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.”
]]>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.
]]>His paper, “Brain Speaks Louder Than Words: Neurophysiological Opinion Inference via Cognitive Dissonance Triggers,” recently received the 2026 Dr. Hermann Zemlicka Award for the most visionary paper at the NeuroIS Retreat.
Wu discusses the research, its potential applications, and the ethical and privacy considerations it raises.
Q&A With D.J.
What was your reaction when you learned your paper received the 2026 Dr. Hermann Zemlicka Award for “most visionary” research?
The Dr. Hermann Zemlicka Award is presented annually at the NeuroIS Retreat to the paper selected as the year’s “most visionary.” I was surprised and deeply honored to receive it. The recognition was especially meaningful because it affirmed our vision of bringing neuroscience and information systems research together to explore a new way of understanding human opinions.
For readers unfamiliar with your research, what problem were you trying to solve?
We wanted to find a better way to understand people’s opinions in situations where traditional methods, like surveys or tracking behavior, may be biased, intrusive, slow, or difficult to use. Our research explores whether brief brain activity measured by electroencephalography (EEG) can reveal when a statement matches or conflicts with a person’s beliefs. If so, it could provide another way to understand opinions without relying only on what people say or how they behave.
Most people are used to answering surveys, reviews, and polls. What made you wonder if there might be a better way to understand what people really think?
Surveys are extremely useful, but they depend on people actively participating. Those answers can be shaped by factors such as memory, the way a question is worded, social pressure, or a person’s ability and willingness to respond. We wondered whether brain activity could offer another way to understand people’s opinions, especially in situations where those challenges come into play. We wanted to see whether we could capture an immediate reaction before a person has to translate it into words.
The paper title, “Brain Speaks Louder Than Words,” is interesting. Can you give an example of what that means in everyday life?
Imagine someone is asked to rate the statement, “I like this product or service.” Even if they feel pressure to give a positive answer, their brain may respond differently depending on whether the statement matches their experience. Our method looks for patterns of brain activity that occur when there is a mismatch between what a person truly believes and what they are evaluating. It is not mind-reading or lie detection. Instead, it uses the brain’s response to conflicting information as one signal that may help us understand whether a statement reflects someone’s genuine opinion.
Were there any findings that surprised you or challenged what you expected going into the project?
One finding was that cognitive conflict depended on more than whether a statement contradicted someone’s opinion. When a heuristic cue, a simple signal that offers a mental shortcut, was present, the brain response to a conflicting statement became weaker. Put intuitively, some contextual cues can soften our immediate sense that “this does not match what I believe,” even when our underlying opinion remains unchanged.
Privacy is a growing concern in today's digital world. Can you compare the privacy implication of the proposed approach versus the traditional methods?
Our approach could reduce the need to store information such as browsing histories, ratings, or written survey responses. Instead of continuously tracking a person’s behavior, it focuses on collecting a specific brain response during a defined interaction with the user’s consent.
At the same time, brain activity is sensitive biometric information and could reveal more than what is needed for a particular task. Any privacy benefits therefore depend on robust safeguards, including data minimization, meaningful informed consent, clearly defined limits on data use and retention, and strong data security through offline, locally controlled processing and calibration.
Our study offers some initial guidance on these issues, but more research is needed to fully understand the privacy implications of this approach.
How could this research change the way organizations understand customers, employees, or other stakeholders in real time?
In situations where surveys may not tell the whole story, receiving feedback is difficult, or past behavioral data are unavailable, our research could eventually help organizations better understand how people are reacting in near real time. For example, it could help identify whether stakeholders' immediate responses suggest agreement or disagreement with a message, product, digital experience, or decision.
These insights would not replace surveys. Instead, they could complement them by showing where people's immediate reactions align with, or differ from, what they later report. Any use at the individual level, however, would require informed consent, a clearly defined purpose, and strong privacy safeguards.
The paper points to opportunities for helping people who may have difficulty communicating through traditional methods. What possibilities do you see for making technology more inclusive?
Brain-computer interfaces have already shown promise in helping people with severe motor or communication impairments interact with technology. Our research explores one possible way to better understand a person's opinions or preferences through brain activity during specific tasks. Over time, this type of work could contribute to technologies that are more responsive to individual needs and more accessible for people who face challenges communicating through traditional methods.
For readers who haven’t thought about neuroscience or brain-computer interfaces before, what’s the one takeaway you hope they remember from this research?
Brain signals do not provide a direct window into a person's thoughts. They reflect many different processes happening in the brain and body, which means they can be difficult to interpret.
That is why theory and experimental design matter. Rather than searching brain data for patterns without a clear explanation, our research focuses on a specific mechanism: how people respond to information that conflicts with their beliefs. By testing that idea in a controlled setting and calibrating the approach for each individual, we can make the results more meaningful and easier to interpret.
Learn More: Information Technology Management
]]>Initial public offering (IPO) filings are dense disclosures submitted to the U.S. Securities and Exchange Commission. They can stretch hundreds of pages and combine legal language, financial data, and visuals like charts and infographics. Even experienced analysts struggle to read them end to end, leaving many everyday investors relying on headlines or summaries.
New research from Georgia Tech’s Financial Services Innovation Lab reveals how artificial intelligence can make these critical documents more accessible and reveals where AI still has room for improvement.
“IPO-Mine: A Toolkit and Dataset for Section-Structured Analysis of Long, Multimodal IPO Documents” authors include Michael Galarnyk, Siddharth Lohani, Vidhyakshaya Kannan, Sagnik Nandi, Aman Patel, Liqin Ye, Arnav Hiray, Rutwik Routu, Prasun Banerjee, Siddhartha Somani, and Sudheer Chava.
Making the Complex Understandable
The authors developed IPO-Mine, an open-source toolkit and dataset designed to break IPO filings into manageable pieces. Instead of treating filings as a single document, IPO-Mine separates them into structured sections and extracts visuals like charts for analysis.
“IPO filings are important public documents, but they are very difficult to work with at scale,” said Siddharth Lohani, a Georgia Tech computer science major ’26, software engineer at Bloomberg, and co-first author of the study. “We wanted to make these filings easier to analyze systematically.”
Why IPO Filings Matter
IPO filings offer investors a detailed look at how a company operates before it goes public.
“IPO filings give investors a clearer look at how a company makes money, what risks it faces, and how it positions itself before entering the public market,” said Arnav Hiray, a Georgia Tech machine learning doctoral student in the Financial Services Innovation Lab, AI for Financial Markets VIP instructor, and co-author. “Even if everyday investors don’t read the full filing, these documents can help them better understand a company’s prospects.”
For retail investors, that insight can help cut through marketing hype and offer a clearer sense of a company’s underlying strengths and risks.
AI’s Strengths
Research shows that large language models and multimodal AI models can use their speed, scale, and pattern-recognition capabilities to offer significant advantages when analyzing IPO filings, including:
Together, these capabilities could help democratize financial analysis, giving more investors access to insights that were once limited to professionals.
A Key Finding
One of the study’s most striking findings is that IPO filings are evolving in two different directions.
“What stood out was that the text and visuals appear to be moving in different directions,” Lohani said. “Investors and researchers need to pay attention to both.”
The research shows that written sections are becoming more standardized, while visual elements such as charts and infographics are becoming more complex and varied.
“As text becomes more standardized, more of a company’s distinctive story shifts into its visuals,” said Vidhyakshaya Kannan, a Georgia Tech Financial Services Innovation Lab intern and co-author. “Because boilerplate language often looks the same across filings, charts and infographics are increasingly where companies differentiate themselves and make their case to investors.”
For investors, that means understanding a company requires studying the visuals as carefully as the text; figures can’t simply be skimmed or skipped over.
Where AI Struggles
Despite its promise, AI is not yet a perfect solution.
“Our results show that even strong multimodal models can disagree with expert human judgments on financial charts, especially when they are misleading,” Kannan said.
The Future of Financial Transparency
AI can accelerate analysis and uncover patterns, but human judgment remains essential for interpreting the visuals companies increasingly use to tell their story. Even with these limitations, researchers believe tools like IPO-Mine could play an important role in shaping the future of financial transparency.
“IPO-Mine can help researchers, regulators, and investors analyze IPO disclosures more systematically, revealing patterns in how companies communicate risk, performance, and strategy before entering public markets,” said Hiray.
By making complex financial disclosures more accessible and actionable, AI-powered tools have the potential to broaden access to critical market information and support more informed investment decisions.
Watch the researchers explain IPO-Mine and its impact
Learn More: Center for Finance and Technology
]]>For communities across the country, this growth hits close to home, and not without controversy. What do data centers actually deliver for the local economies that host them?
New research from Daniel Yue, assistant professor of Information Technology Management at Georgia Tech’s Scheller College of Business, and his co-author, Yiyang Zeng, examines how data center openings affect local jobs, wages, business activity, and electricity prices. Their findings suggest that geography plays a decisive role in whether communities see meaningful economic gains.
“Enormous amounts of capital are flowing into specific communities, much of it tied to new construction, while rigorous evidence on the local benefits of these facilities has been thin,” said Yue. “Community pushback has organized rapidly across the country. Our paper begins to fill that gap by providing new evidence using detailed, facility-level data paired with county-level economic indicators.”
A Historic Investment Wave
There are more than 2,500 data centers, either active or under construction, across the United States. Individual hyperscale facilities often cost more than $1 billion to construct and can consume as much electricity as a small city. Is it all worth it?
On average, Yue and Zeng found that when a data center opens, the host county sees a measurable lift: Employment rises by about 3.5%, wages by 5%, business establishments by nearly 5%, and household income by about 2%. Building permits also increase sharply, reflecting construction activity tied to new facilities.
These are real, economically meaningful gains. But Yue and Zeng discovered that these gains are much smaller than what might be expected from a large investment. And they’re not evenly distributed.
Why Metro Areas Benefit More
The researchers’ clearest finding is that metropolitan areas capture most of the economic benefits from data centers, while rural areas see far fewer spillover effects. While metro counties saw increases in employment and new business growth, non-metro counties saw no measurable gains.
The reason comes down to what economists call “agglomeration,” or economic density.
Data centers don’t operate in isolation. They rely on construction contractors, engineers, equipment suppliers, professional services, and a skilled workforce. Those connections are far easier to build in places that already have deep labor markets and established business networks.
Metropolitan areas are well-positioned to absorb the indirect spending that data centers generate. High-wage technical employees support restaurants, retail, and local services. Suppliers and contractors can scale up quickly. These spillover effects amplify the impact of the initial investment.
In rural areas, that amplification is much harder to achieve. Facilities tend to employ relatively few permanent workers — often fewer than 100 — and many specialized services are imported from outside the county. As a result, the broader economic ripple never materializes.
That doesn’t mean rural communities see no benefit at all. The research finds a small but real drop in unemployment rates in non-metro counties, and local governments may still gain tax revenue or infrastructure investments. But the sweeping job and wage growth often promised during local recruitment efforts will not likely arrive on its own.
“Location, not facility size, determines whether the local benefits show up,” Yue said.
The Hidden Cost: Electricity Prices
Yue and Zeng’s research also uncovers an important trade-off. Data centers use a lot of electricity. A single large facility can use as much power as roughly 80,000 homes. In areas where the researchers can cleanly measure price effects, electricity prices rise by about 5% after a data center enters.
“When the benefits to communities are small, even downsides like higher electricity prices that strain infrastructure will be felt by locals,” Yue shared.
Who pays for the increased cost of electricity isn’t straightforward. Yue and Zeng’s research suggests communities should ask specific questions about who pays for new infrastructure and how those costs will be distributed. Local utility companies divide costs differently among homeowners, businesses, and large industrial users, including data centers. Because cost-sharing systems vary by state, each data center development is unique.
What Communities Should Consider
As states and cities work to attract or push back against data center construction, Yue and Zeng hope their research will encourage more evidence-based decision-making.
For metro areas with strong labor markets and dense business ecosystems, data centers can deliver meaningful, though not transformative, economic gains. For rural communities, the positive impact is more complicated.
“In 10 years, communities will likely wish they had pressed harder on the quality of the decision itself, including whether the debate was evidence-based, whether their local economy was equipped to capture the gains, and whether the fine print aligned with residents' long-term interests,” Yue said. “It’s vital that communities look past flashy, headline incentive packages and focus on the details: tax abatement structures, electricity tariff arrangements, and who ultimately pays for infrastructure upgrades.”
As data centers continue to dot the American landscape, understanding where they create shared value — and where they don’t — will be critical for community leaders and policymakers alike.
Read More: The Local Economic Effects of Data Center Entry
Omar Garcia Urdiales, CREATE‑X’s associate director of Learn, brings a global entrepreneurial experience to Georgia Tech: founder and CEO of a startup operating in the AWS Accelerator Loft, longtime startup coach in Europe’s major innovation hubs, lecturer across multiple universities, and an external doctoral researcher in entrepreneurship and digitalization. He brings this background to his teaching of Startup Lab’s latest iteration – a significant redesign developed by VentureLab’s Director Keith McGreggor. McGreggor created the course and has evolved it over many years, building on its initial success.
“This new iteration of Startup Lab allows us to meet students exactly where they are,” said McGreggor. “By doing this, we give them the strongest foundation possible, providing them with the tools to grapple with uncertainty and build their confidence.”
Startup Lab has long anchored the Institute’s entrepreneurial pathway with clearer structure, a unified language, and a deeper focus on reflective growth, so more Georgia Tech students can discover (and trust) their own entrepreneurial judgment.
Startup Lab is expanding responsibly, with six sections in Atlanta and additional global sections in France and Asia-Pacific taught by faculty trained in the curriculum. Students here benefit from a program that’s learning across borders and bringing that learning back to campus.
“Startup Lab is not about becoming an entrepreneur, but about engaging in the unknown and adopting entrepreneurial behavior, which can be applied to all career paths,” Urdiales said. “Students become better equipped to identify problem spaces and solve them through evidence-based building.”
Urdiales emphasized that Startup Lab is built for students who are still exploring, uncertain, or are simply curious.
“Many students tell us they’re curious about entrepreneurship but feel not ready,” he said. “They worry they’re too introverted for customer interviews or assume Startup Lab is only for people with fully formed ideas. In fact, those are the most common misconceptions.”
The course’s first few weeks focus on training students to see struggles and patterns in the world. Then, they apply those skills on a team, exploring, designing, and testing a concept with real people. The nonnegotiable outcome isn’t the best idea; it’s a more confident, evidence-driven version of you.
“Startup Lab is strengthening that self-awareness. All of us who are entrepreneurs, we don’t grow linearly. We have various iterations of how we see things,” Urdiales said. “This ability to see patterns or to see problems with customer discovery, it’s a learning process and a growth process.”
Urdiales said that students won’t have a passive experience in the lab.
“To become an entrepreneur, you need to do it. You need to engage with customers. You need to get out of the building,” he said. “It gives you the ability to incorporate theoretical frameworks into practical solutions and then understand these more practical outcomes.”
Aligning with CREATE-X’s culture of continuous iteration, Startup Lab is tightening the hands-on core of the course around four simple, repeatable tools so that entrepreneurial thinking becomes muscle memory, not a one-off assignment. The new iteration of the curriculum, developed by McGreggor, helps students learn to:
“What we have is a frontier model in entrepreneurial education,” said McGreggor. “The result is a course that teaches sound decision making and builds entrepreneurial confidence that rewards authentic discovery and iteration over performative polish. It creates a more solid foundation for entrepreneurial thinking and sets students up to engage more deeply with everything that follows in their CREATE-X pathway.”
As a part of Startup Lab, instructors integrate reflection throughout the semester, which helps students notice patterns of work, make small experiments, and adjust based on what’s learned. Students often worry they’re not the founder type or that their introversion will hold them back; Startup Lab reframes those worries as raw material for growth, including communication skill building and one-on-one interactions you won’t always get in higher-level courses.
Startup Lab integrates HaradaLite — McGreggor's adaptation of the Japanese Harada Method — as a weekly reflection practice in which students keep a reflection log, helping them notice patterns of work, run small experiments, and adjust based on what's learned. With this approach, educators are able to measure the growth of entrepreneurial confidence by self-report, leading to a more quantitative approach to teaching.
There’s no mandated order for CREATE-X courses. Startup Lab simply makes the next steps clearer by providing a shared language and milestone structure across sections and instructors, so whatever comes next (I2P, Capstone, Launch, or an internship), you can carry forward a coherent, evidence- aware story of your work.
“All CREATE‑X Learn sections will work with the same milestone objectives,” Urdiales said. “Students trained in Startup Lab are already trained in the muscles of entrepreneurship. They’re more equipped to go into Make and Launch or be a leader within their industry.”
Startup Lab is about becoming the kind of person who can see opportunities, reason from evidence, and make better decisions when the path isn’t obvious.
“Startup Lab helps students see the world’s problems and fill the gaps with fresh ideas, teaching them to see and understand the important difference between evidence and inference,” said McGreggor. “This lays the foundation that leads to good founders, and builds the entrepreneurial confidence needed to succeed.”
Students in Startup Lab can expect a workshop-heavy, conversation-rich semester with weekly artifacts, scenario-based decision prompts, startup reports, and quizzes that keep you honest about what you’re learning. You’ll assemble a Continuity Pack near the end: a compact bundle of your best discovery evidence, decisions, MVEs, economics, and final story slides so your future self (or your I2P/Launch application) can pick up right where you left off.
The course also sets norms for modern tool use. AI is welcomed as a coach and organizer, after your own baseline thinking and research, and as an enhancement of the real conversations you have. That matters because Startup Lab’s promise is that you build solid judgment under the test of uncertainty, critical to the world of today and the future that is being built.
You don’t have to have it all figured out. If you’re a first-year student still exploring, a junior craving real-world projects, or a senior looking to stand out in interviews, Startup Lab is for you.
Seats fill quickly across all sections — and for good reason.
This course gives you the clearest, most supportive on‑ramp into CREATE‑X, with a global methodology, a unified curriculum, and instructors who believe deeply in your potential to grow. Learn how to think entrepreneurially. See the world differently. Build the confidence that will follow you long after the semester ends.
Register for Startup Lab for Fall 2026.
Marketing Strategist
]]>