Latent Structure Benchmark: A New Paradigm for Cultural Domain Analysis Using Large Language Models as 'Participants'

Latent Structure Benchmark is an open-source benchmark project that applies cultural domain analysis methods to large language models (LLMs). Its core innovation lies in treating AI as human participants, using systematic methods to reveal how models organize and understand everyday vocabulary, covering multiple fields such as AI safety, cognitive science, and model alignment, and providing a new perspective for understanding the 'worldview' of AI.

Cultural Domain Analysis (CDA) is a classic anthropological method used to study how cultural groups organize concepts in specific domains. Traditionally, it collects data from human participants through interviews and questionnaires to map cognitive maps. This project proposes applying this method to LLMs, exploring the similarities and differences between models and humans in concept understanding, and filling the gap in AI research regarding the analysis of internal concept structures of models.

Guiding Protocols

• Free listing: Ask models to list relevant vocabulary
• Pairwise comparison: Compare concept similarity to construct distance matrices
• Heap sorting: Classify vocabulary to reveal logic
• Semantic differential scale: Evaluate multi-dimensional perception of concepts

Data Analysis

• Consensus analysis: Test consistency of model responses
• Multidimensional scaling: Visualize concept relationships
• Hierarchical clustering: Discover natural groupings of concepts
• Network analysis: Draw concept association maps

Model Comparison

• Differences in concept domain organization between GPT and Llama series
• Changes in concept structure between base models and aligned models
• Complexity differences among models of different sizes

Traces of Alignment Process

How alignment changes the concept structure of models (original cognition vs. normative cognition) is crucial for AI safety and alignment research.

Revelation of Implicit Structure

The project makes the implicit structure of collective cognition in training data explicit through guiding protocols, helping to understand the concept organization methods inherited by models.

Application Scenarios

• AI safety: Detect potential risks (e.g., proximity between harmful and normal concepts)
• Cross-cultural AI: Study differences in concept organization among models of different languages/cultures
• Cognitive science: Provide new perspectives for human cognitive research

Open Science

The project adheres to the principles of open data, reproducibility, and transparent methods, releasing all test data, code, and process documents, and establishing a cross-model comparison benchmark.

Method Boundaries

LLMs do not have true 'cognition'; their responses are based on probability matching, so inferring cognitive structures from results requires caution.

Impact of Prompt Engineering

Different questioning methods may lead to different results; designing neutral guiding protocols is an ongoing challenge.

Dynamicity Issue

LLM outputs are random and change with version updates, making it difficult to establish a stable benchmark.

Future Outlook

• Expand to more concept domains and languages
• Develop LLM-specific guiding methods
• Establish a cross-model concept structure database
• Explore the relationship between concept structure and model capabilities

Conclusion

This project pioneers a new paradigm for understanding AI: probing the concept organization of models from within. It reveals that AI carries specific cultural perspectives, which is of great significance for the responsible use and development of AI, and serves as a powerful tool for AI researchers.