News Contact

Generative artificial intelligence (AI) is best known for creating images and text. Now, it is helping industries make better planning decisions.

Georgia Tech researchers have created a new AI model for decision-focused learning (DFL), called Diffusion-DFL. Recent tests showed it makes more accurate decisions than current approaches.

Along with optimizing industrial output, Diffusion-DFL lowers costs and reduces risk. Experiments also showed it performs across different fields. 

Diffusion-DFL doesn’t just surpass current methods; it also predicts more accurately as problem sizes grow. The model requires less computing power despite these high-performance marks, making it more accessible to smaller enterprises.

Diffusion-DFL runs on diffusion models, the same technology that powers DALL-E and other AI image generators. It is the first DFL framework based on diffusion models.

“Anyone who makes high-stakes decisions under uncertainty, including supply chain managers, energy operators, and financial planners, benefits from Diffusion-DFL,” said Zihao Zhao, a Georgia Tech Ph.D. student who led the project. 

“Instead of optimizing around a single forecast, the model evaluates many possible scenarios, so decisions account for real-world risk and become more robust.”

[Related: GT @ ICLR 2026]

To test Diffusion-DFL, the team ran experiments based on real-world settings, including:

• Factory manufacturing to meet product demand
• Power grid scheduling to meet energy demand
• Stock market portfolio optimization

In each case, Diffusion-DFL made more accurate decisions than current methods. It also performed better as problems became larger and more complex. These results confirm the model’s ability to make important decisions in real-world scenarios with noisy data and uncertainty.

The experiments also show that Diffusion-DFL is practical, not just accurate. Training diffusion models is expensive, so the team developed a way to reduce memory use. This cut training costs by more than 99.7%. As a result, Diffusion-DFL can reach more researchers and practitioners.

“Our score-function estimator cuts GPU memory from over 60 gigabytes to 0.13 with almost no loss in decision quality, reducing the requirement for massive computing resources,” Zhao said. “I hope this expands Diffusion-DFL into other domains, like healthcare, where decisions must be made quickly under complex uncertainty."

Beyond decision-making applications, Diffusion-DFL marks a shift in DFL techniques and in the broader use of generative AI models. 

In supply chain management, planners estimate future demand before deciding how much product to stock. In this DFL problem, engineers align ML models with predetermined decision objectives, like minimizing risk or reducing costs. 

One flaw of DFL methods is that they optimize around a single, deterministic prediction in an uncertain future.

Diffusion-DFL takes a different approach. Instead of making a single guess, it determines a range of possible outcomes. This leads to decisions based on many likely scenarios, rather than on a single assumed future.

To do this, the framework uses diffusion models. These generative AI models create high-quality data from images, text, and audio. 

The forward diffusion process involves adding noise to data until it becomes pure noise. Models trained via forward diffusion can reverse diffusion. This means they can start with noisy data and then produce meaningful insights from training examples. 

Real-world data is often noisy and uncertain. Traditional DFL methods struggle in these conditions, but diffusion models are designed to handle them.

Because of this, Diffusion-DFL can explore many possible outcomes and choose better actions. Like image-generation AI, the model works well with complex data from different sources. This enables its use across different industries.

“Diffusion models have achieved significant success in generative AI and image synthesis, but our work shows their potential extends far beyond that,” said Kai Wang, an assistant professor in the School of Computational Science and Engineering (CSE).

“What makes Diffusion-DFL unique is that the specific downstream application guides how the model learns to handle uncertainty.

“Whether we are scheduling energy for power grids, balancing risk in financial portfolios, or developing early warning systems in healthcare, we can explicitly train these highly expressive models to navigate the unique complexities of each domain.”

Zhao and Wang collaborated with Caltech Ph.D. candidate Christopher Yeh and Harvard University postdoctoral fellow Lingkai Kong on Diffusion-DFL. Kong earned his Ph.D. in CSE from Georgia Tech in 2024.

Wang will present Diffusion-DFL on behalf of the group at the upcoming International Conference on Learning Representations (ICLR 2026). Occurring April 23-27 in Rio de Janeiro, ICLR is one of the world’s most prestigious conferences dedicated to artificial intelligence research.

“ICLR is the perfect stage for Diffusion-DFL because it brings together the exact community that needs to see the bridge between generative modeling and high-stakes decision-making for real-world applications,” Wang said.

“Presenting Diffusion-DFL allows us to challenge the traditional training framework of diffusion models. It’s about sparking a broader conversation on how we can align the training objectives of generative AI directly with actual, downstream decision-making needs.”

News Contact

When Team Atlanta claimed first place in the DARPA AI Cyber Challenge last year, they weren’t just celebrating a win—they were demonstrating that artificial intelligence (AI) could autonomously detect and patch software vulnerabilities at a scale once considered impossible.

Now, the team is working with the Linux Foundation and the Open Source Security Foundation (OpenSSF) to ensure that its breakthrough doesn’t remain confined to a competition environment. The team’s new initiative, OSS-CRS, aims to standardize and operationalize cyber reasoning systems (CRSs) for real-world use.

“The AI Cyber Challenge pushed the boundaries of autonomous software security, with seven teams developing systems capable of finding and remediating vulnerabilities at scale,” said Andrew Chin, a Georgia Tech Ph.D. student and lead on the OSS-CRS program. 

“However, after the competition’s conclusion, it has been difficult to apply these advancements to the open-source community due to infrastructure incompatibilities and the lack of long-term maintenance for the open-sourced CRS implementations.”

To address this gap, Georgia Tech’s Systems Software Lab (SSLab), directed by Professor Taesoo Kim, is leading the development of OSS-CRS, which provides both a common framework for CRS development and the infrastructure needed to deploy these systems seamlessly across open-source projects.

As part of this effort, the team has ported its competition-winning system, Atlantis, into the OSS-CRS framework. The move makes it compatible with laptops and other everyday machines with flexible resource and budget configurations.

Interoperability is also central to the framework’s design. Atlantis can be combined with other CRSs to improve performance, including systems developed by fellow AIxCC finalists and newer agentic, command-line-based tools. This modular approach reflects a key lesson the team learned from the competition: collaboration between systems can outperform any single solution.

OSS-CRS has been accepted as a sandbox project within OpenSSF’s AI/ML Security Working Group, a milestone that brings added technical guidance and community support to the project. This includes:

• Access to mentorship
• Dedicated working group meetings
• Broader visibility through industry events, publications, and outreach efforts

The collaboration will also foster stronger connections with open-source maintainers, helping streamline vulnerability disclosure and remediation workflows.

News Contact

Walton County, Georgia, didn’t ask to become a test case for the artificial intelligence (AI) infrastructure boom. Meta, the company behind Facebook, Instagram, and WhatsApp, made the decision for them.

In 2018, the company broke ground in Social Circle, a small town an hour east of Atlanta with about 5,000 residents, to build one of its largest U.S. data centers. It opened in 2020.

Local officials called it a win. Shane Short, president and CEO of the Development Authority of Walton County, said the plant generates about $10 million annually in property tax revenue and has led to road improvements and expanded broadband.

Electric vehicle maker Rivian followed Meta’s lead and began construction on a plant near Social Circle in September 2025, adding to the area’s rapid industrial growth.

But for residents, the shift from a largely rural, agricultural economy to an energy-intensive industrial one has put new pressure on power and water systems.

“They’re seeing higher water and power bills, worse air quality, and very few jobs in return for this, while large corporations get tax benefits,” said Ahmed Saeed, an assistant professor in Georgia Tech’s School of Computer Science, describing why residents in some communities push back on new data center development.

Saeed and Josiah Hester, associate professor of interactive computing and computer science and director of the Center for Advancing Responsible AI, have spent the past year studying the energy, water, and financial demands associated with these facilities, and how those costs are distributed.

Betting on Demand

AI data centers run on specialized chips that use large amounts of electricity. That power generates heat, which requires energy- and water-intensive cooling.

The state is adding capacity based on expected demand, not current use.

Last year, the Georgia Public Service Commission approved an estimated $16 billion expansion for Georgia Power to support that growth. It is expected to produce about 10 gigawatts of electricity at a given time. That’s enough energy to power about 7.5 million homes for a year.

If that demand materializes, the electricity is used. If it doesn’t, the cost still has to be paid.

Grid Stability

“Those workloads can put a very large demand on the grid all at once, and then remove it just as quickly,” Saeed said. “That sudden change is difficult for the system to handle.”

That volatility is a separate issue.

Even if data center operators pay for the infrastructure they use, large swings in demand can still strain grid operations, especially during peak periods or extreme weather.

What Comes Next

Back in Walton County, the Meta facility is already attracting additional data centers.

Each new site adds power and water infrastructure designed to operate for decades.

The servers inside need to be upgraded every few years.

Saeed and Hester said if Georgia wants to remain an AI and cloud hub, the state needs to set the terms and companies need to meet them.

That starts with disclosure — how much power data centers draw from the grid, how that demand spikes, and how much water they use. It includes clear expectations for how those facilities respond when the grid is under stress, and protections for the communities where they’re built.

The researchers maintain that “build it and hope” is not a strategy.

News Contact

How did the earliest life on Earth build complex biological machinery with so few tools? A new study explores how the simplest building blocks of proteins — once limited to just half of today’s amino acids — could still form the sophisticated structures life depends on.

The paper, The Borderlands of Foldability: Lessons from Simplified Proteins, is a meta-analysis of six decades of protein research and reveals that ancient proteins may have been far more complicated and dynamic than previously thought. 

Recently published in the journal Trends in Chemistry, the study includes Georgia Tech researchers Lynn Kamerlin, professor in the School of Chemistry and Biochemistry and Georgia Research Alliance Vasser-Woolley Chair in Molecular Design, and Quantitative Biosciences Ph.D. candidate Alfie-Louise Brownless.

Co-authors also include Institute of Science Tokyo graduate student Koh Seya and Liam M. Longo, who serves as a specially appointed associate professor at Science Tokyo and as an affiliate research scientist at the Blue Marble Space Institute of Science.

The research has implications ranging from the origins of life and the search for life in the universe to cutting-edge medical innovation. “One of the biggest unanswered questions in science is how life first began,” says Kamerlin, who is a corresponding author of the study. “Understanding how the first protein-like molecules formed and what the earliest proteins may have been like is a key part of that puzzle.”

“Proteins power our bodies — and all life on Earth,” she adds. “Simply put, the evolution of proteins is the reason that we’re able to have this conversation at all.”

A Protein Folding Paradox

If proteins are the scaffolding of life, amino acids are the components that make up that scaffolding. “Today, an average protein is constructed from a chain of about 300 amino acids, involving 20 different types of amino acids,” Kamerlin shares. Proteins fold when these chains twist into a specific 3-dimensional shape, creating structures critical for biology.

However, while these folds are essential, exactly how a protein knows which way to fold remains a mystery. “We know that proteins didn’t just fold randomly,” Kamerlin shares, “because randomly trying all possible configurations would take a protein longer than the age of the universe.”

It’s a cornerstone problem in biological science called “Levinthal’s Paradox,” and highlights a fundamental mystery: Proteins fold incredibly quickly into very specific combinations — but like a sheet of paper spontaneously folding into an origami swan, researchers don’t know how proteins “choose” the folds they make.

“We can predict what a protein will look like, but can’t tell you how it got there,” Kamerlin adds. “That’s what we’re interested in exploring: how small early proteins developed into the complex proteins that support every living thing on today’s Earth.”

Simple Letters, Sophisticated Structures

Early proteins likely had access to just half of today’s amino acids. “About 10-12 amino acids were likely available on early Earth,” Kamerlin says. Like writing a story with just the letters “A” through “L,” researchers assumed that the ‘vocabulary’ proteins could build from such a limited amino acid alphabet would also be constrained.

“There is a language to protein folding,” Kamerlin explains. “That language is hidden in their structures. Our research is in trying to understand the rules — the grammar and vocabulary that dictate a protein fold.” 

The grammar they discovered was surprising: with a combination of creative techniques and environmental support, complex structures can arise from limited amino acid alphabets. 

“We found that it is possible to develop complex folds with very simple tools — and certain environments, like salty ones, can help support that,” Kamerlin shares. “Early proteins could also cross-link and associate, interacting like LEGO blocks to create more complex structures.”

Pioneering Proteins

Now, the team is conducting research in environments that could mimic conditions on early Earth — aiming to discover more about how these regions could have given rise to today’s complex proteins. “This aspect of our research also ties into the amazing space research happening at Georgia Tech,” Kamerlin says. “While we’re interested in understanding early life on Earth, our work could help inform where best to look for evidence of life beyond our planet.”

Kamerlin specializes in creating computer models that simulate possible scenarios – creating an opportunity to quickly and efficiently test many theories. The most compelling of these can then be tested by her collaborator and co-author at Science Tokyo, Liam Longo, in lab experiments. 

Protein folding is also at the forefront of medical innovation, ranging from diagnostic tools to cancer treatments and neurodegenerative diseases. “In the broader scope, we’re interested in discovering what we can design, what we can stress test, and what we can reconstruct with AI and other computational tools,” Kamerlin says. “Because if you can understand how proteins fold, you gain the ability to design them.”

Funding: NASA, the Human Frontier Science Program, and the Knut and Alice Wallenberg Foundation

DOI: https://doi.org/10.1016/j.trechm.2026.03.001

News Contact

News Contact

Vibe coding programmers are releasing batches of vulnerable code, according to researchers at the School of Cybersecurity and Privacy (SCP) at Georgia Tech, who have scanned over 43,000 security advisories across the web.

The programming style relies on using generative artificial intelligence (AI) to create software code using tools like Claude, Gemini, and GitHub Copilot. According to graduate research assistant Hanqing Zhao of the Systems Software & Security Lab (SSLab), no one had been tracking these common vulnerabilities and exposures before the launch of their Vibe Security Radar.

“The vulnerabilities we found lead to breaches,” he said. “Everyone is using these tools now. We need a feedback loop to identify which tools, which patterns, and which workflows create the most risk.”

The radar extensively scans public vulnerability databases, finds the error for each vulnerability, and then examines the code’s history to find who introduced the bug. If they discover an AI tool's signature, the radar flags it. 

Of the 74 confirmed cases uncovered so far by the tool, 14 are critical risks, and 25 are high. These vulnerabilities include command injection, authentication bypass, and server-side request forgery. Zhao explained that since AI models tend to repeat the same mistakes, an attacker would need to find these bugs just once. 

“Millions of developers using the same models means the same bugs showing up across different projects,” he said. “Find one pattern in one AI codebase, you can scan for it across thousands of repositories.”

Despite its success, the team has only scratched the surface of the problem. The radar can trace metadata like co-author tags, bot emails, and other known tool signatures, but it can't identify an issue if these markers have been removed. 

The next step is behavioral detection. AI-written code has patterns in how it names variables, structures functions, and handles errors. 

“We're building models that can identify AI code from the code itself, no metadata needed,” said Zhao. “That opens up a lot of cases we currently can't touch.”

The team is also improving its verification pipeline and expanding its sources to include more vulnerability databases. The goal is to get a more complete picture of AI-introduced vulnerabilities across open source, not just the ones that happen to leave signatures behind. 

As more programmers rely on vibe coding, Zhao warns that it still needs to be reviewed as thoroughly as any other project. 

“The whole point of vibe coding is not reading it afterward, I know,” he said. “But if you're shipping AI output to production, review it the way you'd review a junior developer's pull request. Especially anything around input handling and authentication.”

When prompting AI, SSLab also recommends providing more detailed instructions to get it closer to production-ready. There are also tools to check the code for vulnerabilities after  code it has been generated. Not double-checking could lead to a catastrophe. 

“The attack surface keeps growing,” said Zhao. “More people running AI agents locally means the attacker doesn't need to break into the company infrastructure. They just need one vulnerability in a model context protocol server that someone installed and never reviewed.”

One reason the attack surfaces are expanding rapidly is AI’s evolution. In the second half of 2025, the Vibe Security Radar found about 18 cases across seven months. Then, in the first three months of 2026, it identified 56. March 2026 alone had 35, more than all of 2025 combined. 

Many tools, like Claude, are now more autonomous, allowing developers to write entire features, create files, and even make architecture decisions. 

“When an agent builds something without authentication, that's not a typo,” said Zhao. “It's a design flaw baked in from the start. Claude Code and Copilot together account for most of what we detect, but that's partly because they leave the clearest signatures.”

News Contact

As students increasingly turn to artificial intelligence (AI) to help with coursework, some worry that their learning could be compromised. Georgia Tech researchers are working to counter this potential decline with an AI tool they hope will promote learning rather than hinder it.  

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.  

News Contact

Manufacturing is undergoing a significant transformation as artificial intelligence reshapes how industrial systems operate, adapt, and scale. The H. Milton Stewart School of Industrial and Systems Engineering (ISyE) has launched its Manufacturing and AI Initiative, which brings together faculty expertise in statistics, optimization, data science, and systems engineering to address emerging challenges and opportunities in modern manufacturing.

ISyE researchers are applying AI to complex manufacturing environments, including multistage production systems, asset management, quality improvement, and human‑centered manufacturing. Faculty leaders emphasize the importance of contextualizing large volumes of manufacturing data so AI can support reliable decision‑making, efficient operations, and sustainable outcomes. At the same time, the initiative acknowledges challenges such as data integration, system complexity, and the need to balance automation with human involvement. Together, these efforts position ISyE at the forefront of shaping AI‑powered manufacturing systems that are innovative, resilient, and socially responsible.

Read the full article in ISyE Magazine 

News Contact

Tickets are now on sale for The Atlanta Opera’s NOW Festival, returning June 12–14, 2026. Celebrating its fifth year, the festival highlights bold, contemporary storytelling and emerging voices in opera through new works and performances.

This year’s festival features the world premiere of Water Memory (Jala Smirti), a chamber opera exploring how artificial intelligence can support individuals living with dementia and provide assistance for aging parents. The festival also includes the 96-Hour Opera Project, a signature competition showcasing newly created opera scenes by emerging artists.

Events will take place across multiple Atlanta venues, including the Ferst Center for the Arts at Georgia Tech and the Ray Charles Performing Arts Center at Morehouse College. Performances include the premiere of Water Memory on June 12, the competition showcase on June 13, and an encore performance on June 14.

Tickets start at $35, with festival passes available for $50. The NOW Festival continues to foster innovation in opera while mentoring the next generation of artists.