Community Spotlight: Niya Stoimenova

Why calling LLMs 'reasoning' models is misleading

By Ioanna Lykiardopoulou

Every month, we spotlight one member for our community. At KickstartAI, we believe that meaningful progress starts with individuals — builders, researchers, founders, and practitioners — who are applying AI in bold, practical ways. Our Community Spotlight series celebrates their work, their challenges, and the lessons they’re learning along the way.

This month, we’re featuring Niya Stoimenova, AI Product & Strategic Foresight Lead at DEUS. With a background spanning mathematics, philosophy, design as well as a PhD in AI reasoning, Niya is focusing on driving enterprise-wide responsible AI transformation. 

Over the past year, leading AI companies have marketed a number of their latest LLMs as “reasoning” models, claiming performance that competes with PhD-level experts in coding and math. Think of OpenAI’s GPT-5,  o3 and o4-mini models, or Anthropic’s Claude 3.7 Sonnet. 

But according to Niya Stoimenova, who holds a PhD in AI reasoning, this labeling is more than hype — it’s misleading. 

During her PhD research, Niya developed a first-of-its-kind theoretical model that maps out the different modes of reasoning one needs to generate novel ideas. She then applied her framework to examine how current AI models can be used to also do that. 

She found that, while AI systems do have certain types of reasoning, they’re missing a very important kind. 

“In order to generate novel insights in general, we make use of three main categories of reasoning,” she says. “Deduction and induction we all know. And then we have a different type of reasoning, abduction, that humans rely on to make explanations and develop new ideas. But that’s missing in AI.” 

This missing link, she argues, makes a big difference not only in what AI can do, but also in how we perceive it. And that’s more than just a matter of semantics. 

The types of reasoning LLMs possess 

According to Niya, deduction, induction, and abduction, all come together to enable a complete reasoning process, or what we perceive as thinking. 

Deduction in the ability to apply a known rule in order to reach a certain conclusion. She provides a very simple example: We know that water boils at 100°C. We see that water is boiling in the pan, and we deduce that its temperature must be 100°C. 

But if we don’t know what the water’s boiling temperature is, we can apply induction, the ability to create rules from a set of observations or patterns. We measure water temperature when it starts to boil, and note that it’s 100°C. We repeat this many times and we induce the general rule: “Water boils at 100°C.”

“Most of machine learning is a combination of deduction and induction,” Niya explains. 

“What LLMs, specifically, are doing is a large-scale induction,” she adds. “They generalize patterns from vast amounts of data and then they try to predict the patterns based on what they've already seen. They're doing induction with some forms of deduction.” 

Deduction and induction are both types of reasoning, Niya says. But this doesn’t necessarily mean that, by possessing them, LLMs can reason at the level needed to reach AGI. “Your calculator is doing deduction. So by the industry’s definition, your calculator is also a reasoning model.”

“Even if you look into OpenAI’s o3, or into some of Anthropic's reasoning models, what they're doing is using chain-of-thought reasoning. They list different types of steps and execute them.” 

But that’s not the type of reasoning that generates novel insights, according to Niya. To do that, they would need abduction.

Why missing abductive reasoning matters 

In its simplest form, (explanatory) abduction is the ability to generate explanations or hypotheses using very limited information. 

Niya provides another example: You walk outside and you see the street is wet. Since you’re in the Netherlands, you assume it probably rained, although it might have actually been that someone cleaned the street. Or a pipe burst.

Although this reasoning isn’t reliable, it helps us navigate everyday situations by identifying the most plausible explanations. Niya argues that the industry confuses this type of abduction with inference to the best explanation (IBE), although IBE is still a form of induction. 

“And when it comes to actually generating novel insights and ideas, then we’re talking about innovative abduction,” she adds. “It works like this: you don’t know what exactly you want to do or how you’re going to do it, but you know why. And there are certain mechanisms humans use to get from that ambiguous “why” to novel solutions.” 

“This is the type of reasoning we need in order to get to general artificial intelligence, and deliver on the promises that AI will create new drugs and invent many different things.” 

Niya emphasizes that new ideas and concepts come from abductive reasoning, while deduction and induction play supporting roles by testing and validating them.

The risks of mislabeling LLMs’ capabilities 

In Niya’s view, reasoning as a whole comes together and leads to novelty when we employ all its different types. This is why she finds the labeling of LLMs as “reasoning” models misleading. 

“A lot of people don't really understand what type of reasoning the model is providing to them,” she says, which means that they might misinterpret their capabilities. 

“What LLMs generate seems very plausible, and can even feel novel, even though there is very little novelty in it,” she explains. “And while their answers seem to be true, it doesn't necessarily mean they are. This comes from the largely inductive nature of LLMs. So,unless you have very deep knowledge of a subject, it’s very difficult to catch mistakes or hallucinations.” 

The way LLMs display confidence adds to the problem: they sound convincing. “Unfortunately, I think not too many people know how to properly manage the pitfalls of these models, precisely because LLMs are so useful.” 

Niya adds that this becomes even more challenging when implementing them into automated workflows and day-to-day operations. “It’s not enough to just give people access to Copilot or ChatGPT. You need education and you need to rethink a lot of parameters on how your organization operates." 

“How do you actually design the workflows? How do you talk about organizational processes? How do you understand outcomes? How do you define roles? And how do you actually put these models into existing systems, or redesign them?” 

Ultimately, she argues, value won’t come from the models alone, but from the systems we build around them. “You need proper guardrails,” Niya says. “And you need to create a blend of human and LLM-augmented type of interaction.” 

💡 Want to learn more about the KickstartAI community and our mission to accelerate AI adoption in the Netherlands? Check out our website, listen to our podcast, or follow us on LinkedIn! Interested in speaking with Niya? Connect with her on LinkdeIn!