A Nurse is Blue and Elephant is Rugby: Cross Domain Alignment in Large Language Models Reveal Human-like Patterns

Abstract: Cross-domain alignment refers to the task of mapping a concept from one domain to another. For example, ``If a \textit{doctor} were a \textit{color}, what color would it be?''. This seemingly peculiar task is designed to investigate how people represent concrete and abstract concepts through their mappings between categories and their reasoning processes over those mappings. In this paper, we adapt this task from cognitive science to evaluate the conceptualization and reasoning abilities of LLMs through a behavioral study. We examine several LLMs by prompting them with a cross-domain mapping task and analyzing their responses at both the population and individual levels. Additionally, we assess the models' ability to reason about their predictions by analyzing and categorizing their explanations for these mappings. The results reveal several similarities between humans' and models' mappings and explanations, suggesting that models represent concepts similarly to humans. This similarity is evident not only in the model representation but also in their behavior. Furthermore, the models mostly provide valid explanations and deploy reasoning paths that are similar to those of humans.

• The paper demonstrates that LLMs achieve significant cross-domain mapping performance, with Llama-7B reaching a 24.3% M@1 score.
• The methodology uses majority voting and greedy decoding to consolidate model responses, aligning them with human conceptual patterns.
• Their human-like explanations, with up to 87% sensibility, pave the way for more transparent and interpretable AI systems.

Can LLMs Think Like Us? A Dive into Cross-Domain Alignment

Introduction

Ever wondered if LLMs can map concepts across different domains—like associating "doctor" with a color—and explain their reasoning like humans do? This paper aims to explore just that. Researchers adapted a psychological task to assess how well LLMs can conceptualize and reason about different domains. They dug into how these models behave similarly to humans and whether they can explain their thought processes in human-understandable terms.

Cross-Domain Mapping

Experimental Setup

Dataset: The dataset used is an intriguing collection of human responses to questions like "If a doctor were a color, what color would it be?" involving twelve different domains.

Models: They tested seven LLMs, including Flan and Llama models, and used prompts to see how these models perform on the cross-domain mapping task.

Prompting Methodology: Different templates ensured the model’s responses were consolidated using a majority vote system. They even employed greedy decoding to approximate the most probable response from the models.

Metrics: They used Match at K (M@K) metrics, i.e., checking if the model's answer matches the top K popular human answers. They also considered the extent of agreement among human participants to deterministically gauge model behavior.

Results

LLMs performed cross-domain mappings at a level significantly above random guess rates. They showed substantial M@1 and M@3 scores, notably with Llama-7B scoring the highest (M@1: 24.3%, M@3: 36.5%). Interestingly, larger models didn't always score higher, suggesting model size isn't the sole determinant of performance.

At the individual level, some LLMs performed better than the average human participant, indicating that some models are more aligned with typical human behavior.

Explanations for Mappings

The researchers assessed the quality of explanations provided by the LLMs for their mappings to understand if they employ similar reasoning to humans.

Model Explanations: Models like Llama and Mistral-7B generated explanations for their mappings. The explanations were tested for validity and coherence.

Explanation Categories: They categorized explanations into seven types such as perceptual similarity, word associations, and thematic associations. The output from LLMs closely mirrored the variety found in human explanations, suggesting they use similar conceptual reasoning.

Explanation Quality: LLMs produced mostly sensible explanations. Mistral-7B, for instance, achieved an 87% sensible explanation rate.

Practical and Theoretical Implications

Practical Implications

1. Improving LLM Utility: The ability of LLMs to perform cross-domain mappings and offer human-like explanations can significantly enhance user interaction with AI. This could improve personalized services, educational tools, and automated content generation.
2. Validation for Task Assignments: The alignment of models with human behavior at both population and individual levels offers a validation layer for AI-driven decision-making processes.

Theoretical Implications

1. Insight into Conceptual Representation: The study offers valuable insight into how LLMs conceptualize different domains, reinforcing the idea that LLMs can mirror some aspects of human cognitive processes.
2. Future of AI Explainability: By demonstrating that LLMs can provide human-like explanations, this research paves the way for more transparent and interpretable AI systems, a crucial step toward wider acceptance and reliance on AI in sensitive fields.

Speculations for Future Developments

1. Enhanced Training Data: Future models might benefit from even larger and more varied datasets to further close the gap between human and model responses.
2. Advanced Semantic Matching: Incorporating more sophisticated semantic matching techniques could elevate the models' ability to produce human-like answers in alignment tasks.

Conclusion

While LLMs are far from being indistinguishable from humans in their explanations, this paper illustrates that they exhibit surprisingly human-like behavior in cross-domain alignments. These findings point to a promising future where LLMs could serve as reliable tools for various applications requiring nuanced understanding and explanation.

The nuances revealed in this paper provide an exciting glimpse into the cognitive capabilities of LLMs and lay the groundwork for further advancements in the field of AI explainability and alignment with human cognition.