Conceptual blending in humans and language models

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Linguistics

Background:

• Concept blending is crucial for human conceptual flexibility, integrating source concepts into context-specific meanings.
• Large language models (LLMs) exhibit blending, prompting investigation into brain versus model mechanisms.
• Human blending is linked to embodied cognition and goal-directed perception of environmental affordances.

Purpose of the Study:

• To analyze the mechanisms of concept blending in humans and LLMs.
• To propose a predictive processing account for human generative-causal blends.
• To contrast human embodied blending with LLM-generated blends.

Main Methods:

• Conceptual analysis comparing human and LLM blending.
• Development of a predictive processing model for human blending.
• Illustration of the model with a vector symbolic architecture example.
• Analysis of LLM blending mechanisms based on next-token prediction and proxy-goals.

Main Results:

• Human blending is framed as a goal-driven predictive process rooted in embodied perception.
• LLMs produce blends constrained by distributional novelty and prompt-derived proxy-goals.
• Key distinctions emerge in conceptual flexibility and agent autonomy between humans and LLMs.

Conclusions:

• Human and LLM blending mechanisms differ significantly, impacting conceptual flexibility and autonomy.
• Understanding these distinctions is vital for designing more capable and autonomous AI agents.
• The predictive processing account offers a framework for understanding embodied concept integration.