TL;DR
This paper introduces a novel "fingerprinting" method to analyze how Large Language Models (LLMs) internally represent psychological constructs, using the Humor Style Questionnaire (HSQ) as a case study. The findings reveal a significant divergence between LLM and human response patterns, suggesting LLMs organize concepts fundamentally differently, raising questions about their use in sensitive psychological contexts.
What Problem Does This Paper Address?
As LLMs become increasingly integrated into human-AI interactions, especially in areas like psychological assessments, understanding their internal representation of complex human constructs is crucial. Traditional evaluation metrics like output accuracy or faithfulness don't capture how LLMs internally relate these constructs. This gap is critical because if LLMs organize psychological concepts differently than humans, it could lead to unexpected failures or misinterpretations when deployed in real-world scenarios, particularly in sensitive domains like mental health. The authors aim to bridge this gap by developing a method to "fingerprint" LLMs based on their internal construct representations.
Methods
The core of the paper is a novel "fingerprinting" approach that analyzes the factor correlation patterns within LLM responses to a psychological assessment. Instead of just looking at individual answers, they examine how the LLM implicitly relates different items on the questionnaire.
- Fingerprinting: For each LLM, they generate 1000 synthetic responses to the 32-item Humor Style Questionnaire (HSQ). They then compute a pairwise correlation matrix (Cx) between all HSQ items across these responses. This matrix serves as the LLM's "fingerprint" of psychological construct organization.
- Similarity Measurement: To compare fingerprints, they vectorize the upper triangular part of the correlation matrices and compute the cosine similarity between them. This yields a score indicating how similar two fingerprints are.
- Exploratory Graph Analysis (EGA): To further understand the latent structure, they apply EGA to the correlation matrices. EGA identifies communities (clusters) of items, revealing how the LLM groups related concepts. This is compared against the known theoretical four-factor structure of the HSQ.
- Baselines: They compare LLM fingerprints against:
- Random: Randomly generated responses to establish a lower bound for structural coherence.
- Human Items: Synthetic responses preserving item-level distributions but disrupting factor structure.
- Human Full: A dataset of 1071 human HSQ responses from a previous study, serving as the gold standard.
- Ablation Study (Context Window): They also conduct an experiment without a conversational context window to assess its impact on LLM responses.
- Cronbach's Alpha: As a complementary validation, they calculate Cronbach's alpha to measure the internal consistency of responses for each humor style dimension.
Datasets
- Humor Style Questionnaire (HSQ): A 32-item questionnaire measuring four dimensions of humor: Affiliative, Self-enhancing, Aggressive, and Self-defeating. This is used as the assessment tool.
- Human HSQ Responses: A dataset of 1071 human responses from Martin et al. (2003) is used as the primary human baseline.
- LLM Models: Six open-weight LLMs were used: Llama 3.1 8B/70B, Mistral 7B/123B, and Qwen 2.5 7B/72B.
Key Insights and Ideas
- Factor Covariance as a Fingerprint: The core idea is that the relationships between items, as captured by correlation matrices, are more revealing of internal construct representation than raw output scores. This provides a stable and distinctive "fingerprint."
- Human vs. LLM Divergence: The most striking insight is the profound difference in how LLMs and humans organize psychological constructs. This suggests LLMs don't "think" about these concepts in the same way humans do, even if their outputs seem plausible.
- Context Matters (for LLMs): The ablation study shows that conversational continuity (a sliding window of previous questions) significantly improves LLM response coherence and reliability, highlighting the importance of interaction history for LLMs.
- EGA Reveals Structural Differences: Exploratory Graph Analysis visually confirms that LLMs fail to recover the established four-factor structure of the HSQ, instead producing fragmented community structures.
Key Results
- Low Human-LLM Similarity: Cosine similarity scores between LLM fingerprints and human fingerprints were consistently low, averaging only 0.026. Even the highest similarity was far below human-to-human consistency.
- Within-Family vs. Cross-Family Similarity: While some within-family similarities were observed (e.g., Mistral models), surprisingly, cross-family similarities between models from different developers (e.g., Qwen and Llama) often exceeded within-family ones. This suggests factors beyond architecture, like training data or optimization, play a significant role.
- Size vs. Consistency: Larger models (70/123B) showed slightly higher similarity to each other than smaller models (7/8B), suggesting parameter count might lead to more consistent representations, but this doesn't necessarily mean more human-like representations.
- Context Improves LLM Reliability: The 5-item context window significantly improved LLM response reliability (Cronbach's alpha) across most models, though still falling short of human levels.
- EGA Findings: Human baselines consistently recovered the four HSQ factors. LLMs produced between 2-8 communities that did not align with these theoretical dimensions, indicating blurred conceptual boundaries.
Limitations and Future Work
- Training Data Contamination: LLMs might have been exposed to the HSQ or similar instruments during training, potentially inflating similarity scores. Future work could use novel instruments not present in training data.
- API/Prompting Variations: Subtle differences in API implementations, prompting, and generation parameters could bias responses. The authors made their code public for transparency.
- Single Instrument: The study used only the HSQ. Future research should explore other psychological instruments (e.g., personality, attitudes) to see if these findings generalize.
- Context Implementation: The sliding window approach is one of many ways to manage context. Future work could systematically vary context management strategies.
- Future Work:
- Extend the fingerprinting methodology to other psychological instruments.
- Investigate how different prompt engineering techniques affect factor correlation patterns.
- Conduct longitudinal studies to track how these patterns evolve across model generations.
- Explore new training methodologies or architectural innovations to better align LLM representations with human psychological constructs.
Summary updated