Reproduce “Emerging Human-like Strategies for Semantic Memory Foraging in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eric Lacosse, Mariana Duarte, Peter M. Todd, Daniel C. McNamee
Researchers investigated semantic memory retrieval in Large Language Models, demonstrating that LLMs exhibit human-like cognitive strategies for convergent and divergent search during semantic fluency tasks. The study identified internal computational signatures for these strategies within LLM architectures, revealing strong behavioral alignment with humans and high decodability from internal representations using mechanistic interpretability techniques.
Key insights
- LLMs spontaneously develop strategic semantic foraging behaviors (clustering and switching) that are strikingly analogous to human cognitive search, suggesting a deeper functional alignment than simple linguistic mimicry.
- Internal computational signatures for both convergent (within-category) and divergent (between-category) semantic search are identifiable and localizable within LLM architectures, predominantly manifesting in mid-to-late layers.
- LLMs can serve as powerful scientific instruments for testing cognitive theories, and understanding their cognitive alignment/disalignment opens avenues for designing AI systems that either seamlessly collaborate with humans or offer complementary cognitive strengths.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.