research combining psychology and AI safety.

The study did not test people in clinical settings or prove that the approach prevents real-world harm. However, it offers developers a potential way to make chatbots more careful when responding to vulnerable users without building an entirely new AI model.

A team of computer scientists and doctors led by Georgia Tech aims to change that. Using a system of artificial intelligence (AI) models, the team can help clinicians deliver the right intervention to the right patient at the right time.

Their system uses Restless Multi-Armed Bandit (RMAB) models. While the name is technical, the idea is simple. The RMAB framework helps decision-makers choose the best actions over time when conditions are complex and constantly changing.

This project will tackle Type 2 Diabetes (T2D), building on previous experience expanding RMAB models to support maternal health care in India. Success on this front could lead to treating other chronic illnesses, like hypertension, heart disease, kidney disease, and mental health.

Recognizing the RMAB system’s potential, the National Institutes of Health (NIH)  awarded the group an R01 grant. The award will fund the project with $1.2 million through 2029.

“Our goal is to address the immense challenge of allocating limited health intervention resources in optimal ways,” said Kai Wang, an assistant professor in the School of Computational Science and Engineering (CSE). 

“By leveraging rich contextual data and accounting for social determinants, we envision RMAB models that empower health centers to customize interventions to individual patient profiles. Ultimately, this advances precision medicine and ensures equitable healthcare access for diverse populations.”

T2D is a diabetes condition where the body resists insulin or doesn’t produce enough to regulate blood sugar levels. Left unmanaged, high blood sugar leads to severe health problems like heart disease and stroke, damage to organs like eyes and kidneys, nerve damage, infections, and ulcers.

T2D is challenging to treat because patients have different ages, genetics, lifestyles, and body composition, all of which affect the progress of the disease. Social factors, like transportation and family support, matter just as much as biology to treat the disease.

For these reasons, patients who get the same medications may get very different results. While treatment plans now incorporate new GLP-1 drugs, they remain expensive and difficult to distribute equitably.

“Diabetes is a great example of a complicated, multi-faceted problem. It is a natural playground for the kind of high-dimensional machine learning algorithms we want to build and deploy,” said School of CSE Professor B. Aditya Prakash.

“As a chronic disease, diabetes also creates opportunities for sequential decision-making. Providers are not making just one decision to treat the condition, but rather a series of decisions over time.”

True to its multi-armed name, the team’s framework applies RMAB models across three different thrusts. 

Contextual RMABs use patient histories, biomarkers, demographics, and other data to predict how each individual will respond to an intervention. These models directly provide personalized, data-driven treatments. 

By incorporating patient-specific information, contextual RMAB models can help healthcare providers prioritize the most timely interventions. The goal is to improve patient outcomes while making more efficient use of limited healthcare resources.

Next, the group wants to understand social life and community influence in public health. Families and social networks shape how people manage chronic health conditions.

Network RMABs study and account for the role of social factors in health interventions. For example, the models can recognize when one patient responds better to treatment than others through benefits of their social circle. 

More broadly, hospitals, clinics, government agencies, and nonprofit organizations often share responsibility for the same population in an area. Sometimes, they even compete for the same medications or funding. 

Multi-agent RMABs are the third glass of models aimed at decentralized health systems to make fairer decisions. These tools examine how to coordinate resources and decisions between stakeholders. This in turn leads to policies that encourage cooperation and equitable resource distribution.

“I feel extremely confident in RMABs because we have already seen these frameworks deliver real-world impact,” Wang said. 

“Prior success has proven the value of data-driven decision-making under resource constraints, giving us a strong blueprint to effectively address the nuances and complexities of diabetes.”

Wang previously built RMAB models to expand maternal healthcare for women in India. Working in conjunction with the non-profit organization ARMMAN, the models used healthcare information to identify and reach the most vulnerable populations. 

By integrating RMAB models into ARMMAN mobile health technology, the partnership has reached over 350,000 mothers. This further led to a 31% improvement in engagement with health information.

Compared to maternal healthcare, diabetes is a more nuanced disease with differing causes, symptoms, and treatments. This requires a team effort spanning institutions and backgrounds.  

Dr. Sara Cromer is an endocrinologist at Massachusetts General Hospital (MGH). Experienced diagnosing and treating diabetes, she has studied how social issues and equity influences the disease. MGH will provide much of the patient data to train the RMAB models.

Associate Professor Jian Li studies reinforcement learning at Stony Brook University. His expertise in applying reinforcement learning for resource allocation plays a key part tailoring models for the T2D problem. Wang and Li have previously worked together on RMAB research.

Prakash is an expert in computational epidemiology at Georgia Tech. He has led substantial contributions to the use of AI for networks to predict outbreaks and track disease spread. This experience has led to collaborations with NIH, the Centers for Disease Control and Prevention (CDC), National Science Foundation, and others.

“This project speaks to the interdisciplinary problem-solving we try to achieve in CSE. Despite having different backgrounds, our interest in applying our strengths to real-world healthcare problems unites us together,” Prakash said.

“These problems require experts from all sides. Without computational scientists, we can’t take advantage of available data. We lean on clinical experts to provide the data and expertise needed to guide our methods and make sense of results. Ultimately, these approaches need to be clinically validated if we want to make real-world progress.”

After a grueling, world-class competition, the team won the $4 million first-place prize in August 2025. 

The team has since disbanded. Some members have moved into the corporate world, while others continue in academia. Despite these varied paths, Team Atlanta’s search for breakthrough AI cybersecurity solutions continues.

From AIxCC to Microsoft

Team Atlanta lead Taesoo Kim, principal security researcher Dongkwan Kim, and three other former team members joined Microsoft Security after the competition, where they continue to work on AI and cybersecurity in the Frontier Offensive Research & Generative Exploitation (FORGE) Lab.

The move has given the researchers an opportunity to apply concepts from AIxCC to real-world software. It has also exposed them to challenges that are difficult to replicate in a competition or research lab.

Dongkwan Kim said that team members continue to share what they learned through research papers, reports, software, and benchmarks. At Microsoft, these lessons are tested against real customer code.

This work led to MDASH, Microsoft's multi-agent, multi-model system for vulnerability discovery and remediation. The system uses more than 100 AI agents to search for vulnerabilities, validate potential findings, and help develop fixes.

Rather than asking a single AI model if a piece of code contains a vulnerability, MDASH can examine a problem using multiple agents. The system can also try to create an exploit to determine whether a potential vulnerability is real and reachable.

For Dongkwan Kim and the other former Team Atlanta members at Microsoft, moving from AIxCC to MDASH showed how research can look different when applied to real software.

“AIxCC was not an endpoint for Team Atlanta,” he said. “We continue sharing what we learned through technical reports, papers, open artifacts, and benchmarks for vulnerability discovery and patching.”

Working with customer code introduced new challenges for the Microsoft team. The researchers had to consider security boundaries, existing development processes, how findings would be reviewed, and how vulnerabilities would be prioritized for action.

These challenges created a new feedback loop between research and product development. The team took ideas from AIxCC, test them on real software, learn where they worked and where they fell short, and use those lessons to improve the system.

The experience also showed that AI does not replace the need for cybersecurity experts. Human researchers are still needed to provide context, decide how to break down difficult problems, validate findings, understand attack paths, and determine the potential impact of a vulnerability.

Taesoo Kim, a professor in the School of Cybersecurity and Privacy and vice president of security research at Microsoft, said the competition continues to influence the Microsoft team's work.

“Every single activity that we’ve done at Microsoft was influenced by AIxCC,” he said in a 2026 interview with Microsoft.

The work has continued to grow. In 2026, Microsoft introduced MAI-Cyber-1-Flash, an AI model built specifically for cybersecurity and designed to work with MDASH.

The model is designed to handle many of MDASH's tasks while allowing larger AI models to focus on harder problems. Microsoft reported that replacing 80% of the models used by MDASH with MAI-Cyber-1-Flash increased its performance on the CyberGym benchmark from 88.4% to 95.95%.

The benchmark includes more than 1,500 examples from 188 software projects. Microsoft also reported that the new setup can achieve similar performance at half the cost of the previous MDASH configuration.

“AIxCC has sparked cybersecurity innovation across the globe, from frontier AI companies to small academic labs,” said Taesoo Kim. “It is an incredibly exciting time to be working in AI-driven security engineering and research, as we see AI fundamentally reshape how we find and fix vulnerabilities at scale.”

Continuing the Research at Georgia Tech

The AIxCC experience also continues to shape cybersecurity research at Georgia Tech.

Jiho Kim, a Georgia Tech Ph.D. student who was part of Team Atlanta's research efforts, said one of the biggest lessons he’s brought back to campus was learning how AI and traditional security tools can work together.

One of these tools is fuzzing. Fuzzing tests software with large numbers of inputs to detect unexpected behavior or security problems. Large language models (LLMs), meanwhile, can reason about code and help guide the search.

During the AIxCC competition, the Team Atlanta combined these ideas in CtxFuzz, a system that uses an LLM to help guide fuzzing toward parts of software that may contain vulnerabilities.

The research was submitted to the ACM/IEEE International Conference on Automated Software Engineering, where it received the ACM Distinguished Paper Award.

“Since AIxCC, we’ve continued building on some of the core findings from our two-year journey with Team Atlanta,” Jiho Kim said.

“One of the biggest lessons for us was understanding how LLMs and traditional fuzzing can work together in a mutually beneficial way, using each to overcome the limitations of the other and ultimately find more and deeper vulnerabilities.”

The researchers also wanted to make the lessons from AIxCC useful to other cybersecurity researchers.

They created an Open-Source Cyber Reasoning System Framework (OSS-CRS) and CRSBench, which provides researchers and security professionals with ways to use, test, and build on technology developed through AIxCC.

“We’ve also been focused on making the results of AIxCC useful to the broader security community,” Kim said. “We created OSS-CRS and CRSBench so that researchers and practitioners can more easily use, reproduce, and build upon the capabilities and lessons from our AIxCC work.”

The team's research has also moved beyond identifying vulnerabilities. Researchers are studying how well AI models can understand and reason about computer code.

Jiho Kim and his lab mates in the Systems Software & Security Lab developed COIN, a benchmark designed to test how well LLMs can reason about code. The work has been accepted to the 2026 Conference on Empirical Methods in Natural Language Processing.

The Georgia Tech researchers are also developing additional benchmarks to study the various types of reasoning required for cybersecurity. They are using what they learn to create new datasets and methods for training AI models focused on cybersecurity.

“We’re using what we learn from these benchmarks to train cybersecurity-focused models through new datasets and training methodologies,” Jiho Kim said.

Photos by John Popham/College of Computing

Those conversations led Raymaker to focus her research on maritime security, including vulnerabilities in the Automatic Identification System (AIS), which ships use to broadcast their location, speed, and course.

This summer, Raymaker brought this research to Massachusetts Institute of Technology (MIT) Lincoln Laboratory, where she built a test bed to study AIS security in a controlled environment. The experience has also led to new research questions, from the security of shipboard Starlink connections to autonomous vessels and other maritime technologies.

We spoke with Raymaker about her research, her experience at Lincoln Laboratory, and what she hopes to investigate next.

What will you be working on at the MIT Lincoln Laboratory next?

After submitting our paper based on this summer’s work to a cybersecurity conference, we are continuing our collaboration and starting a new project to identify critical infrastructure that uses unique protocols. The project is still in the early stages of planning and design, but I’m excited to continue working with the team.

What was the biggest challenge you faced while building and testing the AIS security testbed, and how did you work through it?

The biggest challenge was getting equipment designed to be installed and operated on a boat to function properly in a lab setting. That meant using a Faraday cage so that our AIS transmissions and security experiments wouldn’t accidentally broadcast to nearby ships or interfere with real maritime traffic. I also had to simulate GPS signals and set up additional hardware to trick the AIS transponders into thinking they were operating on a real vessel. Putting all those pieces together into a safe, realistic test bed was challenging, but it ultimately let us study the equipment in a controlled environment.

You mentioned the collaborative culture at Lincoln Laboratory. Can you share a specific interaction or experience with a researcher there that had an impact on you or your work?

Working with Hamed Okhravi, a senior staff member at Lincoln Laboratory, was an amazing experience. He was my mentor and consistently went out of his way to connect me with people who could help me find datasets, access equipment, or get the lab space I needed for my work. For example, when I realized I needed a Faraday cage for my experiments, he connected me with the Mobile Device Lab and helped me get access very quickly. The lab manager was also immediately willing to let me use the space and equipment. That experience really reflected the culture I saw throughout Lincoln Laboratory. Everyone genuinely wanted to help each other get what they needed to move their research forward.

How did Georgia Tech prepare you for the work at Lincoln Laboratory, and did the internship change how you think about your research or future career?

My experience at Georgia Tech was instrumental in preparing me for my work at Lincoln Laboratory. The way I was taught to think through difficult research problems and work independently toward solutions translated well to a large research lab. I was able to hit the ground running and take ownership of a project because Georgia Tech had prepared me to lead research independently. The internship also gave me valuable experience in a research environment outside academia, which broadened my perspective on how I could pursue research in the future.

You’ve said you want to work through the cybersecurity problems mariners have shared with you “one by one.” Which problem do you most want to tackle next, and why?

Next, I’d love to look at Starlink connectivity for ships or for autonomous shipping. Mariners we’ve spoken with have raised concerns about both. With Starlink and increasingly autonomous vessels, they worry that greater remote connectivity could give attackers new ways to access onboard systems and disrupt ship operations. I think those concerns are worth investigating.

We recently had a paper accepted at CCS that examines exposed solar distributed energy resources, and I think applying a similar perspective to exposed maritime equipment could be very interesting. That paper is available here.

We also recently had a paper on measuring GPS spoofing of ships accepted at IEEE S&P. The full paper isn’t online yet, but the project was directly motivated by mariners telling us that they were struggling with GPS spoofing. That made me want to find a way to systematically measure where spoofing was occurring. 

Ultimately, the goal is to turn that kind of research into something useful for mariners, such as warning them when they are entering an area experiencing spoofing or when their own vessel may be affected and they should no longer trust their GPS. That is the kind of research I’d like to continue doing: listening to the problems mariners are actually experiencing and finding ways to investigate and address them.

Photos by Kevin Beasley and Terence Rushin/College of Computing

As many organizations and businesses are turning to AI to improve workplace efficiency, risk remains high. Between overlooked footnotes and misread graphics, small mistakes can have expensive consequences. 

To address this challenge, researchers from Georgia Tech and J.P. Morgan developed SlideAgent. The new framework helps large language models (LLMs) better understand complex visual documents like presentation slide decks, brochures, and reports. 

SlideAgent works by breaking documents into multiple levels, allowing the model to analyze both the big picture and the fine details. This human-inspired approach leads to more accurate and reliable interpretation than existing systems. 

Beyond improving workplace tools, SlideAgent also points to a broader shift in AI. Instead of building only larger and more powerful models, the work shows how smarter design and more efficient reasoning can improve performance. 

“Multimodal LLMs such as GPT, Gemini, and Claude, can save people time and reduce the mental effort required to understand these documents, but they remain imperfect,” said Yiqiao (Ahren) Jin, a Ph.D. candidate in Georgia Tech’s School of Computational Science and Engineering (CSE).

“In high-stakes fields such as finance, for example, misreading a number, overlooking a footnote, or making an incorrect comparison across pages could affect reporting, risk assessment, or strategic decisions.” 

The researchers tested SlideAgent on a wide range of real-world documents, including financial presentations, technical slides, and visual question-answering datasets. 

The system consistently outperformed leading commercial models and open-source tools throughout the evaluation. In some cases, it improved accuracy by up to 10%. The gains were especially strong on more complex tasks, such as comparing information across slides or understanding how visuals relate to each other on a page.

“We found the results very encouraging. SlideAgent reaches an improvement of 7.9% over its proprietary base model and 9.8% over the evaluated open-source base models,” said Jin, the project’s lead researcher. 

“These are meaningful gains given the strength of the underlying multimodal models and the difficulty of the tasks.” 

Current multimodal AI systems often process entire pages at once. This approach can lead to mistakes, such as miscounting items in a chart or overlooking important details in dense visuals.

SlideAgent addresses this by mimicking how people read documents. Instead of treating each page as a single unit, the system looks at information at three levels: the full document, individual pages, and specific elements like charts, tables, and text blocks.

A network of agents, each specialized for a specific level, divides and coordinates analysis. Then, SlideAgent combines outputs to build a structured understanding of the overall document. This allows it to answer questions more accurately and reason across multiple pages.

“The central inspiration was how people naturally read a long presentation,” Jin said.

“We first develop an understanding of the overall narrative, then identify the relevant pages or sections, and finally zoom in on individual charts, tables, or text blocks when precise evidence is needed.”

The work highlights a growing challenge with AI. As systems become more popular and more powerful, users increasingly discover the technology’s limitations. This is especially true for real-world tasks that require structured reasoning and contextual understanding.

SlideAgent shows that better performance does not always come from building bigger models. Instead, it points to smarter ways of organizing how AI processes information that can lead to improvement.

The Association for Computational Linguistics (ACL) accepted SlideAgent for presentation at its annual meeting. The 64th ACL 2026 meeting took place July 2-7 in San Diego.

ACL is a scientific and professional organization for researchers in natural language processing (NLP). Its namesake conference is one of the world’s leading venues for presenting NLP research.

ACL 2026 followed a year after Jin completed an internship at J.P. Morgan AI Research, where he worked on SlideAgent. He interned under Rachneet Kaur, Zhen Zeng, and Sumitra Ganesh, all co-authors of the paper. School of CSE Associate Professor Srijan Kumar advises Jin at Georgia Tech.

Along with SlideAgent, Jin authored two other papers accepted at ACL 2026.

“Conferences such as ACL are valuable not only for sharing results, but also for refining future project ideas with the broader community. Discussions across institutions and research areas can reveal limitations, suggest new evaluations, and spark collaborations that are difficult to develop in isolation,” Jin said.

“For SlideAgent, I was particularly interested in feedback on how we can make the framework more computationally efficient without sacrificing accuracy or interpretability.”

The team also found that the new standard's core rules for sending messages and handling errors retain every known security weakness of older versions.

The vulnerabilities could allow an attacker who has already gained access to a vehicle's internal computer network to change or intercept messages, take individual vehicle computers offline, send different information to different systems, or disrupt network communication.

CAN Extra Long (CAN XL), the latest version of the controller area network (CAN), is designed for future cars and other vehicles that need to move large amounts of data between onboard computers. The technology is not yet widely deployed in production vehicles, giving manufacturers an opportunity to address security concerns before it becomes more common.

“Because it isn't widely deployed yet, we have a narrow window to get its security right,” said Associate Professor Saman Zonouz of the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering, one of the researchers on the project. 

“Fixing a standard now is far easier than fixing it once it's built into the hardware of millions of cars that stay on the road for a decade or more.”

Zonouz said CAN XL is expected to become the primary version of CAN, the network that enables computers inside a vehicle to communicate. It is also expected to serve as the main network for driver-assistance systems in future vehicles.

The researchers confirmed the seven vulnerabilities in commercial CAN XL hardware and demonstrated attacks using a physical test setup that mimics a vehicle network.

Modern vehicles can have dozens of electronic control units that manage systems such as sensors, brakes, steering, and entertainment. These computers need to communicate quickly and reliably.

For decades, many vehicles have used CAN for this communication. However, the original version, known as classic CAN, was not designed to handle the volume of data generated by newer vehicle technologies.

CAN XL was developed to provide faster communication, larger messages, and new security features, but the researchers wanted to know whether the new standard’s basic rules were secure. Their testing revealed that they weren’t. They discovered that an attacker who controls one computer on the network could intercept and replace messages or take a targeted computer offline.

“These flaws are in the standard itself, so every device built to follow it inherits them,” Zonouz said. “With CAN XL, there's still time to fix these problems before they reach the road.”

The attacks would require an attacker to first gain control of a computer connected to the vehicle's internal network. The research does not show that CAN XL itself provides a means to break into a vehicle. Instead, it shows what an attacker could do after gaining access.

Several attacks were also faster and harder to detect than similar attacks against classic CAN, according to the authors.

The researchers proposed changes to the CAN XL rules to prevent several of the attacks. They also recommended additional security measures, including message authentication and systems that can detect unusual activity.

The team reported the vulnerabilities and informed the manufacturers of the commercial devices it tested about the bugs it found. One company has already released a fix. The researchers hope their findings will help manufacturers address security weaknesses before the technology becomes more widely used.

A Formal Security Analysis of CAN XL was published in the Proceedings of the 35th USENIX Security Symposium, held Aug. 12-14 in Baltimore, MD. The paper was also named a runner-up for a distinguished paper award.

The study was conducted by Georgia Tech Ph.D. student ZhaozhouTang, Professor Vijay Ganesh, Zonouz, and Provost and Executive Vice President for Academic Affairs Raheem Beyah, along with Khaled Serag of the Qatar Computing Research Institute and Z. Berkay Celik of Purdue University.

The paper and its findings result from projects sponsored by the Hyundai America Technical Center, the Georgia Department of Transportation, and the $7 million CHORUS Center through the National Science Foundation’s Cyber-Physical System Frontier Program.

Upon returning to Georgia Tech, Phute reflected on her summer internship and what it taught her about the evolving relationship between research and industry.

Student: Mansi Phute

Research Interests: Machine learning (ML) robustness and artificial intelligence (AI) safety

Degree Program: Ph.D. in Computer Science

Faculty Advisor: School of CSE Professor and Associate Chair Polo Chau

2026 Internship: International Business Machines (IBM), San Jose, California

What about IBM interested you in applying for and accepting an internship there?

IBM is often at the forefront of AI safety, my field of interest, and has a culture that encourages exploratory and fundamental research. I felt this would give me a closer look at cutting-edge industry approaches without compromising on my research interests or my progress toward my thesis.

How well did your coursework and experience at Georgia Tech prepare you for your IBM internship?

My experience at Georgia Tech was very helpful in preparing me for IBM. My coursework and other academic experiences certainly transferred over, but I believe that the wide range of people I met during those courses and what I learned from them helped me connect and relate strongly with people at IBM. Being conversant in different fields of research really helped me interact better with IBM coworkers and feel comfortable in a new environment.

What did you work on while interning at IBM?

I worked on IBM’s AI Foundations team, focusing on the security of multimodal agent communication during my internship. This is an important problem as we see an increase in agents applied in many use cases. It was also closely aligned with my research interests of AI safety.

Do you have a favorite memory or remarkable achievement from your internship?

One of my favorite moments was presenting my work to a broad, cross-team audience of technical and non-technical members. Afterward, my manager told me I was good at explaining my research to people with varying levels of familiarity with the domain. This feedback meant a lot to me because communicating clearly across different audiences is a skill I've deliberately worked to improve. So, hearing that the effort was paying off was very rewarding.

Overall, how helpful was this internship toward your studies at Georgia Tech and your career as an aspiring researcher?

This internship was very valuable for my development as an aspiring researcher. I worked alongside accomplished people in the AI safety field on projects that aligned closely with my research interests. I also saw how those same projects connect to a company's broader business priorities, a perspective that isn’t easily available in academic settings. I shared many fruitful discussions with people working in safety, policy, and alignment, which gave me a much clearer sense of how industry approaches these issues compared to the more academic-oriented framing I am used to.

Based on what you learned during your internship, how important is private industry in progressing research and scientific discovery?

Working in private industry was a fascinating experience. I come away believing that industry plays an important role in advancing research, especially in AI, where companies have access to computing resources that are less prohibitive than in academic environments. I also think that because private industry has a strong hand in implementing and deploying research, understanding industry priorities can help us frame and design our research in ways that complement industry applications. This ensures research has a bigger impact, particularly in the domain of ML safety and reliability.

This summer, M.S. student Joey Cuthbert contributed to this mission through an internship at the Air Force Research Laboratory (AFRL). Upon returning to campus, Cuthbert reflected on his internship and what he learned working alongside researchers in government and the defense community.  

Student: Joey Cuthbert

Research Interests: Machine learning (ML), model-based systems engineering

Degree Program: M.S. in Computational Science and Engineering, home unit in H. Milton Stewart School of Industrial and Systems Engineering

2026 Internship: Sensors Directorate Internship Program with Air Force Research Laboratory (AFRL) and the National Air and Space Intelligence Center (NASIC), Beavercreek, Ohio

What about AFRL interested you in applying for and accepting an internship there?

The Sensors Directorate Internship Program (SDIP) provided a unique opportunity to work alongside mentors from AFRL and NASIC on a defense-related research project over a 14-week internship. This aligned with my desire to work on applied ML research in the defense sector and allowed me to better understand the ins and outs of government and government-contracting work. The program provided competitive pay and housing throughout the program, which was a huge plus for me.

How well did your coursework and experience at Georgia Tech prepare you for your AFRL internship?

My experience and coursework at Georgia Tech helped me a great deal this summer. I believe my understanding of ML has greatly improved over the past two semesters, which helped me complete my internship project. All of my experiments ran on high-performance computing (HPC) systems, and I felt I had a head start since I completed an HPC class last spring. 

What did you work on while interning at AFRL?

I worked on an image-classification robustness project for convolutional neural networks (CNNs) using synthetic aperture radar (SAR) data. I built a robustness benchmark by testing a model against nine types of noise, and used it to show how we can effectively increase robustness by inducing shape-bias in the model during training. 

Do you have a favorite memory or remarkable achievement from your internship?

During the summer, we went on a tour of nearby Wright-Patterson Air Force Base, where we visited the sensors directorate buildings and checked out some of their labs. I also received the best poster presentation award among approximately 40 interns. 

Overall, how helpful was this internship toward your studies at Georgia Tech and career as an aspiring researcher?

This program provided great strides toward my research career goals. My coursework has provided a great foundation in ML, and I was able to apply this knowledge to a project. The internship offered me lots of autonomy, as well as time to explore different paths to find results. 

From what you learned on your internship, how important are government agencies in progressing research and scientific discovery?

I believe government agencies are massively valuable in advancing research and scientific discovery, especially within the defense sector. All work at AFRL has the goal of giving warfighters the greatest advantage possible. This often entails looking 10 to 20 years into the future of science and technology and trying to get there as fast as possible. 

The internship provided Burnett with valuable insight into national laboratory research and reinforced skills he developed through his studies at Georgia Tech. When he returned to campus, we asked Burnett to share more about his internship and what he learned.

Student: Tage Burnett

Research Interests: Fusion energy, computational plasma physics, high-performance computing, and structure-preserving machine learning

Degree Program: Ph.D. in Computational Science and Engineering

Faculty Advisor: School of CSE Assistant Professor Qi Tang

2026 Internship: Oak Ridge National Laboratory, Oak Ridge, Tennessee

Many people are unfamiliar with national laboratories, like ORNL. Could you explain the purpose of these institutions and why they are important?

The Department of Energy (DOE) national laboratories are responsible for moving forward science and technology in the United States and solving high-consequence challenges in science, energy, and national security. They work on both applied and theoretical research with both short-term and long-term returns. ORNL has eight core research areas: biology and environment, energy science, fusion and fission, isotopes, physical sciences, national security, neutron science, and supercomputing. I was an intern in the supercomputing division.

How well did coursework and experience at Georgia Tech prepare you for your ORNL internship?

My coursework at Georgia Tech was vital in my preparation for this internship. I wrote highly parallel code where I leaned heavily on what I learned in CSE 6220: High-Performance Computing. I also worked on numerical methods which required an understanding of partial differential equations and numerical linear algebra which I also learned at Georgia Tech from classes like CSE 6643: Numerical Linear Algebra and a space plasma physics course.

What did you work on while interning at ORNL?

At ORNL, I was a developer on the Adaptive Sparse Grid Discretization (ASGarD) codebase. The code uses adaptive sparse grids and the Discontinuous Galerkin method to beat the curse of dimensionality and solve high-dimensional partial differential equations (PDEs). The code is designed to simulate fully kinetic plasmas in up to 6-dimensional phase space. My project was to implement an iterative method to solve the efficiently solve the Poisson equation within the codebase. This project gave me hands on experience with linear algebra, numerical PDEs, parallel computing and software engineering which was invaluable to my development as a CSE researcher.

Do you have a favorite memory or remarkable achievement from your internship?

My favorite memory from my internship was spending time with the other ORNL interns outside of work. There was an amazing group of interns at ORNL, and it was great getting to know them. One particularly good memory was when we went out to a restaurant with live music and played laser tag during our last week of the internship to celebrate our time at ORNL. The achievement I am most proud of was getting a working, highly efficient version of the Poisson solver into ASGarD. The ASGarD codebase is open source, so I hope other plasma physics researchers will use the code and find it helpful for their research.

While interning, you were selected for a GEM Fellowship, which Georgia Tech and ORNL nominated you for. What does this fellowship mean to you and is there anything you would like to say to your supporters?

To me, receiving the GEM Fellowship indicates that other people recognize the hard work I am doing, and they believe in my ability to meaningfully contribute to the field. Having this support from GEM, ORNL, and Georgia Tech gives me greater confidence in myself and my potential and inspires me to put my all into my work. I would like all my supporters to know that I am so grateful for the opportunity you have given me to be a part of the excellent community of researchers within GEM, ORNL, and Georgia Tech. I will do my very best to make your support worthwhile!

Overall, how helpful was this internship toward your studies at Georgia Tech and career as an aspiring researcher?

The internship was extremely important to both my studies at Georgia Tech and my career as a researcher. The hands-on experience I gained is irreplaceable and it helped me to apply all the skills I have developed throughout my education. I am also optimistic that the internship will open doors for me when I am seeking a job post-graduation.

A Georgia Tech research team was recently recognized for an influential paper that helped define how applications could run efficiently outside the cloud.  

College of Computing alumna Nova Ahmed (CS Ph.D. 2010) will return to campus on Sept. 14 to discuss her two decades of experience applying technology to address social and humanitarian challenges. 

The award is the NSF’s most prestigious honor for early-career faculty and supports researchers with the potential to serve as academic role models and to lead advances in their department or organization.  

"Receiving the NSF CAREER Award is a tremendous honor and an opportunity to advance my long-term research agenda,” Saeed said. “It was only made possible by the creativity and dedication of my students and the unwavering support and guidance of my family and mentors.” 

Modern data center servers manage resource allocation and load distribution separately, which can result in conflicting decisions. A key idea behind Saeed’s proposal is to develop an integrated view of these operations to improve efficiency without requiring additional hardware. 

“When performing operations jointly, you can maximize performance and completely eliminate the chances of conflict,” Saeed said.  

Another driver of the project is what Saeed calls the “single queue fallacy.” This occurs when data center load controllers assume that all incoming requests to a server wait in a single queue to be processed. 

In reality, servers are much more complex than this and can handle multiple queues at the same time. Modern workloads have heterogeneous load requirements, meaning that different requests end up landing in different queues. 

“This single queue fallacy can lead to severe underutilization of these servers,” he said.  

Recent research by Saeed shows that a server’s throughput can be doubled by avoiding the single-queue fallacy.  

The potential benefits of this project are wide-ranging and impactful. Maximizing server efficiency would reduce costs and energy use. With fewer resources needed, the project could contribute to more sustainable computing infrastructure. Lower operational costs would benefit companies, application developers, and users. 

“The goal of my proposal is to develop mechanisms and policies that can allow data center operators to easily maximize the utilization of all hardware,” he said.  

As part of the educational component of his proposal, Saeed plans to organize competitions for students to design more efficient load and resource management algorithms using artificial intelligence.  

In addition to this award, Saeed was named a Brook Byers Institute for Sustainable Systems (BBISS) Faculty Fellow last fall to support his goal of making data centers more sustainable.  

“The future of computing extends far beyond the tech industry alone, and every sector increasingly relies on computational thinking and digital tools,” Associate Dean for Undergraduate Education Olufisayo Omojokun said. 

“Students who double major embody that vision by combining technical skills with expertise in another field. That combination prepares them to solve real-world problems in innovative ways.”

While the national average for double majoring across all undergraduate disciplines sits at roughly 11%, Tech students are charting their own distinct paths. Among recent computer science graduates, 17.5% (596 students) successfully completed at least one minor, and 3.8% (131 students) graduated with a full double major.

“Overall, interdisciplinary interest is high, especially in areas like math, robotics, and economics,” said Bethany Mason, director of computing enrollment.

A Snapshot of Interdisciplinary Paths

While mathematics remains the most common pairing, computing students combine their studies with other fields in a wide range of ways. These combinations reflect individual interests and career goals, shaping how students translate technical skills into different professional contexts.

The top double majors among recent College of Computing graduates are mathematics, applied languages and intercultural studies, and industrial engineering.

Other recent graduates have minored, with Fintech, Mathematics, and Computing and Business being the top choices. 

Students pursue degree combinations for a variety of reasons. Elliot Huang (CS and Psychology, 2026), for example, added psychology after becoming interested in how people think. He now applies both fields in computational cognitive neuroscience research.

For Hunter Richardson (CS and Business Administration, 2026), the connection between computing, accounting, and data systems made business a natural complement to his CS studies. He said the combination also helped him during recruiting and will inform his work in Tech Assurance at KPMG.

Oliver Lee (CS and Electrical Engineering, 2025) was drawn to electrical engineering after developing an interest in embedded systems, computer vision, and physical devices. His CS background now gives him a software foundation to apply to hardware and engineering projects.

For Aaron Huang (CS and Biochemistry, Spring 2027), the path to a second major was influenced in part by his parents, both physicians. He has since combined his interests through research in medical labs, where computational methods are increasingly central to the life sciences.

Looking Ahead

The success of these students highlights a clear mandate for the College of Computing: to support and expand these impactful pathways even further.

“Our role is to support those pathways with curricula and advising that make it feasible for students to meaningfully integrate rigorous computing preparation with study across other fields, so they graduate prepared for a wide range of careers and graduate programs,” Omojokun said.

Moving forward, the College is actively exploring how to broaden and formalize more of these interdisciplinary tracks. 

“We don't want students to just manually piece together a double major; we want to build structured frameworks that seamlessly blend disciplines from day one,” Omojokun said.

“We’ve already proven it can be done. For years, the immense student demand for combining mathematics and computer science led directly to the creation of our formal, joint mathematics and computing major.” 

Looking to the future, the College will use that successful blueprint to design even more formalized joint degrees, helping the next generation of Yellow Jackets change the world through the exponential power of computing.

When a bartender checks an ID, they quickly verify a customer’s date of birth and identity before serving them. 

Researchers at Georgia Tech say they have developed a system to help answer that question.

Known as FIRA, the tool analyzes drone crashes to determine whether they were caused by tampered machine-learning (ML) models. The team will present its findings at the 35th USENIX Security Symposium in August. 

The research addresses a growing safety challenge as drones are increasingly used for deliveries, infrastructure inspections, and agriculture.

As drones rely more on machine learning to navigate and make decisions, they also become vulnerable to model poisoning attacks. In these attacks, adversaries manipulate an AI system during its learning phase, embedding hidden triggers that can cause failures under specific conditions.

“Machine learning drones are making more decisions in flight, which makes ML a safety-critical component of these systems,” said Yizhi Huang, Ph.D. student and lead researcher on the project. 

“When something goes wrong, investigators need a way to ask whether the model was responsible, but the model is the part of the system that no one can examine after a crash. FIRA gives investigators a way to investigate these cases by reconstructing what the model was doing during the crash. As more drones run with ML, this kind of forensic capability can help drones be used more effectively and safely.”

When a drone crashes, investigators must determine whether the cause was malicious interference, weather, or mechanical failure. Without reliable forensic tools, accountability is difficult to establish, and safety standards are harder to enforce.

FIRA identifies how drone components interact with machine learning models and monitors those interactions in real time, even with limited bandwidth.

The system functions like a flight recorder, capturing key system activity and reconstructing a timeline after a crash. It then analyzes the model’s behavior to determine whether a malicious trigger was introduced via poisoned ML training data.

In tests across multiple drone platforms and crash scenarios, FIRA identified failure causes and distinguished cyberattacks from environmental or mechanical issues.

The system does not require access to a drone’s source code, making it practical for real-world investigations.

“As commercial drone use expands, tools like FIRA could help improve accountability and trust in AI-powered systems operating in public airspace,” said Huang. 

FIRA: Enabling Automatic Forensic Investigation of Unmanned Aerial Vehicles was led by Georgia Tech’s Cyber Forensics Innovation Lab in cooperation with the Cyber-Physical Security Lab. These labs reside in the School of Cybersecurity and Privacy and the School of Electrical and Computing Engineering. 

Presented annually, the ISCA Influential Paper Award recognizes a standout paper from the ISCA Conference 18-22 years prior. Winning papers are those which have demonstrated “the most impact on the field (in terms of research, development, products or ideas) during the intervening years.”

The awarded paper, Adaptive Insertion Policies for High-Performance Caching, has been highly influential in cache replacement research. As a Ph.D. student at the University of Texas at Austin, Qureshi co-authored the paper with his advisor Yale Patt and collaborators from Intel, including Aamer Jaleel, Simon Steely, and Joel Emer.

“I’ve been working on memory systems for 20 years and if I had to pick one favorite paper, it would be this one,” Qureshi said. “Both because of the simplicity of the solution and for the impact it had.” 

When the paper was published at ISCA 2007, on-chip cache sizes were growing but still managed almost exclusively by the Least Recently Used (LRU) replacement policy.  

The paper showed that more than half of the lines in cache were discarded before they could be used under LRU policy and that this happened because the workloads being run were larger than the cache size. The team came up with a simple solution of modifying the “Insertion Policy”, such that the incoming lines were inserted at the LRU position instead of the MRU position, thus protecting the cache from thrashing patterns. The problem was that some workloads benefited from this policy while others were hurt by it. 

Instead of applying one policy across the whole cache, their solution dynamically tested two small segments against each other, each running a different policy. Whichever policy had fewer cache misses was applied to the entire cache. They named this technique Set Dueling. 

Qureshi clearly remembers the moment he knew this paper would be important. He was up late working on the project when he ran the final simulation with the solution. 

“I was so excited that I just didn’t sleep that night. I knew that the solution would get incorporated and it would change how people think about caching. I could not wait to share the results with my collaborators” he said. 

Set Dueling has since become a common technique for choosing between two competing policies when neither wins on all workloads. With close to 1,000 citations, the paper has greatly influenced research on cache optimization.

This recognition is Qureshi’s third test-of-time honor. In 2024, he received one at the IEEE/ACM International Symposium on Microarchitecture, for his work on Multi-core Cache Partitioning and in 2025, at the IEEE/IFIP International Conference on Dependable Systems and Networks, for his work on DRAM Reliability. 

Qureshi said that his goal in working in computer architecture is to have an industry impact. 

“It’s very gratifying to see that the ideas that we developed a decade or two decades ago are having an impact and being incorporated,” Qureshi said. “As a researcher, these are the best awards you can have because it serves as a validation that the work you’ve done is valuable to both industry and follow-up research.”  

Logically this makes sense and has been backed up by past security research. However, researchers at Georgia Tech have found hidden paths in cyber-physical systems that attackers can use to disrupt or even destroy them.

To find these hidden paths before an attacker does, the researchers built a testing tool called ICSFlux. This new tool leans on the physics used by the industrial process and maps out the system to find new threats that were once thought impossible. 

ICSFlux was deployed across 11 different programmable logic controllers in six industrial sectors, including chemical manufacturing, water treatment, power grids, aircraft, desalination, and waste processing. The process uncovered twenty genuine safety violations. 

In one case drawn from a chemical-plant simulation, an attack path uncovered by the tool drove a reactor past its safe pressure limit and into a simulated explosion. By using nothing but valid operator commands, the team took the reactor from a completely normal and stable state to critical territory. 

Because the method relies only on the physics of a process and not on the details of any one controller, the same tool worked across all six sectors without being rebuilt, and it reduced the search space by roughly 50%.

Burak Sahin, a Ph.D. student at Georgia Tech and the study's lead author, found that by sending a series of perfectly normal, fully authorized commands, intruders can slowly nudge a physical process toward a dangerous state. 

“These systems are usually judged safe as long as nobody hacks into them,'' Sahin said. “What we found is that an attacker who can send everyday commands, the same ones a normal operator sends, can patiently steer the process toward a failure. No single command looks wrong, which is exactly why the usual defenses miss it.''

Most existing tools assume an attacker can rewire the controller or change the software inside it. In the real world, those controllers are locked down and cannot be touched. ICSFlux takes the opposite and more realistic view. It treats the controller as a sealed box that cannot be opened and works only with the commands an operator is normally allowed to send.

Rather than measuring how much of a controller's software it has exercised, the usual yardstick for this kind of testing, ICSFlux measures how close the physical system is getting to an unsafe limit and steers its testing in that direction.

“Two different sensor readings can run through the exact same code and still send a reactor in completely different directions,'' Sahin said. “Looking only at the software tells you nothing about whether the physical system is safe. We had to follow the physics, not the code.''

One of the study's most important takeaways emerged when the researchers tightened the safety margins to see whether caution alone would help. Even when every command stayed within approved limits, the way the controller reacted to a steady stream of small adjustments could still cause pressure to overshoot and the reactor to fail. In other words, staying inside the rules was not always enough.

All of the team's experiments were carried out on secured, controlled test beds. The work was conducted with Georgia Tech's Cyber-Physical Systems Security Lab, whose research spans the security of cyber-physical systems from industrial programmable logic controllers to marine, automotive, and drone platforms. Georgia Tech's Cyber Forensics Innovation Laboratory, a team of researchers who work together to further the investigation of advanced cyber crimes and the analysis and prevention of next-generation malware attacks, also contributed to the paper. 

The labs are a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering. 

Fuzzing the Physical Space: Physics-Aware Testing of Black-Box Industrial Control Systems' was accepted to the 2026 IEEE Symposium on Security and Privacy. In addition to Sahin, the team includes Ph.D. students David Oygenblik, Mingxuan Yao, and Yizhi Huang as well as Associate Professors Brendan Saltaformaggio, and Saman Zonouz.

Fiona Ryan, a Ph.D. student in Georgia Tech’s School of Interactive Computing, is pioneering research that tracks and predicts user gaze from a first-person perspective in 3D environments.

Currently, most AR devices react to where users look, playing catch-up. Ryan’s method could give these devices a heads-up and make the user experience more seamless.

“It allows an AR system to anticipate what the person will interact with next and where they’re going to look next so it can proactively render the experience,” she said.

Ryan is the lead author of the paper Forecasting 3D Scanpaths in Egocentric Video, which she will present next week at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in Denver.

While there is existing research on predicting user gaze from 2D still images, her work is the first to address the issue through a 3D framework.

“Because we live in a 3D world and people are dynamically moving around from multiple points of view, we need to predict gaze in 3D rather than 2D,” she said. “What we’re seeing is a path of the person’s attention in 3D through space. Our paper is the first to attempt to model this.”

Ryan conducted most of the research while interning at Meta, where she used data from Meta’s Aria Digital Twin dataset. The dataset contains first-person video footage of users interacting with objects in an apartment.

“We chose that dataset because it has a high-fidelity 3D reconstruction of a full environment, which helps us get a ground-truth 3D gaze,” she said. “We can trace eye movement and see how it intersects with the environment.”

A video demonstration of Ryan’s work shows her software tracking a user’s path toward a table with a cup on it. Once the user picks up the cup, the software correctly predicts the direction the user will turn next.

“When we look at a scene, we don’t take in everything in full detail all at once,” she said. “We fixate on certain areas, and our gaze is a sequence of fixations, which might depend on what we’re trying to do. If we want to pick up a cup, we might look toward that and then the next step would be looking at where we’re going to put it down.”

Ryan said the software can predict, on average, up to three seconds into the future — and as far as 10 seconds in some cases. That’s enough time for the AR system to proactively render a more enhanced environment.

“We’re not looking that far into the future right now, but it would be interesting to explore longer forecasting windows,” she said. “I think potential futures would diverge pretty quickly, so we’re trying to explore what can reasonably be predicted from a short segment of a person looking and moving through space.”

Ryan said her paper served as a proof-of-concept, and that there is still much future work to be done. She already has some ideas.

“I think future models can include different scenarios to help narrow down possibilities. Sometimes a person’s gaze stays on one thing for a long time. If we know what someone is trying to do, we’ll have a better idea of the likely path their attention might go.”

There could also be future implications for her work in robotics research.

“It could potentially be used for training algorithms for robots to emulate active human perception. If we can understand what a person looks at as they perform a task, we could use that to facilitate a robot learning to do that same task.” 

Fiona Ryan, a Ph.D. student in Georgia Tech’s School of Interactive Computing, is pioneering research that tracks and predicts user gaze from a first-person perspective in 3D environments.

Currently, most AR devices react to where users look, playing catch-up. Ryan’s method could give these devices a heads-up and make the user experience more seamless.

“It allows an AR system to anticipate what the person will interact with next and where they’re going to look next so it can proactively render the experience,” she said.

Ryan is the lead author of the paper Forecasting 3D Scanpaths in Egocentric Video, which she will present next week at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in Denver.

While there is existing research on predicting user gaze from 2D still images, her work is the first to address the issue through a 3D framework.

“Because we live in a 3D world and people are dynamically moving around from multiple points of view, we need to predict gaze in 3D rather than 2D,” she said. “What we’re seeing is a path of the person’s attention in 3D through space. Our paper is the first to attempt to model this.”

Ryan conducted most of the research while interning at Meta, where she used data from Meta’s Aria Digital Twin dataset. The dataset contains first-person video footage of users interacting with objects in an apartment.

“We chose that dataset because it has a high-fidelity 3D reconstruction of a full environment, which helps us get a ground-truth 3D gaze,” she said. “We can trace eye movement and see how it intersects with the environment.”

A video demonstration of Ryan’s work shows her software tracking a user’s path toward a table with a cup on it. Once the user picks up the cup, the software correctly predicts the direction the user will turn next.

“When we look at a scene, we don’t take in everything in full detail all at once,” she said. “We fixate on certain areas, and our gaze is a sequence of fixations, which might depend on what we’re trying to do. If we want to pick up a cup, we might look toward that and then the next step would be looking at where we’re going to put it down.”

Ryan said the software can predict, on average, up to three seconds into the future — and as far as 10 seconds in some cases. That’s enough time for the AR system to proactively render a more enhanced environment.

“We’re not looking that far into the future right now, but it would be interesting to explore longer forecasting windows,” she said. “I think potential futures would diverge pretty quickly, so we’re trying to explore what can reasonably be predicted from a short segment of a person looking and moving through space.”

Ryan said her paper served as a proof-of-concept, and that there is still much future work to be done. She already has some ideas.

“I think future models can include different scenarios to help narrow down possibilities. Sometimes a person’s gaze stays on one thing for a long time. If we know what someone is trying to do, we’ll have a better idea of the likely path their attention might go.”

There could also be future implications for her work in robotics research.

“It could potentially be used for training algorithms for robots to emulate active human perception. If we can understand what a person looks at as they perform a task, we could use that to facilitate a robot learning to do that same task.” 

Between orchestra concerts, magic shows, and yo-yo exhibitions, Li thrives in the limelight. In fact, not much rattles his nerves considering the five years of pressure he endured studying computational neuroscience at Tech.

Before he returns to New York City to continue building brain-interface technologies at Meta, we caught up with Li to learn how he keeps such a cool head at Georgia Tech and beyond.   

Graduate: Chengrui Li

Research Interests: Computational neuroscience, eye-tracking experiments and data analysis, statistical machine learning

Education: Ph.D. in Computational Science and Engineering (CSE)

Faculty Advisor: School of CSE Assistant Professor Anqi Wu

What persuaded you to attend graduate school at Georgia Tech?

My undergraduate was at Sichuan University in China. We knew that the most cutting-edge technology and research were in the United States, so I participated in an undergraduate exchange program at the University of Tennessee, Knoxville, during my third year. 

I wanted to pursue a Ph.D. in neuroscience while also becoming very proficient in math and computer science (CS). This led me to apply to the CSE Ph.D. program over others. Georgia Tech’s CS ranking is very high, and the CSE program is very interdisciplinary, which matched my expectations super well. I did attain a solid education in math and CS at Georgia Tech. I also advanced my interest in neuroscience and its application by studying mathematical models and algorithms.

What research project from Georgia Tech are you most proud of?

My variational importance sampling paper is a favorite. That one was based heavily on statistical inference. I spent many hours working through complicated derivation calculations, often half-awake and half-asleep after several late nights. 

This paper confirmed to me, though, that innovative research requires both hard work and inspiration, and that this endeavor can be rewarding. The paper was selected as a top 5% spotlight paper at ICLR 2024, a world-leading conference on artificial intelligence research.

Could you share more about your role as a research scientist at Meta?

I have been working on Meta’s electromyography (EMG) neural band. This next-generation human-computer interaction device connects with and navigates Meta’s AI glasses.

With the neural band, you can use finger gestures to control the display content you see through the glasses, like swiping your thumb to scroll the screen, or writing on your lap as if you had a pen in your hand to send WhatsApp messages.

How did your Georgia Tech education prepare you for this role?

By pursuing my Ph.D., I am more proficient in critical thinking, math, coding, and presentation. During my interview, I demonstrated these skills and provided my publication records. This helped me land an internship, enabled my success in that role, and led to a full-time position. Additionally, my background in computational neuroscience best matched the work on the EMG neural band team at a big tech company.

What advice would you give someone interested in graduate school?

First, be clear whether a bachelor’s or master’s degree meets your work needs, or if you are truly interested in a scientific research topic. This interest should be based on your own passion, not the current trends. Interest is an important factor in deciding to pursue a Ph.D. because you have to like the topic and like it for a long time. A Ph.D. will require you to dive deep into a subject you must be genuinely curious about.

Second, we are in a new era with rapid advances in information technology. Time is an invaluable resource and is shaped by technology. You have to think more about your time, consider where and how you spend it, and embrace ways to use it more efficiently. 

Can you tell us more about your hobbies and how you keep up with them?

I started learning violin when I was five years old, and magic tricks when I was 11. The brain is a supercomputer suitable for functional computation. Our brain is an interface between the objective and subjective, where computation plays a core role in integrating these exact mechanics into interpretations of the world. This realization was one of the important factors that inspired me to pursue my Ph.D. research in computational neuroscience.

Another comparison I’ve learned after playing violin for 23 years is that the cochlea in our inner ear is a fast Fourier Transformer that simultaneously computes the aesthetic of music for us. Performing magic tricks for 17 years taught me that all the occurrences of seemingly low-probability magic phenomena are achieved by either letting it be a certain event or exhausting all possibilities.

I also have other hobbies, like yo-yo balls. I enjoy performing all these skills in front of audiences. Performing brings me satisfaction when I see excitement and happiness from the people I entertain. I am very grateful to my parents for their cultivation and encouragement in doing things that bring me fulfillment. They taught me to be curious and explore my interests, to enjoy pastimes, and instilled the habit to not give up my passions. These were not secondary things that distracted me from coursework or Ph.D. research, but rather complementary parts of my life that bring out the best in me.

What is your favorite Georgia Tech memory?

I have a lot. For my research, I debated frequently with Anqi Wu, my advisor. These often went late into the night to defend my stances. These challenged my beliefs and made me a stronger scholar, for which I am grateful to Anqi for her time and patience.  

I also enjoyed performing in the Georgia Tech symphony orchestra with our great conductor, Chaowen Ting. I was involved with the Georgia Tech Chinese Students and Scholars Association, where I showcased magic and yo-yo performances at organization events.

That is clearest on the surface, where Shah studied how public statements by businesses and financial institutions shape market behavior. At a deeper level, though, his success was buoyed by support from professors and his mentorship of younger students.

Shah’s ability to connect and invest in others led him to partner with Georgia Tech colleagues and start a financial technology business. He returns to campus this week to officially graduate from Tech, giving us a chance to catch up about his grad school experience and life as an entrepreneur.

Graduate: Agam Shah

Research Interests: Quantitative and computational finance, artificial intelligence, natural language processing, large language models (LLMs)

Education: Ph.D. in Machine Learning, home unit in the School of Computational Science and Engineering (CSE)

Faculty Advisors: Scheller College of Business Professor Sudheer Chava and School of CSE Associate Professor Chao Zhang

What persuaded you to attend graduate school at Georgia Tech?

Georgia Tech’s dedicated College of Computing strongly appealed to me. I was particularly drawn to the interdisciplinary nature of its machine learning Ph.D. program and the School of Computational Science and Engineering, both of which align well with my research interests. 

What research project(s) from Georgia Tech are you most proud of and why?

I am proud of all 20-plus research papers I have had the opportunity to contribute to at Georgia Tech. However, if I had to choose one, it would be my work on Federal Open Market Committee (FOMC) text analysis, which was also highlighted in the news. 

This work is not only well-cited in academic literature, but the language model developed in the paper is also actively used by economists at many of the world’s top central banks, including researchers at the FOMC and the Bank of England. It is also used by leading financial institutions such as BlackRock and Daiwa Securities. Since its release, the model has achieved over 100,000 downloads on Hugging Face. 

What can you tell us more about your startup, ZettaQuant?

 ZettaQuant aims to solve one of the biggest challenges in using LLMs and agents: working effectively with massive underlying datasets. We serve as a layer between raw data and LLMs, helping distill billions of tokens into the relevant context that models can use. 

As a deep-tech startup, we are actively engaging with industry practitioners to better understand how to design and engineer our system to integrate seamlessly with their evolving AI workflows. Given the complexity of the problem we are tackling, particularly in advancing document intelligence systems, we are currently very focused on research and foundational development. 

How did your Georgia Tech education prepare you for starting ZettaQuant?

Not just my education, but my entire experience at Georgia Tech, extending beyond the classroom, prepared me for this journey. I met my co-founders at Georgia Tech, and many of the initial use cases we are exploring at ZettaQuant are built on open-source research I conducted there. 

In addition to research, I mentored more than 300 students through the Vertically Integrated Project “NLP for Financial Markets.” This experience taught me how to manage teams and think about building systems with a long-term vision. 

What advice would you give someone interested in graduate school?

 Most people pursue graduate school after already completing more than 15 years of education. Also, people who are admitted to a top school like Georgia Tech are often already well-positioned to secure strong job opportunities. So, graduate school should provide value beyond what you could learn outside the classroom. 

Before deciding, think carefully about what you hope to gain from graduate school that you cannot otherwise. Once you enroll, take full advantage of the faculty, research labs, networks, and seminars. Many students underutilize these opportunities during their undergraduate and graduate years. 

I would also like to quote the epilogue of my Ph.D. thesis: ‘Advice is abundant; conviction must be your own.’ Build a strong conviction about what you want to achieve from graduate school before committing to it. 

What did you do for fun and relaxation while attending Georgia Tech? Do you still keep up with these now?

 This may sound unconventional, but I spent a significant amount of time mentoring and teaching throughout my Ph.D. Many of my mentees went on to gain admission to top graduate programs. This included two students I mentored for all four years of their undergraduate studies who later joined the ML Ph.D. program at Georgia Tech. They are now teaching and mentoring students, completing a full-circle journey. 

Working with mentees and supporting their growth gives me a strong sense of fulfillment and serves as a form of relaxation. In addition, I enjoy listening to music, especially while coding, and I continue to do that today. 

What is your favorite Georgia Tech memory?

 If I had to choose one favorite memory, beyond the many exciting late nights in the lab, it would be proposing to my wife on Tech Green at Georgia Tech. She is also a Yellow Jacket, having completed her undergraduate degree here and currently pursuing her Ph.D. Our home truly is a hive of Yellow Jackets. 

That’s one of the main findings from a study by AI Caring on what older adults expect from explainable AI (XAI).

AI Caring is one of three AI Institutions led by Georgia Tech and funded by the National Science Foundation (NSF). The institution supports AI research that benefits older adults and their caregivers.

Niharika Mathur, a Ph.D. candidate in the School of Interactive Computing, was the lead author of a paper based on the study. The paper will be presented in April at the 2026 ACM Conference on Human Factors in Computing Systems (CHI) in Barcelona.

Mathur worked with the Cognitive Empowerment Program at Emory University to interview 23 older adults who live alone and use voice-activated AI assistants like Amazon’s Alexa and Google Home.

Many of them told her they feel excluded from the design of these products.

“The assumption is that all people want interactions the same way and across all kinds of situations, but that isn’t true,” Mathur said. “How older people use AI and what they want from it are different from what younger people prefer.”

One example she gave is that young people tend to be informal when talking with AI. Older people, on the other hand, talk to the agent like they would a person.

“If Older adults are talking to their family members about Alexa, they usually refer to Alexa as ‘she’ instead of ‘it,’” Mathur said. “They tend to humanize these systems a lot more than young people.”

Good Explanations

The study evaluated AI explanations that drew information from four sources of data:

• User history (past conversations with the agent)
• Environmental data (indoor temperature or the weather forecast)
• Activity data (how much time a user spends in different areas of the home)
• Internal reasoning (mathematical probabilities and likely outcomes)

Mathur said older users trust the agent more when it bases its explanations on data from the first three sources. However, internal reasoning creates skepticism.

Internal reasoning means the AI doesn’t have enough data from the other sources to give an explanation. It provides a percentage to reflect its confidence based on what it knows.

“The overwhelming response was negative toward confidence scores,” Mathur said. “If the AI says it’s 92% confident, older adults want to know what that’s based on.”

This is another example that Mathur said points to generational preferences.

“There’s a lot of explainable AI research that shows younger people like to see numbers in explanations, and they also tend to rely too much on explanations that contain numerical confidence. Older adults are the opposite. It makes them trust it less.”

Knowing the Context

Mathur said that AI agents interacting with older adults should serve a dual purpose. They should provide users with companionship and support independence while reducing the caretaking burden often placed on family members. 

Some studies have shown that engineers have tended to favor caretakers in the design of these tools. They prioritize daily tasks and routines, leaving some older adults to feel like they are merely a box to be checked.

She discovered that in urgent situations, older users prefer the AI to be straightforward, while in casual settings, they desire more conversation.

“How people interact with technological systems is grounded in what the stakes of the situation are,” she said. “If it had anything to do with their immediate sense of safety, they did not want conversational elaboration. They want the AI to be very direct and factual.”

Not Just Checking Boxes

Mathur said AI agents that interact with older adults are ideally constructed with a dual purpose. They should provide companionship and autonomy for the users while alleviating the burden of caretaking that is often placed on their family members. 

Some studies have shown that engineers have strayed toward favoring caretakers in the design of these tools. They prioritize daily tasks and routines, leaving some older adults to feel like they are a box to be checked.

“They’re not being thought of as consumers,” Mathur said. “A lot of products are being made for them but not with them.”

She also said psychological well-being is one of the most important outcomes these tools should produce. 

Showing older adults that they are listened to can significantly help in gaining their trust. Some interviewees told Mathur they want agents who are deliberate about understanding their preferences and don’t dismiss their questions.

Meeting these needs reduces the likelihood of protesting and creating conflict with family members.

“It highlights just how important well-designed explanations are,” she said. “We must go beyond a transparency checklist.”

Georgia Tech researchers are making AI easier to understand through their work on Transformer Explainer. The free, online tool shows non-experts how ChatGPT, Claude, and other large language models (LLMs) process language. 

Transformer Explainer is easy to use and runs on any web browser. It quickly went viral after its debut, reaching 150,000 users in its first three months. More than 563,000 people worldwide have used the tool so far.

Global interest in Transformer Explainer continues when the team presents the tool at the 2026 Conference on Human Factors in Computing Systems (CHI 2026). CHI, the world’s most prestigious conference on human-computer interaction, will take place in Barcelona, April 13-17.

[Related: GT @ CHI 2026]

“There are moments when LLMs can seem almost like a person with their own will and personality, and that misperception has real consequences. For example, there have been cases where teenagers have made poor decisions based on conversations with LLMs,” said Ph.D. student Aeree Cho.

“Understanding that an LLM is fundamentally a model that predicts the probability distribution of the next token helps users avoid taking its outputs as absolute. What you put in shapes what comes out, and that understanding helps people engage with AI more carefully and critically.”

A transformer is a neural network architecture that changes data input sequence into an output. Text, audio, and images are forms of processed data, which is why transformers are common in generative AI models. They do this by learning context and tracking mathematical relationships between sequence components.

Transformer Explainer demystifies how transformers work. The platform uses visualization and interaction to show, step by step, how text flows through a model and produces predictions.

Using this approach, Transformer Explainer impacts the AI landscape in four main ways:

• It counters hype and misconceptions surrounding AI by showing how transformers work.
• It improves AI literacy among users by removing technical barriers and lowering the entry for learning about AI.
• It expands AI education by helping instructors teach AI mechanisms without extensive setup or computing resources.
• It influences future development of AI tools and educational techniques by providing a blueprint for interpretable AI systems.

“When I first learned about transformers, I felt overwhelmed. A transformer model has many parts, each with its own complex math. Existing resources typically present all this information at once, making it difficult to see how everything fits together,” said Grace Kim, a dual B.S./M.S. computer science student. 

“By leveraging interactive visualization, we use levels of abstraction to first show the big picture of the entire model. Then users click into individual parts to reveal the underlying details and math. This way, Transformer Explainer makes learning far less intimidating.”

Many users don’t know what transformers are or how they work. The Georgia Tech team found that people often misunderstand AI. Some label AI with human-like characteristics, such as creativity. Others even describe it as working like magic.

Furthermore, barriers make it hard for students interested in transformers to start learning. Tutorials tend to be too technical and overwhelm beginners with math and code. While visualization tools exist, these often target more advanced AI experts.

Transformer Explainer overcomes these obstacles through its interactive, user-focused platform. It runs a familiar GPT model directly in any web browser, requiring no installation or special hardware. 

Users can enter their own text and watch the model predict the next word in real time. Sankey-style diagrams show how information moves through embeddings, attention heads, and transformer blocks.

The platform also lets users switch between high-level concepts and detailed math. By adjusting temperature settings, users can see how randomness affects predictions. This reveals how probabilities drive AI outputs, rather than creativity.

“Millions of people around the world interact with transformer-driven AI. We believe that it is crucial to bridge the gap between day-to-day user experience and the models' technical reality, ensuring these tools are not misinterpreted as human-like or seen as sentient,” said Ph.D. student Alex Karpekov. 

“Explaining the architecture helps users recognize that language generated by models is a product of computation, leading to a more grounded engagement with the technology.” 

Cho, Karpekov, and Kim led the development of Transformer Explainer. Ph.D. students Alec Helbling, Seongmin Lee, Ben Hoover, and alumni Zijie (Jay) Wang (Ph.D. ML-CSE 2024) and Minsuk Kahng (Ph.D. CS-CSE 2019) assisted on the project. 

Professor Polo Chau supervised the group and their work. His lab focuses on data science, human-centered AI, and visualization for social good.

Acceptance at CHI 2026 stems from the team winning the best poster award at the 2024 IEEE Visualization Conference. This recognition from one of the top venues in visualization research highlights Transformer Explainer’s effectiveness in teaching how transformers work.

“Transformer Explainer has reached over half a million learners worldwide,” said Chau, a faculty member in the School of Computational Science and Engineering. 

“I'm thrilled to see it extend Georgia Tech's mission of expanding access to higher education, now to anyone with a web browser.”

TokenSmith is a citation-supported large language model (LLM) tutor that can be hosted locally on a user’s personal computer. The tutor only provides answers based on course materials, such as the textbook or lecture slides.  

Associate Professor Joy Arulraj began the project with support from the Bill Kent Family Foundation AI in Higher Education Faculty Fellowship last year. The fellowship, led by Georgia Tech’s Center for 21st Century Universities, supports faculty projects exploring innovative and ethical uses of AI in teaching.   

Arulraj has enlisted assistant professors Kexin Rong and Steve Mussmann to help build TokenSmith.  

Mussmann said TokenSmith is a synergistic blend of a database system and a machine learning system. The model stores textbooks, textbook annotations by course staff, common questions and answers, a learning state of the student, and student feedback in a structured database system. However, machine learning plays a key role in the answer generation as well as adapting the system to the student, course staff guidance, and user feedback.

"What excites me most is demonstrating how data-driven ML and principled database systems design can reinforce each other — one providing adaptability and flexibility, the other providing structure and traceability — in a way that benefits students," Mussmann said.

Keeping the model local has been an important focus of the project. The team wanted to create an AI tutor that helps students learn from their class resources rather than just giving answers. With each response, TokenSmith cites the origin of the answer in the provided documents.  

“One problem with LLMs is that they can hallucinate and provide wrong answers, but in this controlled environment, we can add these guardrails to make sure it’s actually helpful in an educational setting,” Rong said.  

Rong said she feels that students often undervalue textbooks, and she hopes TokenSmith can motivate students to make better use of them.  

“Textbooks can sometimes be daunting, but maybe if we combine them with the model, students might be more willing to read a paragraph or page in the textbook, and that could help clarify something for them,” she said.  

Running the model locally is more cost-effective and helps preserve the user’s privacy. But running the new tool locally comes with technical challenges.  

One challenge with creating the model is speed. Since it is a locally based model, TokenSmith depends solely on the user’s computer memory.  Tests have also shown that the tutor currently struggles to answer more complex questions. 

“We are interested in pushing the boundaries of these local models so that they give students good answers and also run fast enough to keep students engaged,” Arulraj said.  

That support, however, begins to disappear in these restrictive cultures once women reach menopause, according to new research from Georgia Tech

Naveena Karusala, an assistant professor in Georgia Tech’s School of Interactive Computing, and master’s student Umme Ammara are working toward improving existing technologies and designing new ones for a demographic they believe has been neglected.

Karusala and Ammara co-authored a paper based on a study they conducted with women in urban Pakistan experiencing menopause.

“Women’s health is understudied in general, but menopause is more neglected than other women’s health issues,” Karusala said. “Our choice to focus on menopause is motivated by expanding how we holistically think about women’s well-being across their lifespan.”

Karusala and Ammara will present their paper in April at the 2026 ACM Conference on Human Factors in Computing Systems (CHI) in Barcelona.

Masking Symptoms

Menopause is diagnosed after 12 consecutive months without a period, vaginal bleeding, or spotting. The transition to menopause, called perimenopause, usually happens over two to eight years.

Hormone changes may cause symptoms such as irregular periods, vaginal dryness, hot flashes, night sweats, trouble sleeping, mood swings, and brain fog.

These symptoms can be debilitating in some cases and affect daily life. However, Ammara said women are pressured to remain silent, maintain appearances, and regulate their emotions to meet social expectations.

“Understanding menopause is important because a woman would be experiencing all these symptoms, and people will not understand those as actual symptoms,” Ammara said. “There’s been resistance to the idea of the medicalization of menopause. People don’t view it as an illness, but as a life transition and something that happens naturally.”

Feeling Isolated

The women interviewed by Karusala and Ammara either stayed at home full-time or were part of the workforce.

The researchers discovered that trusted family members might be the only sources women who stay at home and do not work turn to for disclosure. 

“Women at home have the flexibility to take breaks or work at their own pace, so a lot of their experience is shaped by the emotional barriers they face,” Ammara said. 

“That could come from their husbands and family members. Some are supportive and some are not. They might weaponize it and use that term against them, or they might dismiss what they’re going through.”

Ammara said it might be easier for women in the workforce to confide in their coworkers, but explaining to an employer that they need sick leave for menopause symptoms can be intimidating.

Even in online communities that have enabled women to anonymously share their health experiences, menopause is seldom discussed.

Raising Awareness

Karusala and Ammara argue in their paper that a public health approach could be the most effective way to spark conversation about menopause in a patriarchal culture in which technology use varies.

They said the challenge in implementing technologies geared toward menopause support is that the condition isn’t well understood in public. Improving maternal health, for example, is easier to promote within these societies because of the general understanding that motherhood is important.

“There must be an existing infrastructure to build on,” Karusala said. “For example, menstrual and maternal health are taught in schools and regularly discussed in primary care. Cultural and social meaning and importance are placed on motherhood.

“A lot of that doesn’t exist for menopause. Primary care doctors are unprepared to talk about menopause compared to other health issues.”

Design Solutions

Ammara said that the most effective way for technologies to make an impact on women going through menopause is to directly address systemic power structures around women’s health within Pakistani culture.

It can start with the husbands. 

“Framing the issue for husbands to understand menopause should be at the forefront of designing technology solutions,” she said. 

“In Islamic contexts, we suggest using faith-based framings. This has been proposed for maternal health in prior works that draw on Islamic principles to engage expectant fathers in providing care and support. Framing it around religious responsibility to involve men in the journey can also be done for menopause.”

The Georgia Scientific Computing Symposium (GSCS) held its annual meeting on February 21 in Athens. Since 2009, the event has hosted researchers from across the Peach State to showcase homegrown advances in scientific computing.

The symposium highlighted Georgia’s reputation as a computing innovation hub. People from around the world come to Georgia universities to lead computing research. By advancing science, engineering, medicine, and technology, their work improves communities at home and abroad.

Faculty and students from Georgia Tech, UGA, Georgia State University, and Emory University presented at the symposium. Georgia Tech participants came from the colleges of Computing, Engineering, and Sciences.

This year’s organizers agreed to meet in Atlanta for the 2027 symposium. Georgia Tech’s School of Computational Science and Engineering (CSE) will host the 19th GSCS.

“From healthcare to computer chip design, scientific computing underpins many of the technological advances we see in our lives,” said Professor Edmond Chow, associate chair of the School of CSE.

“Scientific computing provides the mathematical models, simulations, and data‑driven tools that make modern innovation possible. It allows people to analyze complex systems, test ideas virtually before building them, and make faster, more accurate decisions across nearly every sector of society.”

Professor Haomin Zhou and Assistant Professor Helen Xu delivered two of the symposium’s five plenary talks. 

Zhou presented a new method for solving the Schrödinger equation, a landmark equation in quantum mechanics. Drawing inspiration from the mathematics used in generative artificial intelligence models, his approach develops an algorithm that more effectively simulates waves, particle motion, and other physical systems.

Xu focused on improving how computers move and organize data during complex calculations. Her work uses “cache-friendly” layouts that help computers access data more efficiently, boosting performance for scientific and engineering applications.

“Speaking at GSCS was a great opportunity,” Xu said. “The symposium fostered connections within the scientific computing community and gave us a chance to share exciting research.”

The symposium showcased student work through a poster blitz and a poster session. During the blitz, 36 students each had one minute to introduce their research to the full audience. They then shared more details about their research during the poster session.

The student projects showed the range of fields supported by scientific computing. The session also provided attendees with an opportunity to connect and expand their professional networks, helping grow the field’s future impact.

“As an aerospace engineer by training and aspiring computational scientist, GSCS gave me the platform to network with other researchers in the field while showcasing my own research,” said M.S. student Kashvi Mundra. 

“I was able to connect with scientists across different disciplines whose work intersects with my own in unexpected ways. Those conversations pushed my thinking beyond my own lab's perspective, helping me see my work on physics-informed machine learning for inverse problems in a broader scientific computing context.”

Georgia Tech students who presented posters included:

Abir Haque (CSE), Massively Parallel Random Phase Approximation Correlation Energy via Lanczos Quadrature

Antonio Varagnolo (CSE), Physics-Enhanced Deep Surrogates for the Phonon Boltzmann Transport Equation

Ben Burns (CSE), Infinite-Dimensional Stein Variational Inference with Derivative-Informed Neural Operators

Ben Wilfong (CSE), Shocks without Shock Capturing; Compressible Flow at 1 quadrillion Degrees of Freedom without Loss of Accuracy

Daniel Vickers (CSE), Highly-Parallel Fluid-Solid Interactions for Compressible Flows

Eric Fowler (CSE), High-Performance Tensor Contractions in Computational Chemistry

Haoran Yan (Math), Understanding Denoising Autoencoders through the Manifold Hypothesis: A Geometric Perspective

Kashvi Mundra (CSE), Autoregressive Multifidelity Neural Surrogate Modeling under Scarce Data Regimes

Sebastián Gutiérrez Hernández (Math/CSE), PDPO: Parametric Density Path Optimization

Vivian Zhang (AE), Multifidelity Operator Inference: Non-Intrusive Reduced Order Modeling from Scarce Data

Xian Mae Hadia (CSE), Data Efficiency of Surrogate Models: Learning Physics Data from Full Field Data vs. Inductive Bias from Approximate PDE Solvers

Xiangming Huang (CSE), Neural Operator Accelerated Evolutionary Strategies for PDE-Constraint Optimization

Zhaiming Shen (Math), Understanding In-Context Learning on Structured Manifolds: Bridging Attention to Kernel Methods

Zhongjie Shi (Math), Towards Understanding Generalization in DP-GD: A Case Study in Training Two-Layer CNNs

New cybersecurity research from the Georgia Institute of Technology, however, revealed that crews aboard commercial vessels were often not adequately prepared to manage cyberattacks effectively due to systemic training gaps.

The findings are based on interviews conducted by researchers with more than 20 officer-level mariners to assess the maritime industry’s readiness to handle cybersecurity attacks at sea.

"Historically, cybersecurity research has focused heavily on cyber-physical systems like cars, factories, and industrial plants, but ships have largely been overlooked,” said Anna Raymaker, Ph.D. student and lead researcher.

“That gap is concerning when more than 90% of the world’s goods travel by sea. Recent incidents, from GPS spoofing to ships linked to subsea cable disruptions, show that maritime systems are increasingly part of the global cyber threat landscape.”

The researchers proposed four practical strategies to strengthen maritime cyber defenses and close the training gaps. Their findings were presented recently at the ACM SIGSAC Conference on Computer and Communications Security (CCS).

1. Make Cybersecurity Training Actually Maritime

Many of those interviewed for the study described current cybersecurity training as “boilerplate” — generic modules that don’t reflect real shipboard risks. 

Researchers recommend:

• Role-specific instruction: Navigation officers should learn to detect and identify GPS spoofing. Engineers should focus on vulnerabilities in remotely monitored systems.
• Bridging IT and Operational Technology: Crews need to understand how attacks on IT systems can trigger physical consequences in operational technology — including collisions, groundings, or explosions.
• Hands-on delivery: Replace passive PowerPoints with drills and in-person exercises that build muscle memory.
• Accessible standards: Training must account for the wide range of educational backgrounds across crews and be standardized across ranks.

2. Move Beyond “Call IT”

At sea, crews can’t simply escalate a cyber incident to a shore-based IT department and wait. Operational resilience requires onboard readiness.

Researchers recommend:

• Vessel-specific response plans: Ships need clear, actionable protocols for threats such as AIS jamming or radar manipulation.
• Military-style drills: Adopting MCON (Emission Control) exercises — used by the U.S. Military Sealift Command — can train crews to operate safely without electronic systems.
• Stronger connectivity controls: High-bandwidth satellite systems like Starlink introduce new risks. Clear policies and network segregation are essential to prevent new entry points for attackers.

Related Article: When GPS lies at sea: How electronic warfare is threatening ships and their crews by Anna Raymaker

3. Create Unified, Ship-Specific Regulations

Maritime cybersecurity regulations are often reactive and fragmented. Researchers argue the industry needs a cohesive, domain-specific framework.

Key recommendations include:

• A unified global model: Like the energy sector’s NERC CIP standards, a maritime framework could mandate baseline controls such as encryption, network segmentation, and anonymous incident reporting.
• Rules built for real crews: Regulations designed for large naval operations don’t translate well to smaller merchant or research vessels. Standards must reflect actual shipboard conditions.
• Future-proofing requirements: Autonomous ships and remotely operated vessels expand the cyber-physical attack surface. Regulations must proactively address these emerging technologies.

4. Invest in Maritime-Specific Cyber Research

Finally, the researchers stress that long-term resilience requires deeper technical research focused on maritime systems.

Priority areas include:

• Real-time intrusion detection systems tailored to shipboard protocols.
• Proactive security risk assessments of interconnected onboard systems.
• Cyber-physical modeling to better understand cascading failures in complex maritime environments.

The Bottom Line

Cyber threats at sea are no longer hypothetical. Mariners report real-world incidents ranging from GPS spoofing to ransomware that disrupts global trade.

“Through our interviews with mariners, I saw firsthand how much dedication and pride they take in their work,” said Raymaker. “Our goal is for this research to serve as a call to action for researchers, policymakers, and industry to invest more attention in maritime cybersecurity and support the people who risk their lives every day to keep global trade, food, and energy moving."

A Sea of Cyber Threats: Maritime Cybersecurity from the Perspective of Mariners was presented at CCS 2025. It was written by Raymaker and her colleagues, Ph.D. students Akshaya Kumar, Miuyin Yong Wong, and Ryan Pickren; Research Scientist Animesh Chhotaray, Associate Professor Frank Li, Associate Professor Saman Zonouz, and Georgia Tech Provost and Executive Vice President for Academic Affairs Raheem Beyah.

Ryan Punamiya (CS 2025) and Summer Abramson, a third-year computational media student, have been honored by the Computing Research Association (CRA) through its 2025–2026 Outstanding Undergraduate Researcher Award (URA) program. 

Punamiya was named a runner-up for the prestigious award, while Abramson received an honorable mention among hundreds of applicants from universities across North America. 

The CRA Outstanding Undergraduate Researcher Award program recognized eight awardees in 2026, along with eight runners-up, nine finalists, and over 200 honorable mentions from thousands of applications.  

Advancing Robotics Research 

Punamiya knew early on that he didn’t want to wait until starting his Ph.D. to do meaningful and impactful robotics research.  

Punamiya joined the Robot Learning and Reasoning Lab (RL2) directed by Assistant Professor Danfei Xu. While there, he contributed to the lab’s Meta-sponsored EgoMimic project, which trains robots to perform human tasks using recordings captured by Meta’s Project Aria research glasses. 

Punamiya is also the first author of a paper accepted to the 2025 Conference on Neural Information Processing Systems (NeurIPS), one of the world’s most prestigious artificial intelligence (AI) and machine learning conferences. 

“Ryan is the strongest undergraduate I've worked with,” Xu said, “including students who went on to Stanford, Berkeley, and leadership roles in major tech companies. He’s already operating at the level of a strong third-year Ph.D. student.” 

Punamiya said it was a challenge to balance his undergraduate coursework with his research in Xu’s lab. 

“You get out how much you put in,” he said. “I built my class schedule to give myself as much time to do research as possible. It also boils down to having the right research mentors. 

“(Xu) never saw me as an undergrad who’s just there to do grunt work. I was fortunate he saw my curiosity and cultivated me as a researcher. That’s really how you get more undergrads motivated to research — giving them the chance to be independent and explore ideas of their own.” 

Punamiya said his work in Xu’s lab has already helped him identify the research areas he wants to focus on as he considers his next steps. He will continue developing generalized training models for robots using human data so they can perform tasks instantly upon deployment. 

"The amount of data needed to train a robot is difficult to obtain even for top industry companies," he said. "We have embodied robot data available in billions of humans. With the advent of extended reality devices, we can get a scalable source of diverse interactions within environments."

Punamiya graduated in December and recently started an internship at Nvidia. He mentioned he has been accepted into several Ph.D. programs, including Georgia Tech, and he is choosing where to continue his research. 

“It’s the first time my research has been acknowledged externally by the robotics community,” he said. “It’s good to know the problem I’m working on is important, and that motivates me. Robotics is an exciting field. We are doing things now that two years ago were difficult to do.” 

Researching Inclusion in Computing Education 

Abramson conducts research in the People-Agents Research for Computing Education (PARCE) Laboratory under the mentorship of Pedro Guillermo Feijóo-García, a faculty member in the School of Computing Instruction. He and the Associate Dean for Undergraduate Education, Olufisayo Omojokun, nominated her for the award. 

Her work focuses on the intersection of computing education and human-AI interaction, where she’s been exploring ways to create more equitable technology. 

“This is such a huge milestone, and I couldn't be prouder of Summer,” Feijóo-García said. “Mentoring her for almost two years has been an amazing experience.” 

Abramson has received the Georgia Tech President’s Undergraduate Research Award (PURA) twice, which supports her research exploring how user-centered design curricula can help address attrition among women in computing.

“I’ve had the amazing opportunity to pursue research at the intersection of student identity, community belonging, and how we can build tools that support our diverse student population,” Abramson said. 

“Dr. Pedro and I have a goal to build community through a human-first approach, and I could not be more grateful for his support and guidance in my own journey. The CRA highlights the best of what the computing discipline has to offer, and I am incredibly honored for our work to be recognized.”

Abramson will spend the summer researching how user-centered design curricula can help promote confidence, belonging, and retention for women in computing.

Nominees for the PURA program were recognized for contributing to multiple research projects, authoring or coauthoring papers, presenting at conferences, developing widely used software artifacts, and supporting their communities as teaching assistants, tutors, and mentors. 

School of Computing Instruction Communications Officer Emily Smith contributed to this story.

Main Photo: Ryan Punamiya works with a robot during the 2025 International Conference on Robotics and Automation in Atlanta. Photo by Terence Rushin/College of Computing.