Say less: Adobe Research GEM Fellow’s EMNLP paper shows cost savings of shorter AI responses 

In 2024, Morayo Adeyemi was an undergraduate studying computer science, with a deep interest in AI and efficiency, at Howard University. As she and her classmates dug into their senior thesis capstone, Adeyemi reached out to Adobe researchers for mentorship. Soon, she was corresponding with a senior researcher on how to shape the group’s project. As Adeyemi moved on to graduate work, the relationship evolved into a years-long mentorship, a published research paper at the prestigious Empirical Methods in Natural Language Processing (EMNLP) 2026 conference in Budapest, and a productive Adobe Research GEM Fellowship. 

“Mentoring Morayo has been a fantastic experience,” says Senior Researcher Franck Dernoncourt.  “Her research aligns directly with Adobe Research’s focus on model efficiency, scalability, and practical AI innovation. As large language models grow in size and capability, making them lighter, faster, and more resource-efficient is critical for real-world deployment. Her work helps bridge high-performing model capabilities with scalable, real-world application.” 

Adobe Research and the GEM Fellowship 

Adeyemi is one of more than 100 Adobe Research GEM Fellows who’ve participated in the program since 2012. Adobe is matched with GEM Fellows through the National GEM Consortium, a network of organizations that provide STEM students with internships and financial support for their graduate work. The GEM Fellowship application is open now. 

 “I like how the GEM Fellowship provides a lot of mentorship through the process of doing research,” says Adeyemi. “I got to meet with my mentors a lot before I even started, coming up with the research project that would fit me as an intern and was a fit for my research interests.” 

In addition to supporting the pipeline of technology talent, the GEM Fellowship brings creativity and new insights to Adobe Research. “The energy the GEM interns bring is contagious, and their diverse academic backgrounds introduce perspectives that broaden how we think about our research,” explains Mentor and Research Scientist/Engineer Viet Lai. “They’ve encouraged us to consider the user experience and the value our technologies bring to users. We find ourselves asking questions like: How useful is this product in practice? Are we adding cognitive load for the user? Does the technology actually make the user’s experience better? That broader perspective complements our technically intensive research and helps us think more holistically about the impact of what we build.” 

An internship focused on AI efficiency 

Adeyemi’s first Adobe Research internship project focused on language and efficiency with AI. “Given the increasing costs of AI, I was curious about how to save money without compromising AI’s performance,” explains Adeyemi. 

Adeyemi and her mentors were inspired by the online trend of telling an LLM to talk to users like a caveman. The idea was that reducing vocabulary and simplifying sentence structures might improve LLM efficiency. But Adeyemi wanted to bring a scientific lens to this approach, so the team looked at eight AI models to measure what happens if you tell them to be more concise in their responses, or if you give them more concise prompts. Their research measured accuracy, cost, and reference-text agreement compared with unrestrained generation. They found that output compression (asking a model to be concise in its response) can cut costs, while input compression (giving AI more concise prompts) raised costs and reduced accuracy.  

“For every word that an LLM gives back to you, there are associated costs,” says Adeyemi. “So my research project explored whether we could tell the LLM to be concise, and if that would save us money—and it actually does, while keeping the accuracy.”

 

Adeyemi’s findings were shared at the Empirical Methods in Natural Language Processing (EMNLP 2026) conference in October.

Investigating vision-language models 

This past summer, during her GEM Fellowship, Adeyemi worked with Lai on using vision-language models to understand image quality while editing and generating images. Adeyemi developed a novel approach that can reason about image quality and explain its assessment using professional imaging expertise, all while still preserving access to the model’s underlying probability distribution. The combination offers interpretable, expert-like reasoning and a quantitative quality measure.  

“As image generation and AI-assisted image editing become increasingly capable, an important challenge is understanding the quality of the result,” explains Lai. “Morayo’s research addresses exactly that: how can a model evaluate an image with the kind of nuanced, professional understanding that a creative expert would bring? This kind of research can help bridge advances in foundation models with practical tools that help people create better visual content.”  

Looking ahead

According to Adeyemi, one of the biggest benefits of her collaborations with Adobe Researchers has been connecting with talented, smart people. “I get to be around geniuses all the time—people with really big expertise in creativity with AI. I get to talk to them about my ideas,” she explains. “I think that’s the best part of it.”

Inspired by her time with Adobe Research, Adeyemi plans to continue her work toward a PhD in computer science with a focus on AI for efficiency.

Looking to take your research further? An Adobe Research internship offers the chance to publish at top conferences, work alongside world-class scientists, and see your ideas influence real products. Applications for the 2027 program launch this fall, so keep an eye on our program page for details. In partnership with the GEM Fellowship, Adobe also continues to support a strong pipeline of future STEM innovators.

Wondering what else is happening in Adobe Research? Check out our latest news here.