How do large language models actually "think"? Interview with Erik Arakelyan

Just a few years ago, the capabilities that large language models possess today seemed more like science fiction. These systems can now process vast amounts of information, analyse texts, answer complex questions, and perform tasks that previously required human involvement. But behind the familiar chatbot interface lies a far more complex process: how exactly does artificial intelligence learn to understand language, acquire and process information, and, in a certain sense, "reason"?

Erik Arakelyan, Senior Research Scientist Group Manager at NVIDIA, discussed this and much more on the programme "Did You Know?"

The partner for this series of interviews on science and technology is Team Telecom Armenia.

How a language model is trained

According to Arakelyan, for a long time the foundation of training large language models was text data collected from across the internet. Today, however, the dataset is significantly broader: models are also trained on images, video, and audio. At the same time, the basic training principle can be explained quite simply—the model must learn to predict which word is most likely to appear next, given the existing context.

He noted that each such prediction is a mathematical problem. The model analyses many possible continuations, calculates the probability of each, and then selects the most suitable option. It is precisely through the enormous number of examples that it gradually learns to capture language patterns, styles, and context.

Moreover, the model does not simply memorise a single continuation. In natural language, there are almost always several possible ways for a phrase to develop, and the system evaluates the probabilities of many variants. The more diverse the data it receives during training, the more accurate this evaluation becomes. Errors do not disappear entirely, but as training progresses, the model becomes increasingly better at predicting the most likely continuation given the context.

How the model understands context

However, the ability to predict the next word alone is not enough for the model to understand context and follow instructions. As Arakelyan explains, in large language models, each word is effectively represented as a set of numbers, and the relationships between words are expressed through the mathematical proximity of these numerical representations. This creates a kind of language map, where words are connected to other words, concepts, and contexts. For example, the concepts of "love" and "hate", various emotions, or the names of capital cities have certain mathematical relationships with each other.

When the model receives a phrase, it essentially evaluates which sequence of numerical representations is most likely to appear next. To do this, it uses mathematical and optimisation methods that allow it to select the most appropriate continuation given the entire preceding context.

However, the next important stage in the evolution of language models was instruction tuning—training to follow instructions. In the initial phase, the model primarily learns to predict the next element in a sequence, but that does not yet mean it is capable of correctly carrying out a user's request. Therefore, after the initial training, the model is further trained on examples that show it instructions and the desired responses.

Who defines an acceptable answer

Then another layer of training is introduced—reinforcement learning, including reinforcement learning from human feedback (RLHF). As Arakelyan explains, for a single question, several response variants are collected, and people then evaluate them, indicating which one is more acceptable or preferable. Based on such data, the model learns not simply to continue text with high probability, but to choose a response that, according to human evaluation, better fulfils the task.

It is at this stage that the question arises of who, and by what criteria, determines which response is considered "correct," acceptable, or desirable. According to Arakelyan, differences in such evaluations can manifest in how the model refuses to discuss certain topics, how rigidly it adheres to particular logic, or how it displays certain cultural and ideological stances. Therefore, the model's characteristics depend not only on the initial training data but also on the people and criteria involved in the subsequent stages of its training.