Principal Research Scientist Branislav Kveton is interested in the ways people interact with technology—and how AI tools can learn from interactions with humans to get better and better. Kveton has collaborated with Adobe Brand Concierge, which lets brands have AI-driven conversations with their customers, and he’s leading a team of researchers at Adobe Research. A theorist at heart, Kveton talked to us about his work on reinforcement learning and agentic AI models and his insights on the evolving relationship between humans and AI.

How would you describe your research interests?

I’ve always been passionate about algorithms that learn online, and incrementally, from interactions with people.

Here’s an example: You’re putting ads on your website and your objective is to make product recommendations that people will respond to. As people interact with your site and click, you promote things that are aligned with their interests based on their behavior. This is a standard formulation of online learning and reinforcement learning which goes back to the 1980s—and it’s the work that I started doing back in grad school in 2002.

What’s changed recently—and it’s made this work more engaging and exciting—is the interface between a person and a website. It used to be just a click, but clicks are a very limited form of interaction. Today we have generative AI models that people talk to. This is a much richer form of interaction than we had before, and it opens up so many new opportunities.

How are you thinking about generative AI and agentic AI right now?

I’m very excited about the progress we’ve made on generative AI over the past few years. The principle behind the text generated by large language models (LLMs) is that the next word is sampled from a probability distribution defined by previous words. And it turns out that if you take all of human knowledge, you’ve got a lot of examples to learn to do this extremely well. It’s amazing. I never expected to see this in my lifetime.

As far as agentic AI, I look at it like this: I remember my father showing me punch cards he was putting into the computer—we’ve made a lot of progress in programming since then. In particular, we’ve realized that it’s useful to have functions and to have people work on different parts of a program. Later you can call them and use them again—reusability is huge in programming. And the same thing is happening with agentic AI.

The main benefit to agentic AI is that you don’t need to train one big model to do everything. For example, if a model is trained to predict the next word, it doesn’t mean it’s good at solving systems of equations. That’s a very different problem and we already have tools for that. So it makes sense to use an existing tool to do that—and you can think of that existing tool as an agent. Then you just need a way of invoking those agents to get their output when you need it.

One of your main areas of interest is creating seamless interactions between humans and machines. Can you tell us more about your work in that area?

This gets to something I’m very passionate about these days—human simulation using LLMs. Think of how we develop technology that people love to interact with. We do user studies—we come up with a set of prototypes and give them to people to test. This process is extremely slow.

Now imagine you could replace or partially enhance these studies with synthetic evaluation. These days I can take a pre-trained LLM and ask it to imagine it’s a sales assistant in a store. It will look at its pretrained knowledge and give recommendations for things that might not even be in the store. But I can enhance it. For instance, I could fine-tune the model with information about the products in the store so it will behave more like a sales assistant in that particular store.

But how do we build a good model of the people in the store? Again, you can take a large language model and do the exact same thing, pretending to be a person in the store.

From there, we can let the simulators interact with each other and use the generated interactions to learn to be a better sales assistant using reinforcement learning.

Can you tell us about what you’ve been working on lately?

At Adobe Summit, back in March, we announced Brand Concierge. It’s a new application for brands to configure and manage conversational AI agents for product search and discovery, delivering personalized, immersive, and conversational experiences for consumers.

In the past, we gave businesses charts and tables to help them understand their customers. Today, we offer generative AI for marketing, making it easier to personalize every interaction. The next step is adding conversational AI agents to websites, so that people can simply ask questions in natural language and get what they need. And in the future, we’ll see more agents being deployed on websites, because no one wants to just click through links anymore. They want interactive, conversational experiences that feel more human.

How are you thinking about the future of AI?

We keep hearing that people are worried that generative AI is already superhuman. Now, it’s been true for the last 25 years that computers are better at chess than humans, and for 30 years that computers are better at checkers. And for so many of these problems, computers are better because they have more computational power.

But there is much more to being a human—there is empathy, being able to work in groups, and creativity, for example. There are good reasons that evolution brought us to this place, with these capabilities. And I’m looking forward to computers helping us to enhance them.

So the question is how do we find the right spot, where we advance and don’t feel like technology is a threat? Where we feel, instead, that the tools help, and that we can create and innovate faster because of who we are and the tools we’ve built? I really look forward to seeing how far this paradigm between humans and AI will go.

Interested in joining our innovative team? Adobe Research is hiring! Check out our openings for full-time roles and for internships.