The Grounding Gap

In the property-generation paradigm, participants — humans in the original studies, models in ours — are given a concept and asked to list the properties, situations, or associations that come to mind. Individual answers vary, but the coded distributions across human participants are highly consistent: people anchor abstract concepts along grounding dimensions such as sensorimotor experience, emotion, and social context in a stable way.

Relative to humans, frontier models generate substantially more Verbal Association properties and substantially fewer Internal State & Emotion properties. Sensorimotor frequencies stay closer to the human range in aggregate, and Social frequencies are broadly comparable — though that category is sparse for humans and models alike.

Aggregate frequency alone does not show whether models ground the same concepts in the same way. A model could produce a human-like number of Sensorimotor properties overall while attaching them to different words. Measuring alignment directly — Pearson correlation between human and model category frequencies across the stimulus set — human–model alignment stays far below the estimated human–human ceiling for all 21 models.

The best model reaches only mean r ≈ 0.38 against an estimated human ceiling near 0.97. Scale does not explain the gap: larger models and stronger general-purpose systems do not show reliably higher human alignment, so the mismatch is not simply a capability bottleneck that shrinks with size.

Models are also much closer to one another than to humans: model–model correlations are systematically higher than model–human correlations. Current LLMs exhibit a shared mode of grounding abstract concepts that is systematically different from the human one.

Does the gap arise because models lack an understanding of the categories that structure human grounding, or because that understanding is not recruited when concepts are generated freely? To find out, we replicate a rating experiment (Troche et al. 2017) in which 751 English nouns are rated on 14 dimensions using a 1–7 Likert scale.

The dimensions cover sensory content (Color, Taste/Smell, Tactile, Visual Form, Auditory), motor content (Self-Motion), internal and evaluative content (Emotion, Polarity, Morality, Thought), social content (Social), and magnitude-related content (Space, Quantity, Time). Using the original prompt “I relate this word to [X]”, we evaluate the same 21 models over 10 shuffled runs each.

The results differ sharply from property generation. The strongest model reaches mean r ≈ 0.76 against an estimated human ceiling of about 0.87, and alignment improves substantially with model scale. The grounding gap is therefore not an inability to recognise sensation, emotion or sociality — those dimensions are available when the task makes them explicit.

If models can recover grounding dimensions when queried, is that information also reflected in their internal representations? We analyse Gemma-3 models with sparse autoencoders (SAEs), searching for features whose activations track the human grounding dimensions. Three prompt phrasings are extracted and averaged, so the surfaced features are stable across wording rather than artefacts of a single prompt.

Both models show fairly strong behavioural alignment with human ratings across most dimensions. Sensory dimensions are the main exception — model ratings stay far below the human ceiling. Nevertheless, the SAE analysis identifies features correlating with all evaluated dimensions, the behaviourally weaker sensory ones included. For most dimensions the best-matching features reach r = 0.6 to 0.8.

The main exception is the magnitude category — space, quantity and time — which shows weaker feature alignment, possibly reflecting the more abstract nature of those dimensions. These features track several sub-dimensions of grounding and align with human judgments to a degree; they should not be read as complete or fully human-like representations.

How the features are found

We take 425 nouns from four psycholinguistic norm datasets, each tied to one of the four grounding categories. For every noun we generate ten natural sentences that end with it, remove the noun, and read the SAE activations at the last remaining token: the point where the model is about to produce the word. The median over the ten sentences keeps the signal specific to the noun rather than to any one context. A feature becomes a candidate for a category when it fires for at least five nouns of that category.

Across layers of Gemma 3 4B, features tied to specific next words grow quickly after layer 17 and level off around layer 24. The category candidates behave differently: they rise slightly after layer 15, stay flat, and jump in the last layer. We read this as features tracking the theme of the context rather than the exact next word.

What the features look like

Most of these features have no single obvious meaning. Below are representative Gemma 3 4B features that you can inspect on Neuronpedia, each with an automatic description written by Claude Opus 4.6 from our activation data and its correlation with the human ratings. Neuronpedia hosts only some layers (9, 16, 22 and 29), so these are the best features available there, not always our strongest. Neuronpedia's own top logits can also differ from ours, because we look at the word predicted right after the active feature.

• Sensory (r = 0.575): layer 22, feature 193. Edible foods and taste or mouth-related content.
• Self-Motion (r = 0.426): layer 29, feature 132. Forceful or momentous events and qualities.
• Internal (r = 0.647): layer 29, feature 472. Emotional and motivational states, positive and negative.
• Social (r = 0.375): layer 22, feature 166. Family and kinship relations.
• Time (r = 0.373): layer 29, feature 733. Value-laden outcomes, advantages and setbacks.

Steering

We validate the features by steering: clamping a feature to a high activation during generation shifts the properties the model produces toward that feature's category, while the other categories fall or stay flat. In the paper we show this for the sensory and internal features above. You can explore the features and run the clamping yourself in our interactive demo on Hugging Face.

Per-word share of each property category for Gemma 3 4B over 291 abstract nouns and 10 runs, unsteered (grey) and with one feature clamped. Boxes span the middle 50% of words and the line is the median; numbers above each pair are the change in the mean, in percentage points.

All code and data needed to reproduce the paper are public. The supplementary repository ships a preloaded sample, so both leaderboards reproduce without an API key and without a GPU.

• Supplementary repository — property-generation experiments, the rating experiment, and the SAE feature analysis, each with its own walkthrough.
• Interactive demo — an explorer for the SAE features that track grounding dimensions on Gemma-3-4B-it, with live feature clamping.
• Camera-ready paper (PDF) and the arXiv preprint.

Human norms come from Harpaintner et al. (2018) and Troche et al. (2017) under CC-BY 4.0; the Kelly et al. (2024) norms are fetched on demand from the authors' OSF project and are not redistributed.