TLDR: A study evaluated 77 large language models on a human-like conceptual representation task, finding that instruction-finetuning and larger architectural dimensions significantly improve alignment with human thinking. Factors like multimodal pretraining, activation functions, and vocabulary size had limited impact. Existing LLM benchmarks do not fully capture this human-AI representational alignment.
Large Language Models (LLMs) have made incredible strides in understanding and generating human language, leading many to wonder if these powerful AI systems are developing internal representations similar to our own. This question is crucial for both advancing AI and for cognitive science, which seeks to understand human thought. However, current ways of evaluating LLMs often fall short in measuring this ‘human-likeness’ in their internal workings.
A recent research paper, UNCOVERING THE COMPUTATIONAL INGREDIENTS OF HUMAN-LIKE REPRESENTATIONS IN LLMS, by Zach Studdiford, Timothy Rogers, Kushin Mukherjee, and Siddharth Suresh, tackles this challenge head-on. The researchers set out to identify which specific design choices and training methods – or ‘computational ingredients’ – are most important for building LLMs that develop representations aligned with human conceptual understanding.
How They Explored Human-Like Representations
To measure how human-like an LLM’s representations are, the team used a method well-established in cognitive science: the triplet similarity task. In this task, participants (or models) are shown three items and asked to identify which two are most similar, or which one is the ‘odd one out’. This helps map out how concepts are related in a high-dimensional space. For this study, they used concepts from the THINGS database, a large collection of object concepts with rich human behavioral data.
The researchers evaluated a diverse suite of 77 open-source LLMs, varying widely in their computational ingredients. These included differences in model architecture, activation functions, whether they were pre-trained on multimodal data (like images and text), their parameter size, the amount of training data, and the dimensionality of their internal components like attention heads and Multi-Layer Perceptrons (MLPs).
Key Discoveries: What Drives Alignment with Human Thinking
The study yielded several fascinating insights into what makes LLMs think more like humans:
-
Instruction-Finetuning is Key: Models that underwent instruction-finetuning – meaning they were specifically trained to follow instructions – showed significantly better alignment with human representations. This suggests that the process of teaching models to understand and execute commands also helps them organize their internal knowledge in a more human-like way.
-
Architectural Depth Matters: The dimensionality of certain architectural components, particularly MLPs and attention heads, had a strong positive relationship with human-model alignment. Essentially, models with more ‘expressive’ internal spaces tended to be more human-aligned. This indicates that the capacity for rich, complex internal processing is important.
-
Larger Models and More Data Help: In line with some expectations, larger models (in terms of parameter count) and those trained on more data generally exhibited greater human-like representations. Similarly, models with longer maximum context lengths also showed better alignment, possibly because the techniques enabling long-context reasoning also enhance conceptual understanding.
What Didn’t Seem to Make a Difference (or even hindered)
Not all computational ingredients had a positive impact on alignment:
-
Multimodal Pretraining’s Limited Role: Surprisingly, pre-training models on both images and text did not independently improve human-model alignment. In fact, when other factors were accounted for, multimodal pretraining sometimes predicted reliably lower alignment, suggesting that current multimodal training approaches might not always foster more human-like conceptual structures.
-
Activation Functions and Vocabulary Size: The choice of activation function (e.g., SwiGLU, GeGLU) and the size of the model’s vocabulary had no significant bearing on alignment.
Alignment and Existing Benchmarks
The study also examined how this new measure of representational alignment correlated with performance on standard LLM benchmarks (like MMLU, BigBenchHard, and IFEval). While some benchmarks, particularly those probing broad semantic knowledge and instruction-following, showed correlations with alignment, no single existing benchmark fully captured the variance in human-model representational alignment. This highlights a significant gap in current LLM evaluation methods, suggesting that simply optimizing for benchmark scores might not lead to truly human-aligned AI.
Also Read:
- AI’s Unexpected Leap in Visualizing Concepts
- AI Models Show Strong Potential for Automating Metaphor Identification in Texts
Conclusion
The findings from this research underscore the critical role of instruction-finetuning and the dimensionality of architectural components in fostering human-like conceptual representations in LLMs. It also reveals that while model scale and training data amount are important, other factors like multimodal pretraining might need re-evaluation in the context of human alignment. This work provides valuable guidance for developing LLMs that not only perform tasks well but also think in ways that resonate more closely with human cognition.
