ASU Graduate Wins IBM Fellowship for Structured AI Research

Large language models can summarize documents with ease. But ask them to extract precise information and organize it into something structured — like a table a doctor or analyst could rely on — and they often stumble. Important details get missed, information becomes inconsistent, and models generate claims unsupported by the original text. The output looks polished, but it isn't always reliable.

Naman Ahuja wants to change that. This May, he will graduate from the School of Computing and Augmented Intelligence at Arizona State University, earning a master's degree in computer science. Over the past two years, his research has focused on getting AI systems to convert long, unstructured text into accurate, usable tables. That work earned him an IBM Infrastructure Master's Fellowship Award, a prestigious honor recognizing research with strong real-world and industry impact.

The problem exposes a key limitation of modern AI. In the real world, a lot of data exists in complex and semi-structured formats, like PDF documents or Wikipedia pages. These documents have some structure, but they're still complex and contain a lot of information. Ahuja's solution, developed through his master's thesis, rethinks how AI should approach the problem.

Instead of asking a model to generate a table in one pass, Ahuja breaks the task into steps. First, the system extracts atomic facts from the text. Then it builds a plan for how the table should be organized. Finally, it fills in the table incrementally, updating entries as new information appears. The approach mirrors how a human might do the same work: read carefully, decide what categories matter, and then populate the table piece by piece.

His thesis argues that reliable structured generation is about breaking complex tasks into smaller, verifiable steps that reduce errors and improve traceability. That traceability matters most in high-stakes environments like health care, where clinicians often conduct systematic reviews, reading large volumes of research and extracting key findings into tables for decision-making. It's a time-consuming, manual process where mistakes can carry real consequences (and nobody wants to be the person who missed a critical drug interaction).

According to the ASU News report, Ahuja's approach is also designed for what researchers call living data, or information that evolves over time. As new studies are published, systems like the one he is developing can update existing tables instead of rebuilding them from scratch, maintaining consistency while incorporating new evidence.

Vivek Gupta, a Fulton Schools assistant professor of computer science and engineering and head of ASU's Complex Data Analysis and Reasoning Lab (CoRAL), where Ahuja conducted his research, sees the work as part of a broader shift in AI. "Naman's work really captures what we're trying to do in CoRAL. We're focused on complex structured data, especially how to generate it and evaluate it correctly, so we can build AI systems people can trust in real-world settings," Gupta says.

LinkedIn documentation from Gupta provides additional technical context. Ahuja's thesis, titled "Strategies for Structured Data Generation," introduces a schema-guided, plan-then-fill ("map–make") framework for transforming long, noisy narratives into accurate, auditable tables that can power real-world analytics and decision-support systems. Rather than relying on brittle single-pass generation, his approach first constructs an explicit schema and then incrementally populates and revises entries as new evidence emerges.

He pairs this with rigorous reliability evaluation, measuring factual coverage and state consistency — which reduce hallucinations and enhance robustness to perturbations. This work led to a main conference paper at ACL 2025 and directly supports an ASU-Mayo Clinic collaboration on living, interactive clinical practice guidelines, where his methods enable Living Systematic Review and Living Patient Evidence Summary tables that remain updateable, traceable, and consistent as new clinical findings emerge.

That emphasis on reliability and real-world usability is part of what drew attention from IBM's fellowship program. The IBM Academic Awards Program has supported university research since 1951, though the Master's fellowship differs from the more widely publicized PhD track. The Master's award recognizes exceptional graduate students whose work addresses focused areas of interest in technology with clear industry applicability.

For Ahuja, the path to that work began in Hyderabad, India, where he completed his undergraduate degree in computer science before coming to ASU in 2024. His focus has consistently been on turning research into something usable beyond the lab. At ASU, he served as a teaching assistant for a graduate-level natural language processing course, delivered a guest lecture on neural networks, and presented his research at an international conference in Vienna.

Outside the lab, Ahuja tries to stay balanced. He plays basketball regularly, explores new music, and unwinds by watching stand-up comedy in both Hindi and English. As he prepares to graduate, Ahuja is already stepping into the next phase of his career, having accepted a full-time role at Amazon in Seattle, where he will continue working on large-scale systems.

"I'm interested in working on the core systems, how these models are actually built and how they can be used to solve real-world problems," Ahuja says. The move aligns closely with his long-standing interest in applying research to industry challenges.

And while his immediate future is now set, the broader problem he's focused on isn't going away. The world is producing more text than ever, and the need to turn that text into usable knowledge is only growing. For now, that still often means someone, somewhere, building a table by hand. Ahuja's work suggests they don't have to — but whether the industry actually adopts these methods remains the real question.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn