TLDR: Large Language Models (LLMs) are being used as “model organisms” to study how humans form associations. Researchers found that LLMs exhibit a “U-shaped” pattern of memory reorganization, where moderately similar items become more distinct after learning, consistent with a theory called the Non-Monotonic Plasticity Hypothesis. This process is also influenced by “vocabulary interference,” meaning how much new associations compete with existing knowledge in the model, with higher interference leading to greater differentiation.
Understanding how our brains form connections between co-occurring items, a process known as associative learning, is fundamental to human cognition. This learning reshapes our internal representations in incredibly complex ways. However, directly testing hypotheses about these representational changes in biological systems like the human brain is notoriously difficult. Challenges include precisely controlling the similarity between items before learning, the need for extensive data sampling, and practical limitations like cost and participant fatigue.
A recent study, titled “Large Language Models as Model Organisms for Human Associative Learning” by Camila Kolling, Vy Ai Vo, and Mariya Toneva from the Max Planck Institute for Software Systems, proposes a novel approach: using large language models (LLMs) as scalable alternatives to investigate these intricate memory dynamics. The researchers argue that LLMs, with their ability to learn rapidly within context (in-context learning), offer a controllable and accessible computational model for generating new hypotheses about how memory is reorganized in the brain.
The Non-Monotonic Plasticity Hypothesis
The study focuses on three main hypotheses regarding how associative learning alters representations. The classical Hebbian learning rule suggests that repeated associations strengthen connections, leading to more integrated representations. Conversely, theories like pattern separation propose that memories differentiate to reduce overlap and minimize interference. To reconcile these seemingly opposing dynamics, the Non-Monotonic Plasticity Hypothesis (NMPH) posits a U-shaped curve of representational change: items that are either very similar or very dissimilar tend to integrate or remain stable, while moderately similar pairs differentiate after learning.
LLMs as a New Frontier for Cognitive Neuroscience
The researchers adapted a cognitive neuroscience associative learning paradigm for LLMs. They repeatedly presented token pairs (like words or sub-word units) within the model’s input context, then measured how the model’s internal representations of these tokens changed. This was done by comparing the cosine similarity of their hidden representations before and after learning. Six open-source LLMs, including various Llama models, Gemma, and Mistral, were put to the test.
The initial findings were striking: LLMs exhibited a non-monotonic pattern consistent with the NMPH. Specifically, moderately similar token pairs showed significant differentiation after learning, mirroring human-like patterns of representational change. This U-shaped curve was most prominent during the ‘Consolidation phase’ of learning, a period when the models maintained stable, high accuracy on the associative task.
The Role of Vocabulary Interference
Beyond simple pair similarity, the study introduced a crucial new factor: vocabulary interference. This concept captures how new associations compete with the model’s vast prior knowledge, which is encoded during its pre-training. Imagine trying to learn a new association for a word that is already very similar to many other words in your vocabulary; this creates high interference. The researchers found that higher vocabulary interference consistently amplified differentiation, meaning that when there was more competition from other similar items in the model’s vocabulary, the learned associations became even more distinct.
This interaction between pairwise similarity and vocabulary interference revealed a richer dynamic than previously understood. For low-similarity pairs, integration occurred regardless of interference, as there was ample room to bring them closer without confusion. For high-similarity pairs, differentiation was crucial under high interference to prevent entanglement, while under low interference, they remained stable. Mid-similarity pairs, however, proved to be a “sensitive zone” where both factors strongly interacted, leading to differentiation that intensified with greater interference.
Also Read:
- Mapping the Dynamics of Human Thought: A Quantitative Framework for Reasoning
- Enhancing LLM Knowledge Distillation with Counterfactual Explanations in Low-Data Settings
Implications and Future Directions
These results position LLMs as powerful tools for studying representational dynamics in human-like learning systems. The ability to systematically manipulate factors like pairwise similarity and vocabulary interference in LLMs offers a level of experimental control rarely achievable in human studies. This could help reconcile diverging findings in neuroscience and generate new hypotheses about memory reorganization in the brain.
While LLMs are not direct replicas of human brains, they provide a valuable platform for exploring complex cognitive phenomena. Future work could delve into how representational changes occur across different layers of the models, explore more naturalistic semantic stimuli, and investigate how these processes emerge during developmental-like learning trajectories. You can read the full research paper here: Large Language Models as Model Organisms for Human Associative Learning.
