For every research paper that gets published, several more languish in digital folders. They’re victims of the painstaking process of academic writing. A new AI-powered platform by two USC researchers aims to change that, automating the tedious work of formatting, citations, and revisions that keeps scientists from sharing their discoveries.

The Mudpile Problem

Mayank Kejriwal has a folder on his computer called “obsolete-mudpile.” Inside are more than 50 unfinished research papers, each representing months of work that will likely never see the light of day.

“These are all things I started with graduate students,” says Kejriwal, research associate professor in the Daniel J. Epstein Department of Industrial and Systems Engineering at USC’s Viterbi School of Engineering and principal scientist at USC Information Sciences Institute. “The experiments are done. The results are there. But writing the paper takes so long that we just run out of time.”

It’s a problem that plagues researchers across academia. The bottleneck isn’t the science itself: it’s the formatting, the citations, the endless revisions to satisfy reviewers, and the game of submitting to different journals with different requirements.

David vs. Goliath: Launching One Day After OpenAI

Now Kejriwal thinks he has a solution. On January 27, he and his co-founder Zhisheng Tang, a PhD student in industrial and systems engineering, launched GRAIL, an AI-powered platform that promises to transform rough notes, spreadsheets, and voice recordings into submission-ready academic papers in under an hour. The timing was uncanny. OpenAI had announced Prism, its own LaTeX editor for scientific writing, just one day earlier.

“We didn’t know OpenAI was working on the exact same problem,” Kejriwal says. “It kind of thrust us into the middle of this David versus Goliath thing.”

But GRAIL has already shown it can compete. In its first 48 hours online, the platform processed over 300 million tokens, roughly equivalent to hundreds of thousands of pages of text. One PhD student spent 12 hours on the platform writing papers through the night. Another researcher uploaded a years-old Excel spreadsheet filled with notes from reading papers and turned it into a fully formatted survey paper in less than an hour.

From Casual Conversations to a Weekend Prototype

Tang first met Kejriwal in 2021 in his advisor’s text analytics class. After publishing their first journal paper together, “It was a lot of work,” Tang recalls, they began imagining a better way. “We should really release both the graduate student and the professor’s whole energy into doing the actual research, instead of being bogged down on details,” Tang says.

The idea remained just talk until one weekend in early 2024. Tang had nothing to do and remembered their conversations. Using Cursor, an AI-powered coding tool, he built a simple chat interface just for fun. “I used vibe coding to build vibe science,” he says, laughing.

That weekend prototype became the foundation for GRAIL.

For Tang, the journey has meant sacrifice. He’s spent roughly two months of nights and weekends building GRAIL, sometimes coding five hours a night when inspiration struck. “Eventually I have no personal life,” he jokes. Now he’s dividing his time between finishing his PhD and scaling GRAIL, building a tool to free scientists from tedious work while pouring his free time into making it happen.

The platform is free for now. It’s a strategic decision after OpenAI’s announcement. But Kejriwal, who initially funded the project with $15,000 to $20,000 of his own money, recently secured backing from Gutter Capital, a competitive accelerator that accepts less than two percent of applicants.

GRAIL works by integrating AI directly into LaTeX, the typesetting language that scientists use to create formatted documents. While platforms like Overleaf already help researchers write in LaTeX, they don’t eliminate the tedious back-and-forth of copying text into ChatGPT, editing it, and pasting it back.

“It’s not flowing. It’s very cumbersome,” Kejriwal says. “We were inspired by Cursor (the AI coding platform that revolutionized software development) and we wanted to create the same experience for science writing.”

The Citation Problem That Plagues AI Writing

The platform’s secret weapon is how it handles citations: the Achilles heel of AI-generated academic writing. Language models like ChatGPT are notorious for inventing fake references, a problem known as “hallucination.” GRAIL claims to have solved this with a proprietary method. The idea came to Kejriwal not in the lab, but while relaxing on vacation in Puerto Vallarta, Mexico.

“We have zero percent hallucination in citations,” he says. “We don’t ask the language model to directly generate references. We use a more old-fashioned method, but it’s still fully automated.”

Kejriwal won’t reveal the technical details, but the approach is simple: when citations can’t be verified, the system leaves them out. “This is a major advantage we have over Prism,” he notes, “because Prism has the AI generate references directly, and it’s not trustworthy at all.”

From Mudpile to Published: Real Results

Tang himself has become one of GRAIL’s most devoted users, a PhD student using the tool he built to accelerate his own research. “I definitely use GRAIL a lot to speed up my own process,” he says.

The results have surprised even Kejriwal. One of his master’s students had completed research but lacked the writing skills to turn it into a paper. Kejriwal had resigned the work to his mudpile folder. After the student used GRAIL, the output was good enough to submit to a journal.

“His weakness just wasn’t a problem anymore,” Kejriwal says.

The implications go beyond saving time. Right now, negative results (experiments that disprove a hypothesis) rarely get published because they’re not exciting enough to warrant the effort of writing them up. But that information is valuable. Other researchers could avoid making the same mistakes if they knew what didn’t work.

GRAIL could also level the playing field for underrepresented researchers. Kejriwal points to faculty at historically Black colleges and universities who often carry double or triple the teaching loads of professors at R1 institutions.

“They have things to contribute, but there’s this gatekeeping effect,” he says. “There’s a certain way you have to write, and if you don’t know how to present your work to the right venue, it’s not going to make it.”

The Road Ahead: Autonomous Scientists by Summer

The bigger vision is even more ambitious. By this summer, Kejriwal plans to launch what he calls an “autonomous scientist,” an AI agent that doesn’t just write papers, but actually conducts computational experiments on its own.

“Imagine you give it one of your papers and tell it to improve your work scientifically,” he explains. “It would suggest improvements, run the experiments, come back with results, and if you like them, write the paper for you.”

It’s the kind of future that raises thorny questions about authorship and the nature of scientific contribution. What if AI starts writing 90 percent of scientific content? Will it reduce the quality of research? How should scientists disclose AI use to journals that now require it?

Kejriwal’s answer is straightforward: whoever uses the tool is the author, and they bear responsibility for verifying the results. “The AI is still just a tool,” he says. “You can’t blame the AI if there’s something wrong. That’s on you.”

But he also argues the concern may be misplaced. According to a recent study by GPTZero, over a hundred papers accepted to NeurIPS, the premier machine learning conference, contained hallucinated references. Those papers were written and reviewed before GRAIL even existed.

“The cat is already out of the bag,” Kejriwal says. “AI use in scientific papers has significantly increased. We need to grapple with the reality of technological change.” Many computer science conferences now allow AI use in papers, with the caveat that authors remain responsible for integrity and originality. With tools like GRAIL, consciously built with scientists in mind, Kejriwal argues the quality of work would actually be higher than with uncontrolled AI use.

For now, GRAIL is focused on computer science and mathematics, fields where research is computational rather than done in wet labs. But Kejriwal sees a world where the platform expands to all of science, partnering with robotics companies that are automating laboratory experiments.

Tang envisions an even broader future. “In two years, maybe, there will be technology that allows natural language or some form of AI control over robots,” he says. “Then we’re talking about doing bioscience experiments, chemistry experiments—any type of science.”

Back in his office at USC, Kejriwal pulled up a paper he started in graduate school in 2015. Using GRAIL, he finally completed it after eleven years and called his old advisor to ask if they should still collaborate.

“I told him, ‘You won’t believe what I have to show you,'” Kejriwal says, laughing. “I finally delivered.”

Published on

Last updated on