Uncovering the Computational Ingredients of Human-Like Representations in LLMs

TL;DR

This work investigates what computational ingredients drive human-like semantic representations in large language models. By employing a large-scale triadic similarity framework anchored to the THINGS concept set, it compares 70+ open-weight transformers against human semantic embeddings derived from SPoSE, enabling a direct representational alignment analysis. The study finds that instruction-tuning and higher architectural dimensionality strongly predict human alignment, whereas overall size and multimodal pretraining have limited independent impact. It also shows that existing benchmarks only partially reflect alignment, revealing a benchmarking gap and providing concrete guidance for designing cognitively grounded LLMs.

Abstract

The ability to translate diverse patterns of inputs into structured patterns of behavior has been thought to rest on both humans' and machines' ability to learn robust representations of relevant concepts. The rapid advancement of transformer-based large language models (LLMs) has led to a diversity of computational ingredients -- architectures, fine tuning methods, and training datasets among others -- but it remains unclear which of these ingredients are most crucial for building models that develop human-like representations. Further, most current LLM benchmarks are not suited to measuring representational alignment between humans and models, making benchmark scores unreliable for assessing if current LLMs are making progress towards becoming useful cognitive models. We address these limitations by first evaluating a set of over 70 models that widely vary in their computational ingredients on a triplet similarity task, a method well established in the cognitive sciences for measuring human conceptual representations, using concepts from the THINGS database. Comparing human and model representations, we find that models that undergo instruction-finetuning and which have larger dimensionality of attention heads are among the most human aligned, while multimodal pretraining and parameter size have limited bearing on alignment. Correlations between alignment scores and scores on existing benchmarks reveal that while some benchmarks (e.g., MMLU) are better suited than others (e.g., MUSR) for capturing representational alignment, no existing benchmark is capable of fully accounting for the variance of alignment scores, demonstrating their insufficiency in capturing human-AI alignment. Taken together, our findings help highlight the computational ingredients most essential for advancing LLMs towards models of human conceptual representation and address a key benchmarking gap in LLM evaluation.

Figures (13)

• Figure 1: Method for estimating human-model alignment. We collected 35k triplet similarity judgments from each of the 77 models in our suite. We computed semantic embeddings based on these judgments and compared the representational geometry of model embeddings to human embeddings derived from hebart2023things.
• Figure 2: Computational ingredients predictive of human model alignment. A., B., C., and D. show the relationship between instruction-tuning, MLP dimensionality, context length, and embedding dimensionality on Procrustes $R^2$. These were the four ingredients most predictive in the mixed-effects regression model. E.$t$-scores for each predictor, which highlight the relative contribution of each ingredient towards alignment when considered in a single statistical model.
• Figure 3: Effects of post-training on alignment. The black points show mean alignment (across models) at each post-training stage.*Instella only.
• Figure 4: Relationship between alignment and other LLM benchmarks. A.BigBenchHard scores for each model in our suite vs. their Procrustes $R^2$ value w.r.t. human embeddings. B. Correlation between alignment and the six LLM benchmarks evaluated.
• Figure 7: Correlations between measures of human-model alignment and model benchmarks, for models where benchmark scores are publicly are available. (Procrustes $R^2$, CKA, RSM)