Revealing emergent human-like conceptual representations from language prediction

TL;DR

This work shows that next-token language prediction alone can give rise to human-like conceptual representations in LLMs. By inferring concepts from definitional descriptions in-context, models develop a shared, context-independent structure that predicts performance across tasks and aligns with human behavior and brain activity. The findings bridge symbolic and connectionist accounts of concepts, demonstrating a meaningful, brain-relevant organization that emerges without real-world grounding, while also highlighting gaps in visually grounded perceptual properties. Overall, the study positions LLMs as a window into human concepts and a route toward improving alignment between artificial and human intelligence, with clear avenues for grounding and cross-modal integration.

Abstract

People acquire concepts through rich physical and social experiences and use them to understand and navigate the world. In contrast, large language models (LLMs), trained solely through next-token prediction on text, exhibit strikingly human-like behaviors. Are these models developing concepts akin to those of humans? If so, how are such concepts represented, organized, and related to behavior? Here, we address these questions by investigating the representations formed by LLMs during an in-context concept inference task. We found that LLMs can flexibly derive concepts from linguistic descriptions in relation to contextual cues about other concepts. The derived representations converge toward a shared, context-independent structure, and alignment with this structure reliably predicts model performance across various understanding and reasoning tasks. Moreover, the convergent representations effectively capture human behavioral judgments and closely align with neural activity patterns in the human brain, providing evidence for biological plausibility. Together, these findings establish that structured, human-like conceptual representations can emerge purely from language prediction without real-world grounding, highlighting the role of conceptual structure in understanding intelligent behavior. More broadly, our work suggests that LLMs offer a tangible window into the nature of human concepts and lays the groundwork for advancing alignment between artificial and human intelligence.

Figures (28)

• Figure 1: Illustration of the reverse dictionary task as a conceptual probe. A Transformer-based LLM is presented with $N$ description--word pairs as demonstrations in context, followed by a query description. The model is then prompted to encode the query description into a conceptual representation and predict the term that best matches the described concept.
• Figure 2: Performance of LLaMA3-70B on the reverse dictionary task measured through exact match accuracy. Black lines show performance when the model is given $N$ correct demonstrations sampled from the training set and evaluated on an independent test set. Blue and red lines show performance when one misleading demonstration---a description paired with a proxy label (capital letter, random string, or random word, as specified in the figure titles)---is added to $N - 1$ correct demonstrations of other concepts. The model is queried with the description identical to the misleading one. The blue line shows the frequency with which the model reproduces the proxy label, while the red line indicates how often it generates the correct word for the query concept given contextual information from other concepts. Shaded areas denote $95\%$ bootstrapped confidence intervals, calculated from 10,000 resamples over five independent runs.
• Figure 3: LLMs converge toward a similar representational structure of concepts. (A) Pairwise alignment (RSA) of LLaMA3-70B conceptual representations across different contextual demonstrations. Axes indicate the number of demonstrations, and each cell shows the alignment from a single run. (B) LLM performance on the reverse dictionary task reflects alignment with the representations formed with 120 demonstrations. Each point corresponds to a representation space formed with $N$ demonstrations, with the x-axis indicating exact match accuracy and the y-axis denoting alignment with the 120-demonstration space. Error bars indicate $95\%$ confidence intervals from 10,000 bootstrap resamples across five runs. (C) A t-SNE visualization of conceptual representations formed by different LLMs with 24 demonstrations. Distances are calculated as $1 - \textrm{alignment}$, averaged over five runs. Each point corresponds to an LLM, plotted in proportion to its complexity and color-coded by exact match accuracy on the reverse dictionary task. Better-performing models (blue) exhibit more aligned representations. (D) Alignment with LLaMA3-70B conceptual representations predicts model performance across downstream tasks. The x-axis shows the alignment with LLaMA3-70B representations, and the y-axis indicates overall performance, computed as accuracy averaged across a range of language understanding and reasoning tasks. Each point represents a different model and is color-coded by model complexity. Linear fits are shown as straight lines, with shaded areas representing $95\%$ confidence intervals derived from 10,000 bootstrap resamples. Overall, models with higher complexity align more closely with the LLaMA3-70B. An exception is Qwen2-0.5B, which shows relatively high complexity but low alignment, possibly due to constraints in model scale or training data quality.
• Figure 4: Alignment between LLM-derived conceptual representations and psychological measures of similarity. (A--B) Performance of LLaMA3-70B conceptual representations compared with static word embeddings (FastText) and sense vectors (DeConf) in predicting human similarity judgments. (A) Spearman's rank correlation with human similarity ratings for concept pairs from SimLex-999. (B) Prediction accuracy for human triplet odd-one-out judgments in THINGS. The noise ceiling reflects the upper bound of performance based on inter-subject consistency. Error bars denote $95\%$ confidence intervals calculated from five independent runs. (C--E) t-SNE visualizations of LLM-derived conceptual representations (C), static word embeddings (D) and sense vectors (E). Data points are color-coded by human-labeled categories from THINGS. Conceptual representations derived from LLaMA3-70B exhibit a clearer alignment with category structure than word embeddings or sense vectors.
• Figure 5: Performance of conceptual representations derived from LLaMA3-70B in predicting context-dependent human ratings across 52 category--feature pairs. Scatter plots illustrate the relationship between predicted ratings from conceptual representations (x-axis) and the average human ratings (y-axis). Linear fits are shown as straight lines, with shaded regions representing $95\%$ confidence intervals derived from 10,000 bootstrap resamples. Category--feature pairs with statistically significant correlations (Spearman's rank correlation, FDR $P < 0.01$) are displayed against a white background.