Revealing emergent human-like conceptual representations from language prediction

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• Humans form concepts through experience.
• Large language models (LLMs) show human-like behaviors despite text-only training.
• The nature of LLM conceptual representations is unknown.

Purpose of the Study:

• Investigate LLM concept formation during in-context inference.
• Determine how LLMs represent and organize concepts.
• Assess the relationship between LLM concepts and behavior.

Main Methods:

• LLMs performed concept inference tasks with linguistic descriptions and contextual cues.
• Analyzed LLM internal representations for structure and convergence.
• Correlated representations with LLM task performance, human behavior, and neural data.

Main Results:

• LLMs flexibly derived concepts from text and context.
• Representations converged to a shared, context-independent structure.
• Alignment with this structure predicted LLM performance and human judgments.
• LLM representations aligned with human brain activity.

Conclusions:

• Structured, human-like conceptual representations emerge in LLMs from language prediction alone.
• LLMs provide insights into human conceptual structure and intelligence.
• Findings support advancing alignment between artificial and human intelligence.