Human-like object concept representations emerge naturally in multimodal LLMs
a year ago
- #AI cognition
- #neuroscience
- #machine learning
- The study explores how large language models (LLMs) and multimodal LLMs develop human-like object representations.
- Researchers collected 4.7 million triplet judgments from LLMs to derive 66-dimensional embeddings for 1,854 natural objects.
- The embeddings showed semantic clustering similar to human mental representations and were interpretable.
- Model embeddings aligned with neural activity patterns in brain regions like the extrastriate body area and fusiform face area.
- Findings suggest LLMs develop conceptual representations that share fundamental similarities with human cognition.
- The research advances understanding of machine intelligence and informs the development of human-like AI systems.
- Data and code are publicly available, supporting further research in this area.