Research map / Large language models / Language, safety and interpretability
Large Language Models as Model Organisms for Human Associative Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts a neuroscience associative-learning paradigm to in-context learning in six open LLMs, finding a U-shaped (NMPH-consistent) representational differentiation during the consolidation phase that is amplified by "vocabulary interference".
Problem
Testing how associative learning reshapes internal representations (integration vs. differentiation as a function of pre-learning similarity, per the Non-Monotonic Plasticity Hypothesis) is hard in biological systems; the paper asks whether LLMs show analogous dynamics and whether global vocabulary
Approach
Present token pairs repeatedly in-context (no weight updates), measure change in last-layer cosine similarity between paired tokens relative to first occurrence, select pairs via GCG-style gradient search across the similarity spectrum, and define vocabulary interference as the median similarity of a paired token to a sampled vocabulary subset.
Key results
- Mid-similarity token pairs (about 0.55-0.75 pre-learning cosine similarity) significantly differentiate during the consolidation phase, producing a U-shaped pattern consistent with NMPH across six models.
- Higher vocabulary interference is associated with greater differentiation; under controlled sampling, curves for high-interference pairs lie below low-interference curves across similarity levels.
- Models reach 90-100% accuracy on the in-context association task; only Llama2-7b and Mistral-7b show a forgetting phase (Llama2-7b at r=40, Mistral-7b at r=3000).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Repeatedly present associated token pairs through in-context learning | Six open-source base LLMs · Associative learning task | accuracy: 90-100% | Helped |
| Learn moderately similar pairs during Consolidation | Six open-source base LLMs · Token pairs with before-learning similarity $0.55-0.75$ | pairwise representational similarity | Hurt |
| Learn low-similarity pairs during Encoding | Six open-source base LLMs · Token pairs with before-learning similarity between $0.1$–$0.6$ | pairwise representational similarity | Helped |
| Learn high-similarity pairs during Encoding or Consolidation | Six open-source base LLMs · High-similarity token pairs | pairwise representational similarity | No clear effect |
| Continue repeated pair exposure into the Forgetting phase | Llama2-7b and Mistral-7b · Associative learning task | accuracy: $r=40$; $r=3,000$ | Hurt |
| Increase vocabulary interference | Associative learning task, extended token-pair set · Mid- and high-similarity pairs | pairwise representational similarity after learning | Hurt |
| Increase vocabulary interference | Associative learning task, extended token-pair set · Low-similarity pairs | pairwise representational similarity after learning | Mixed |
| Increase vocabulary interference | Associative learning task, original token-pair set · High-similarity pairs | pairwise representational similarity after learning | Mixed |
| Compare representation changes across similarity groups during Forgetting | Six open-source base LLMs · Moderately similar pairs | pairwise representational similarity | No clear effect |
Surprising findings
- The non-monotonic differentiation appears only during consolidation and disappears during encoding and forgetting phases.
- Forgetting in Llama2-7b coincides with pair shifts toward the peak of the vocabulary interference distribution (speculative).
Limitations
- Token pairs are defined by geometric similarity rather than naturalistic semantics; WordNet analysis is only preliminary.
- Vocabulary interference uses a random 1,000-token subset, which may underrepresent full-vocabulary competition.
- LLM results are not direct stand-ins for human brains and require empirical validation in humans.
Experimental setup
Datasets: WordNet
Metrics: cosine similarity change between paired token representation associative recall accuracy paired t-test with Benjamini-Yekutieli correction
Baselines: Hebbian learning prediction pattern separation prediction
Builds on: Non-Monotonic Plasticity Hypothesis Greedy Coordinate Gradient in-context learning
Closest papers by meaning
- Scaling and context steer LLMs along the same computational path as the human brain (NeurIPS 2025 · similarity 0.87)
- Scaling Laws for Associative Memories (ICLR 2024 · similarity 0.87)
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.87)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.87)
- Unveiling the Learning Mind of Language Models: A Cognitive Framework and Empirical Study (NeurIPS 2025 · similarity 0.87)
- Linking In-context Learning in Transformers to Human Episodic Memory (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.