The language capabilities of today’s artificial intelligence systems are surprising. Now we can have natural conversations with Chatgpt, Gemini and many other systems and have almost the same fluency as humans. However, we still know very little about the internal processes in these networks, which have led to such excellent results.
One published in Journal of Statistical Mechanics: Theory and Experiment ((Jat) reveals part of this mystery. It shows that when training is performed using a small amount of data, neural networks initially rely on the position of words in the sentence. However, as the system is exposed to enough data, it transitions to new strategies based on the meaning of the word. The study found that once critical data thresholds are crossed, this transition suddenly occurs – just like a phase change in a physical system. These findings provide valuable insights into understanding the functioning of these models.
Like a child who learns to read, neural networks start by understanding sentences based on word position: depending on the position of words in the sentence, the network can infer their relationship (are they subjects, verbs, objects?). However, as the training continues – the Internet “continuously goes to school” – changes occur: the meaning of words becomes the main source of information.
This is what happens in a simplified self-acting mechanism model – the core building blocks of the transformer language model, such as the models we use every day (Chatgpt, Gemini, Claude, etc.). Transformers are neural network architectures designed to process sequences of data such as text and form the backbone of many modern language models. Transformers specifically understand relationships in sequences and use self-acting mechanisms to evaluate the importance of each word relative to other words.
“To evaluate the relationship between words, the network can use two strategies, one of which is to take advantage of the position of words,” explains Hugo Cui, a postdoctoral researcher at Harvard University. For example, in languages like English, topics usually precede verbs before verbs, and verbs precede objects. “Mary Eats Apples” is a simple example of this sequence.
“This is the first strategy that appears spontaneously when networks are trained,” Cui explained. “However, in our study, we observed that if the training continues and the network receives enough data at some point (once it crosses the threshold), the policy suddenly shifts: the network starts relying on meaning.”
“When we designed this work, we just wanted to study what strategies or hybrids of strategies the network would adopt. But what we found was surprising: Under a certain threshold, the network relies solely on location, and above that location, just on meaning.”
CUI describes this transition as a phase transition, borrowing a concept from physics. Statistical physics research systems are composed of large numbers of particles (such as atoms or molecules) that statistically describe their collective behavior. Similarly, neural networks (the basis of these AI systems) are composed of a large number of “nodes” or neurons (named similar to the human brain), each connected to many other people and performing simple operations. The intelligence of the system comes from the interactions of these neurons, which can be described statistically.
This is why we can say that network behavior is a sudden change in phase transition, similar to the change of water under temperature and pressure conditions, changing from liquid to gas.
Cui stressed: “It is important to understand the strategy shift from a theoretical perspective that happens in this way.” “Our networks are simplified compared to the complex models people interact with every day, but they can prompt to start understanding the conditions that lead to the model stabilization of one strategy or another. Hopefully, this theoretical knowledge can be used in the future to make the use of neural networks more efficient and safer.”
The study by Hugo Cui, Freya Behrens, Florent Krzakala and Lenka Zdeborová is titled “Stage Transition between Position and Semantic Learning in Resolvable Dot-Prododuct Poasine” is published in JSTAT as part of machine learning as part of machine learning and included in the 2025 Special Isseage and included in Neurips 2024 Cression Inscement in Neurips 2024 Chardings.
