Abstract
This research investigates whether Large Language Models (LLMs) can serve as "model organisms" to understand how the human brain reorganizes information during learning. By analyzing representational changes during in-context learning (ICL), researchers found that LLMs mirror biological activity described by the Non-Monotonic Plasticity Hypothesis (NMPH). Specifically, items with moderate initial similarity undergo differentiation—becoming more distinct—to mitigate memory interference, providing a controllable and scalable framework for cognitive neuroscience.
Introduction
When we learn new associations, such as pairing a face with a name, our brain's internal representational "map" undergoes significant reorganization. Cognitive scientists have long debated whether this process involves integration (increasing similarity) or pattern separation (increasing distinctness). Testing these theories in humans is notoriously difficult due to the challenges of controlling prior similarity and the inherent noise in neuroimaging. This study proposes that LLMs, which can learn associations rapidly through in-context learning without permanent weight updates, offer a transparent laboratory to test these biological hypotheses.
Methods
The researchers adapted a classic associative learning paradigm for LLMs, utilizing the internal vocabularies of six prominent open-source models, including the Llama series, Gemma2, and Mistral-7b. Instead of traditional training, they utilized In-Context Learning (ICL), presenting the model with repeated pairs of tokens $(x, y)$ within a single prompt. The model's "learning" is driven purely by its attention mechanism.
To measure Representational Change, the team calculated the cosine similarity between the hidden vector representations of tokens at the model's final layer before and after the learning repetitions. They employed Greedy Coordinate Gradient (GCG) to identify token pairs across the full spectrum of initial similarity. Furthermore, they introduced a "Vocabulary Interference" metric to quantify how much a new association competes with the model's pre-existing knowledge based on the density of similar tokens in its vocabulary.
Key Insights: The Non-Monotonic Plasticity Hypothesis
The study's primary finding is that LLMs adhere to the Non-Monotonic Plasticity Hypothesis (NMPH). During the consolidation phase of learning, items that are already highly distinct or highly similar tend to maintain their relative positions. However, items with moderate initial similarity undergo significant differentiation, being pushed further apart in the model's internal space.
The researchers identified three distinct phases in LLM in-context learning:
- Encoding: Rapid accuracy gains accompanied by integration (low-similarity pairs moving closer).
- Consolidation: High accuracy is maintained while representations reorganize, specifically showing differentiation for mid-similarity pairs (0.55–0.75 cosine similarity).
- Forgetting: A phase where accuracy eventually declines.
The Role of Competition
Differentiation is not merely a function of the pair itself but is driven by "global" competition. High vocabulary interference—where a token is surrounded by many similar "distractor" tokens—increases the pressure to differentiate representations to maintain accuracy. When using real-world words from WordNet, which tend to have higher natural interference, the U-shape is less pronounced, and the models show a more monotonic trend toward differentiation.
Conclusion and Future Work
By demonstrating that LLMs reorganize associations similarly to biological systems, this work establishes AI models as viable "transparent brains." While LLMs are not perfect biological replicas, their accessibility allows researchers to probe the drivers of representational change in ways impossible with human subjects. Future research will explore dynamics in earlier layers, developmental trajectories during pre-training, and the influence of semantic meaning on representational reorganization.
Summary updated