The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans

# The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans
Source: [https://arxiv.org/html/2605.08837](https://arxiv.org/html/2605.08837)
Odysseas S\. Chlapanis Department of Informatics Athens University of Economics and Business Archimedes, Athena Research Center Greece &Orfeas Menis Mastromichalakis Instituto de Telecomunicações Portugal &Christos H\. Papadimitriou Department of Computer Science Columbia University United States

###### Abstract

Abstract concepts — justice, theory, availability — have no single perceivable referent; in the human brain, their meaning emerges from a web of experiences, affect, and social context\. Do large language models \(LLMs\) ground abstract concepts in a similar way? We study this by replicating property\-generation experiments from cognitive science on 21 frontier and open\-weight LLMs\. Across models and experiments, we find a consistent pattern: when compared to humans, models rely too heavily on word associations, and underproduce properties tied to emotion and internal states\. This yields a large and consistent grounding gap: no model exceeds a Pearson correlationr=0\.37r=0\.37with human responses, compared to a human\-to\-human ceiling abover=0\.9r=0\.9\. To better interpret this gap, we also replicate a rating experiment on grounding categories and find that here LLMs align more closely with human judgment, and alignment improves as models get larger\. We then use sparse autoencoders \(SAEs\) to inspect whether this information is also reflected in the models’ internal features, and we do identify features connected to grounding dimensions such as “sensorimotor” and “social”\. These findings suggest that current LLMs can recover grounding dimensions when explicitly queried, but do not recruit them in a human\-like way when words are generated freely\.111All data and code are publicly available at[https://github\.com/odychlapanis/grounding\-gap/\.](https://github.com/odychlapanis/grounding-gap/)

## 1Introduction

Concrete concepts such ascatandtableare anchored in shared experience through identifiable referents that can be perceived through the senses\(Borghiet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib1)\)\. In contrast, abstract concepts such asart,adventure, orjustice, lack such referents\. When humans process abstract words, they reconstruct their meaning from grounded experiences\. Consider how a person makes sense ofart: a colorfulpainting, a feeling ofawe, a favoriteconcert, as well as more abstract ideas such ascreativity\. Grounded\-cognition theories argue that such sensorimotor, emotional, and social anchors are not incidental but constitutive of meaning\(Lakoff and Johnson,[2003](https://arxiv.org/html/2605.08837#bib.bib7); Barsalou,[1999](https://arxiv.org/html/2605.08837#bib.bib5),[2026](https://arxiv.org/html/2605.08837#bib.bib57)\)\. Whether LLMs represent abstract concepts in a similarly grounded way, or rely primarily on patterns of linguistic association\(Bender and Koller,[2020](https://arxiv.org/html/2605.08837#bib.bib65); Benderet al\.,[2021](https://arxiv.org/html/2605.08837#bib.bib66)\), remains an open question\.

To measure how closely LLMs align with humans in the grounding of abstract concepts, we replicate on LLMs two property\-generation experiments from cognitive science\(Harpaintneret al\.,[2018](https://arxiv.org/html/2605.08837#bib.bib8); Kellyet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib9)\)\. In these experiments, participants — in these references humans, in this paper models — are given a concept and asked to list the properties, situations, or associations that come to mind\. Although individual responses vary, 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 similar and consistent way\. We then compare the distributions produced by models with those produced by humans\. Across 21 frontier and open models, we find a large and consistent grounding gap\. In both experiments, no model reaches a human–model correlation abover=0\.37r=0\.37, while the human–human ceilings exceedr=0\.9r=0\.9\. Models are also much closer to one another than to humans: model–model correlations are systematically higher than model–human correlations\. Taken together, these results show that current LLMs exhibit a shared mode of grounding abstract concepts that is systematically different from the human one\.

But does the gap arise because models lack an understanding of the categories that structure human grounding, or because this understanding is not recruited in a human\-like way when concepts are generated freely? To answer, we turn to a rating experiment from cognitive science, in which words are evaluated directly along grounding\-related dimensions\(Trocheet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib2)\)\. There, recent LLMs align much more closely with human judgments, and this alignment improves with model scale, leaving a gap of only about \.1 in Pearson correlation with humans\. This contrast suggests that the relevant dimensions are at least partly available to the models when they are explicitly queried, but they are not spontaneously recruited in a human\-like way during property generation\.

This raises a mechanistic question: Can we find in LLMs internal representations that align with these grounding\-related dimensions? To investigate this, we analyze models with sparse autoencoders \(SAEs\), a technique that helps identify interpretable internal features\(Brickenet al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib38); Hubenet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib37); Templetonet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib39)\)\. We search for features whose activations track grounding\-related dimensions, and we do find evidence of such structure: some internal features correlate with human grounding dimensions or sub\-dimensions\. We then validate these features by showing that steering the model along them increases generation of properties from the corresponding target categories\. Taken together, our findings suggest that current LLMs understand the basic grounding categories nearly–but not quite–as well as humans do, and yet they do not recruit them in a human\-like way when generating freely\.

#### Contributions\.

- •We replicate two property\-generation experiments from cognitive science on 21 frontier and open LLMs, and show a large and consistent gap between humans and models in the grounding of abstract concepts\.
- •We replicate a rating experiment on grounding\-related dimensions and show that LLMs align much more closely – but not fully – with human judgments there, indicating that the gap is not simply due to failure to recognize the relevant dimensions\.
- •We use sparse autoencoders to probe Gemma models for internal features related to grounding dimensions, and we do find evidence of such structure\.

## 2Related work

#### Grounded cognition and grounding norms\.

Traditional theories of*grounded cognition*in language posit that abstract categories are not represented as arbitrary symbols but are grounded in sensorimotor and affective experience\(Barsalou,[1999](https://arxiv.org/html/2605.08837#bib.bib5); Barsalou and Wiemer\-Hastings,[2005](https://arxiv.org/html/2605.08837#bib.bib6)\)\. Neurocognitive evidence supports that action words engage motor and premotor cortex\(Hauket al\.,[2004](https://arxiv.org/html/2605.08837#bib.bib68); Pulvermüller,[2005](https://arxiv.org/html/2605.08837#bib.bib69)\), and emotional and social dimensions of abstract concepts have been linked to interoceptive processing\(Critchleyet al\.,[2004](https://arxiv.org/html/2605.08837#bib.bib34); Mancano and Papagno,[2026](https://arxiv.org/html/2605.08837#bib.bib72)\)\. Behavioral norms operationalize these dimensions at scale: datasets in which human participants rate words on Likert scales for dimensions such as sensorimotor strength, arousal, valence, and socialness\.Lynottet al\.\([2020](https://arxiv.org/html/2605.08837#bib.bib11)\)provide modality\-specific strength ratings for 40,000 English words, whileScottet al\.\([2019](https://arxiv.org/html/2605.08837#bib.bib12)\)andDiveicaet al\.\([2023](https://arxiv.org/html/2605.08837#bib.bib13)\)cover arousal, valence, and socialness dimensions\. These corpora enable grounding research to move beyond binary concrete/abstract distinctions\(Brysbaertet al\.,[2014](https://arxiv.org/html/2605.08837#bib.bib10)\)\. The property\-listing paradigm\(McRaeet al\.,[2005](https://arxiv.org/html/2605.08837#bib.bib3); Barsalou and Wiemer\-Hastings,[2005](https://arxiv.org/html/2605.08837#bib.bib6)\)provides a richer signal: rather than scalar ratings, participants freely generate properties of concepts, yielding a distributional profile over experiential dimensions\.Harpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)applied this paradigm specifically to abstract nouns, showing that they elicit rich sensorimotor, social, and affective features\.Kellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)extended the paradigm to test whether emotion concepts form a distinct subcategory, finding that they rely more strongly on interoceptive features than other abstract concepts\. These two studies form the empirical basis of our behavioral experiments\.

#### LLM cognitive experiments\.

Early work showed that language models often fail psycholinguistic diagnostics and do not reliably ground semantic properties\(Ettinger,[2020](https://arxiv.org/html/2605.08837#bib.bib35)\)\.Pezzelleet al\.\([2021](https://arxiv.org/html/2605.08837#bib.bib19)\)compared pre\-trained transformer representations against human concept norms and found that visual co\-training improves alignment for concrete word pairs, but offers little benefit for abstract ones\. Recent work extends this question to frontier and open LLMs by prompting them for human\-style ratings and comparing them to established norms\.Xuet al\.\([2025](https://arxiv.org/html/2605.08837#bib.bib4)\)report strong alignment with Glasgow and Lancaster norms on non\-sensorimotor dimensions, but weaker alignment on sensory and motor domains\.Wanget al\.\([2025b](https://arxiv.org/html/2605.08837#bib.bib20)\)similarly find persistent gaps for abstract concepts, embodied modalities, and emotion or social dimensions across nine LLMs, with alignment improving with scale\. Related studies also show that LLMs can approximate affective and sensorimotor ratings\(Trott,[2024](https://arxiv.org/html/2605.08837#bib.bib21)\)and recover aspects of perceptual color structure from text alone\(Abdouet al\.,[2021](https://arxiv.org/html/2605.08837#bib.bib22)\)\. Closest to our setup,Sureshet al\.\([2023](https://arxiv.org/html/2605.08837#bib.bib18)\)use property listing rather than ratings, but focus on concrete concepts and feature\-overlap similarity against existing norms\. In contrast, we target abstract concepts and classify generated properties into experiential categories\.

#### Mechanistic interpretability of semantic concepts\.

A prominent line of mechanistic interpretability work uses linear probes to localize non\-linguistic latent structure in LLM residual streams, recovering concept axes for board game states, spatiotemporal coordinates, and sentiment\(Liet al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib28); Nandaet al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib29); Gurnee and Tegmark,[2024](https://arxiv.org/html/2605.08837#bib.bib25); Tiggeset al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib26)\)\. Some works extend this paradigm by using these latent vectors to steer models towards the detected concepts\(Turneret al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib45); Panicksseryet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib44)\)\. Recent studies have identified circuit\-level emotion components to control emotional expression\(Wanget al\.,[2025a](https://arxiv.org/html/2605.08837#bib.bib32)\)and recovered emotion directions in models like Claude Sonnet 4\.5\(Sofroniewet al\.,[2026](https://arxiv.org/html/2605.08837#bib.bib71)\)\. In this work, we analyze features learned by sparse autoencoders \(SAEs\) to probe whether grounding\-related information is reflected in model representations\. SAEs decompose polysemantic LLM representations into monosemantic, interpretable features\(Brickenet al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib38); Hubenet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib37); Templetonet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib39)\)\. Recently, SAE features have been applied for analyzing emotion\(Wuet al\.,[2025](https://arxiv.org/html/2605.08837#bib.bib31)\)\. In our work, we leverage the Gemma Scope 2 SAE models\(McDougallet al\.,[2025](https://arxiv.org/html/2605.08837#bib.bib30)\), the successor to Gemma Scope\(Lieberumet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib42)\), and we probe sensorimotor, internal\-state, and social concepts\.

## 3The grounding gap

We study LLM grounding of abstract concepts through*property generation*, a behavioral paradigm from cognitive science in which participants are given a concept and asked to list the features, situations, or associations that come to mind for it\. Unlike scalar rating tasks\(Trocheet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib2)\), which require participants to map conceptual content onto a fixed set of predefined dimensions, property generation captures the content that people spontaneously recruit when representing a concept\. For this reason, it has been used as a richer and more faithful way to study grounding, especially for abstract concepts\(Barsalou and Wiemer\-Hastings,[2005](https://arxiv.org/html/2605.08837#bib.bib6); McRaeet al\.,[2005](https://arxiv.org/html/2605.08837#bib.bib3)\)\.

To measure how closely LLMs align with humans on conceptual grounding, we replicate two human property\-generation studies on abstract concepts:Harpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)andKellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)\. We selected these studies because both provide publicly available data and explicit coding taxonomies, enabling controlled human–model comparison under complementary experimental designs\.Harpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)\(Experiment 1\) organizes properties of abstract concepts intosensorimotor,internal\-state,social, andverbal\-associationcontent, reflecting dimensions of abstract\-concept representation that are well established in the broader literature\(Trocheet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib2); Villaniet al\.,[2019](https://arxiv.org/html/2605.08837#bib.bib49)\)\.Kellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)\(Experiment 2\) provides a more specialized setting, with a different stimulus design and coding taxonomy with a particular focus on emotion concepts\. Using both experiments allows us to test whether human–model grounding gaps persist across distinct property\-generation paradigms rather than arising from a single dataset or coding scheme\.

### 3\.1Experimental setup

![Refer to caption](https://arxiv.org/html/2605.08837v1/x1.png)Figure 1:Per\-word property\-category frequency distributions for three frontier LLMs vs\. the human baseline on the 293\-abstract\-noun benchmark\(Harpaintneret al\.,[2018](https://arxiv.org/html/2605.08837#bib.bib8)\)\. Each box spans the 25th–75th percentile of word\-level frequencies; median annotated\.We apply the same procedure in both experiments, following their original designs\. For each stimulus word, we prompt the model to generate the properties, situations, or associations that come to mind \(four properties for Experiment 1 and five for Experiment 2\), using the original human instructions \(prompts in Appendix[G](https://arxiv.org/html/2605.08837#A7)\)\. We then map the generated properties to the taxonomy of the corresponding human study and compare the resulting per\-word category generations against the human ones\. All scores are averaged over 10 runs per model; see Appendix[E](https://arxiv.org/html/2605.08837#A5)for convergence analysis\.

#### Stimulus sets\.

The stimulus set consists of the target words presented to participants—or, in our case, to the models—for property generation\. For the first experiment\(Harpaintneret al\.,[2018](https://arxiv.org/html/2605.08837#bib.bib8)\), we derive a 293\-word English stimulus set from the original 296 German abstract concepts\. Two items were excluded as translation duplicates \(the original paper provided English translations for all words\), and one additional item was excluded because it is used as a one\-shot example in the original prompt\. In the second experiment\(Kellyet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib9)\), the stimulus set contains 357 words split into three subsets: 118 abstract emotion words, 118 abstract non\-emotion words, and 119 concrete words\. We use the 236 abstract words for our analysis; see Appendix[C](https://arxiv.org/html/2605.08837#A3)for findings for concrete words\.

#### Property coding\.

The generated properties are open\-ended, so they must be assigned to the grounding categories defined by each study\. Following the original coding schemes, generated properties forHarpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)are assigned to:Sensorimotor,Internal State & Emotion,Social, andVerbal Association, while forKellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)toTaxonomic,Entity,Situation, andIntrospective\. We use LLMs\-as\-coders, after validating candidate models on human\-annotated word\-to\-category pairs\. The selected coders achieve agreement with the ground truth that is comparable to, or higher than, human coders, while remaining practical to use at scale\. We useGemini\-2\.5\-Flash\-Liteas the primary coder for Experiment 1 andGemini\-2\.5\-Flashfor Experiment 2\. Detailed coder comparisons, annotation protocols, and reliability analyses are reported in Appendix[A](https://arxiv.org/html/2605.08837#A1)\.

#### Alignment metric and human ceiling\.

After coding, each stimulus word is represented by the relative frequencies of its responses across the four target categories\. We compare human and model frequency profiles using per\-category Pearsonrr, and reportMeanrr, the average across the four categories, as our main alignment metric\. The corresponding human ceiling is the Pearsonrrthat two independent splits of raters would obtain on the same words: for Experiment 2 we compute it directly fromKellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)’s participant responses, and for Experiment 1 we estimate it analytically fromHarpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)’s per\-word aggregates using standard inter\-rater reliability formulas; the two methods agree to within0\.010\.01on Experiment 2, validating our estimator \(Appendix[F](https://arxiv.org/html/2605.08837#A6)\)\.

#### Models and settings\.

We evaluated a diverse set of 21 frontier and open\-weight LLMs spanning several families; we refer to Appendix[B\.1](https://arxiv.org/html/2605.08837#A2.SS1)for the full list of models\. We used the default temperature for all generations and to ensure coding reliability LLM\-coders used a temperature ofT=0\.0T=0\.0\.

Table 1:Per\-category Pearson correlations on Experiment 1 for three representative frontier models\.

### 3\.2Results

#### Experiment 1\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x2.png)Figure 2:Mean r for all LLMs vs\. the human ceiling on Experiment 1\(Harpaintneret al\.,[2018](https://arxiv.org/html/2605.08837#bib.bib8)\)\. Parameter count for closed models is based on estimations fromLi \([2026](https://arxiv.org/html/2605.08837#bib.bib74)\)\.Figure[1](https://arxiv.org/html/2605.08837#S3.F1)illustrates the per\-word and per\-category distributions for human responses and for three representative frontier models: Claude Opus 4\.6, Gemini 3\.1 Pro Preview, and GPT\-5\.4\. Relative to humans, all three models generate substantially moreVerbal Associationproperties and substantially fewerInternal State & Emotionproperties\. The same overall pattern holds across all models \(see Appendix[B](https://arxiv.org/html/2605.08837#A2)for the full set of models\)\. The similarity between the frequency of the different models also indicates that this is not a coincidental property of a single model family, but a consistent behavior across current architectures\. By contrast,Sensorimotorfrequencies remain closer to the human range in the aggregate, andSocialfrequencies are also broadly comparable, although this category is relatively sparse for both humans and models and is therefore less informative at the frequency level alone\.

The frequency distributions indicate a systematic mismatch between human and model grounding, but aggregate frequency alone does not show whether models ground the*same*concepts in the*same*way as humans\. For example, a model may produce a human\-like number ofSensorimotororSocialproperties overall, while assigning those properties to different words\. We therefore measure alignment directly by computing Pearson correlations between human and model category\-frequency generations across the stimulus set, as described in Section[3\.1](https://arxiv.org/html/2605.08837#S3.SS1)\.

As shown in Figure[2](https://arxiv.org/html/2605.08837#S3.F2), human\-model alignment remains far below the estimated human–human ceiling across all 21 models\. The best model reaches only a meanr≈0\.37r\\approx 0\.37, compared with an estimated human ceiling of about0\.970\.97, with the same pattern across individual categories as shown in Table[1](https://arxiv.org/html/2605.08837#S3.T1)\. Thus, even when aggregate frequencies appear relatively human\-like, models often associate grounding dimensions with different concepts than humans do\. This gap is not explained by scale: larger models and stronger general\-purpose systems do not show reliably higher human alignment, suggesting that the mismatch is not simply a capability bottleneck that shrinks with model size\.

Inter\-model correlations reveal a complementary pattern\. As shown in Figure[3\(a\)](https://arxiv.org/html/2605.08837#S3.F3.sf1), models correlate substantially more with one another than with humans, despite spanning different providers, data mixtures, and architectural choices\. This suggests that current LLMs share a rather common mode of abstract\-concept grounding, but one that is systematically distinct from the human pattern\. The grounding gap is therefore not idiosyncratic to a single model family, but shared across current autoregressive LLMs\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x3.png)\(a\)Experiment 1 \(Harpaintneret al\.\([2018](https://arxiv.org/html/2605.08837#bib.bib8)\)\)
![Refer to caption](https://arxiv.org/html/2605.08837v1/x4.png)\(b\)Experiment 2 \(Kellyet al\.\([2024](https://arxiv.org/html/2605.08837#bib.bib9)\)\)

Figure 3:Pearson\-rrcorrelation heatmaps for three representative frontier models and humans\.
#### Experiment 2\.

Experiment 2 confirms the findings of Experiment 1 under a different experimental setup\. Human\-model alignment remains low in the combined abstract condition, which pools theabstract emotionandabstract non\-emotionitems: even the strongest models reach only a meanr≈0\.33r\\approx 0\.33, far below the human\-to\-human correlation ofr≈0\.91r\\approx 0\.91\. Full results for all models are reported in Appendix[B](https://arxiv.org/html/2605.08837#A2)\.

Inter\-model correlations also follow the same pattern as in Experiment 1\. As shown in Figure[3\(b\)](https://arxiv.org/html/2605.08837#S3.F3.sf2), models correlate substantially more with one another than with humans\. Together, the two experiments reveal a robust behavioral grounding gap: current LLMs share a common model\-like pattern of abstract\-concept representation, but one that remains systematically distinct from humans\.

### 3\.3LLM\-coder error margins

The Pearson correlations and category frequencies reported above are based on labels produced by Gemini2\.52\.5Flash\-Lite\. To estimate whether this affects our conclusions, we manually re\-coded the outputs of three frontier models \(Claude Opus4\.64\.6, Gemini3\.13\.1Pro, and GPT\-5\.45\.4\) and compared the resulting metrics with those obtained from Flash\-Lite labels on the same property pairs\. The differences are minor: replacing Flash\-Lite with human coders changes Meanrrby onlyΔ=\+0\.014±0\.019\\Delta=\+0\.014\\pm 0\.019on average, with no consistent per\-model shift \(per\-model95%95\\%confidence intervals contain zero\), and changes per\-category mean frequencies by roughly±2%\\pm 2\\%\. These margins are far smaller than the human–model gap reported in Section[3\.2](https://arxiv.org/html/2605.08837#S3.SS2.SSS0.Px1), indicating that our main results are not driven by the use of LLMs as coders\. Detailed analyses are reported in Appendix[A](https://arxiv.org/html/2605.08837#A1)\.

## 4Rating analysis of grounding dimensions

To better understand the grounding gap identified through the previous experiments, we study whether this gap arises because current LLMs do not properly recover the grounding dimensions that structure human concept representations, or because these dimensions are not recruited in a human\-like way when concepts are generated freely\. To distinguish between these possibilities, we replicate a rating experiment from cognitive science\(Trocheet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib2)\)\. Unlike property generation, rating experiments do not test the spontaneous associations for a concept; instead, they test whether the model can score a word on semantic dimensions once these dimensions are explicitly specified\.

#### Setup\.

We follow the setup ofTrocheet al\.\([2017](https://arxiv.org/html/2605.08837#bib.bib2)\), in which human participants rated 751 English nouns 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 evaluated the same 21 models from previous experiments over 10 shuffled runs each\. Performance is reported as Mean r, the Pearson correlation between model and human ratings averaged across all 14 dimensions\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x5.png)Figure 4:Mean r for all LLMs vs\. the human ceiling on the rating experiment\(Trocheet al\.,[2017](https://arxiv.org/html/2605.08837#bib.bib2)\)\. Parameter count for closed models is based on estimations fromLi \([2026](https://arxiv.org/html/2605.08837#bib.bib74)\)\.
#### Results\.

The results differ sharply from those of property generation \(compare Figures[2](https://arxiv.org/html/2605.08837#S3.F2)and[4](https://arxiv.org/html/2605.08837#S4.F4)\)\. On this task, recent frontier LLMs align much more closely with human judgments, with the strongest model reaching Meanr≈0\.76r\\approx 0\.76against an estimated human ceiling of about0\.890\.89\. Unlike property generation, alignment here improves substantially with model scale\. This contrast suggests that the grounding gap is not due to an inability to recognize dimensions such as sensation, emotion, or sociality\. Rather, current LLMs can recover these dimensions when they are made explicit in the task, but do not recruit them in a human\-like way when asked to generate the content of abstract concepts freely\. Extended experimental details and results are provided in Appendix[D](https://arxiv.org/html/2605.08837#A4)\.

## 5Mechanistic analysis

Having established that LLMs can predict grounding dimension ratings, we investigate whether they internally encode features tracking these dimensions by analyzing the 4B and 12B Gemma 3 instruction\-tuned models using Gemma Scope 2 sparse autoencoders \(SAEs\)\(McDougallet al\.,[2025](https://arxiv.org/html/2605.08837#bib.bib30)\)\. We isolate latent features correlated with grounding dimensions, focusing on those that align strongly with human experiential categories\. Remarkably, some of these internal features surpass the model’s own behavioral baseline in human alignment\. While steering these representations predictably increases the generation of target\-category properties, we view this merely as a methodological sanity check; artificially amplifying feature activations does not close the grounding gap, as it overrides rather than reflects the model’s natural internal structure\.

### 5\.1Methodology

#### Feature identification\.

We ask whether the model contains SAE features that correspond to grounding categories or sub\-groups beyond individual words\. First, we collect a pool of 426 labeled nouns from four psycholinguistic norm databases\(Lynottet al\.,[2020](https://arxiv.org/html/2605.08837#bib.bib11); Diveicaet al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib13); Scottet al\.,[2019](https://arxiv.org/html/2605.08837#bib.bib12); Brysbaertet al\.,[2014](https://arxiv.org/html/2605.08837#bib.bib10)\), each associated with one of the 4 grounding dimensions of interest \(sensorimotor, internal\-state, social, and abstract content\)\. For each nounww, we generate 10 English sentences ending inwwusing Gemini\-3\.1\-pro\-preview, and then removeww, creating a form of cloze \(fill\-the\-blank\) questions to study model activations when generatingww\. We record the SAE activation at the final token of these cloze questions, i\.e\. the position where the model is about to generateww, and take the per\-feature median across the ten sentences as the noun’s activation signature\. A feature enters categorycc’s candidate pool if it has non\-zero median activation on at least five distinct nouns fromcc\. The approach is similar to the one used by\(Sofroniewet al\.,[2026](https://arxiv.org/html/2605.08837#bib.bib71)\)for emotion vectors\. Construction details are reported in Appendix[H\.1](https://arxiv.org/html/2605.08837#A8.SS1)\. This approach gives us a set of candidate features for all 4 grounding dimensions\.

#### Tracking grounding dimension norms with SAE features\.

The previous stage is designed to select features that their activations are correlated with generation of nouns related to their category, but their activations are not necessarily aligned with humans\. Our goal is to find features that not only respond to a grounding dimension, but do so in a way that is aligned with human norms in a grounding dimension\. To achieve this we evaluate these candidate features on the same dataset we used in our rating experiment in Section[4](https://arxiv.org/html/2605.08837#S4)\. For the*feature*measurement, the original rating prompts are unsuitable: they target a specific dimension and elicit a numeric answer, whereas our candidate features were identified in the regime of free noun generation\. We instead use two simple and generic property\-generation prompts and the original complex property\-generation prompt from Section[3](https://arxiv.org/html/2605.08837#S3)and average the results, record each candidate feature’s activation at the final prompt position, and correlate it with the human ratings on each dimension\. This measures how well the feature itself tracks the human grounding signal while simulating the property\-generation task\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x6.png)Figure 5:Percentage of mean r for all LLMs and SAE features vs\. the estimated human ceiling mean r on the rating experiment dimensions\.

### 5\.2Results

Overall, both models show a rather strong behavioral alignment with human ratings across most grounding dimensions \(Figure[5](https://arxiv.org/html/2605.08837#S5.F5)\)\. Sensory dimensions are the main exception: model ratings remain far below the human ceiling, which is consistent with the findings ofXuet al\.\([2025](https://arxiv.org/html/2605.08837#bib.bib4)\)\. Nevertheless, the SAE analysis identifies features that correlate with all evaluated dimensions, including the behaviorally weaker sensory ones\. For most dimensions, the best\-matching SAE features reach Pearson correlations in the range ofr=0\.6r=0\.6to0\.80\.8, suggesting that these grounding\-related dimensions are at least partially reflected in the models’ internal representations\. The main exception is the magnitude category, which includes space, quantity, and time and shows weaker feature alignment, possibly reflecting the more abstract nature of these dimensions\. Overall, these results suggest that the models contain internal features that track several sub\-dimensions of grounding, with these features being aligned with human judgments to some extent, and should not be interpreted as complete or fully human\-like representations\.

### 5\.3Feature steering

To interpret and validate these features, we conduct a steering experiment on Gemma 3 4B using the highest\-correlation SAE feature for each dimension: sensory, motor, internal, social, and magnitude\. We replicate Experiment 1 \(Section[3\.2](https://arxiv.org/html/2605.08837#S3.SS2.SSS0.Px1)\) under feature steering, using feature clamping\(Templetonet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib39)\)to a fixed high activation value\. Steering increases target\-category property generation in all cases: sensory \(\+5\.3%*Sensorimotor*\), motor \(\+4\.7%*Sensorimotor*\), internal \(\+9\.4%*Internal*\), social \(\+14\.7%*Social*\), and magnitude \(\+1\.7%*Verbal*\)\. These effects support the feature–category associations identified in the correlation analysis\. We further inspect the highest\-activating examples for each feature using Neuronpedia,222Neuronpedia is an open\-source platform for SAE interpretability and visualization:[https://www\.neuronpedia\.org](https://www.neuronpedia.org/)\.and find that the features generally activate on words related to their grounding dimensions\. However, they are not perfectly category\-specific: some also activate on words outside the narrow target dimension, suggesting that they capture broader semantic patterns rather than clean, isolated grounding categories or sub\-dimensions\. For details, see Appendix[H\.2](https://arxiv.org/html/2605.08837#A8.SS2)\.

## 6Discussion

Grounding is important for humans because it connects language to lived experience and the physical world\. As LLMs take roles originally meant for humans and move closer to agentic autonomy, their apparent differences from humans in grounding acquire some significance\. The grounding gap we identified is unlikely to go away through scaling and text\-based training alone, since \(a\) our experiments show that scaling\(Hoffmannet al\.,[2022](https://arxiv.org/html/2605.08837#bib.bib75); Weiet al\.,[2022](https://arxiv.org/html/2605.08837#bib.bib41)\)does not appear to help, with state\-of\-the\-art models remaining remarkably far from human ceilings, and \(b\) the excessive reliance on verbal associations we observe is a natural result of training that is focused on prediction from linguistic context, a core characteristic of autoregressive LLMs\. Genuine progress on this front may require fundamental shifts, such as novel training techniques or alternative architectures\(Funget al\.,[2025](https://arxiv.org/html/2605.08837#bib.bib82); Brooks,[1991](https://arxiv.org/html/2605.08837#bib.bib83)\)\.

Our interpretability analysis reveals that LLMs possess latent grounding structures that sometimes exhibit stronger human alignment than the model’s overall behavioral output\. However, these features do not exclusively encode sensory, internal, or social experiences \(see Appendix[H\.2](https://arxiv.org/html/2605.08837#A8.SS2)for details\)\. According to one mainstream theory of grounded cognition\(Barsalou,[1999](https://arxiv.org/html/2605.08837#bib.bib5)\), true conceptual understanding requires the capacity to internally simulate physical experiences\. Whether LLMs possess, or can acquire through training, internal circuits capable of such simulation remains an open challenge for future research\.

#### Limitations\.

Our behavioral analysis is limited by the generality of the property\-generation experiments we replicate, although both are well established in the literature\. To our knowledge, these are the only extant property\-generation studies on abstract\-concept grounding with both openly available stimulus sets and coding taxonomies suitable for direct replication and comparison\. The fact that we observe highly consistent results across these two independently designed experiments, with different stimuli, taxonomies, and participant samples, suggests that the observed gap is not specific to a single experimental setup\.

A second limitation is the use of LLMs as coders, which could introduce model\-specific biases into the estimated category frequencies and correlations\. We mitigate this by validating candidate coders against human\-annotated word\-to\-category pairs and selecting models with human\-level or stronger agreement\. We also show in Section[3\.3](https://arxiv.org/html/2605.08837#S3.SS3)that replacing the LLM coder with human coders changes the main metrics only marginally, with differences far smaller than the human–model gaps reported above\. This makes it highly unlikely that our conclusions are driven by the coding procedure\.

Finally, our SAE analysis is limited to Gemma models due to the limited availability of high\-quality SAEs for recent LLMs and the cost of scaling the analysis\. More generally, SAE features are interpretable approximations of model representations, not complete decompositions\. The absence of a feature in an SAE therefore does not guarantee that the corresponding information is absent from the underlying model\. We therefore treat the SAE results as exploratory evidence that complements the behavioral findings, rather than as the basis for strong mechanistic claims\.

## 7Conclusion

We investigate how well frontier LLMs are cognitively aligned with humans in their grounding of abstract concepts\. We replicate two property\-generation experiments on 21 LLMs, and find strong evidence thatLLMs are not grounded in the same way humans are— but of course they were not trained to be\. In particular, LLMs exhibit systematic under\-production of internal\-state properties and over\-production of abstract ones, and overall very poor Pearson correlation with humans, while they correlate well with each other\. These failures are uniform across LLM families and do not improve with scale\. In a replicated rating experiment, LLMs exhibit a reasonable understanding of concept categories, and in fact SAE features selective for such categories exist in Gemma\-3\-4B and Gemma\-3\-12B, in evidence that grounding structure is encoded internally in LLMs even though it may not behaviorally expressed\. Going forward, our results suggest that grounding alignment with humans may be an interesting and novel quantity that is invisible to current testing benchmarks, and may be worth watching in the future\. And finally, further scaling or better text\-based training will probably not close the gap, so we may need to rethink the way we design and train LLMs\.

## References

- Can language models encode perceptual structure without grounding? a case study in color\.InProceedings of the 25th Conference on Computational Natural Language Learning \(CoNLL\),External Links:[Link](https://aclanthology.org/2021.conll-1.9/)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- L\. W\. Barsalou and K\. Wiemer\-Hastings \(2005\)Situating abstract concepts\.InGrounding Cognition: The Role of Perception and Action in Memory, Language, and Thought,D\. Pecher and R\. A\. Zwaan \(Eds\.\),pp\. 129–163\.External Links:[Document](https://dx.doi.org/10.1017/CBO9780511499968.007)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2605.08837#S3.p1.1)\.
- L\. W\. Barsalou \(1999\)Perceptual symbol systems\.Behavioral and Brain Sciences22\(4\),pp\. 577–660\.External Links:[Document](https://dx.doi.org/10.1017/S0140525X99002149)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2605.08837#S6.p2.1)\.
- L\. W\. Barsalou \(2026\)Grounded cognition\.InOpen Encyclopedia of Cognitive Science,M\. C\. Frank and A\. Majid \(Eds\.\),External Links:[Link](https://oecs.mit.edu/pub/9iq4376o)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1)\.
- E\. M\. Bender, T\. Gebru, A\. McMillan\-Major, and S\. Shmitchell \(2021\)On the dangers of stochastic parrots: can language models be too big?\.InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency,FAccT ’21,New York, NY, USA,pp\. 610–623\.External Links:ISBN 9781450383097,[Link](https://doi.org/10.1145/3442188.3445922),[Document](https://dx.doi.org/10.1145/3442188.3445922)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1)\.
- E\. M\. Bender and A\. Koller \(2020\)Climbing towards NLU: On meaning, form, and understanding in the age of data\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,D\. Jurafsky, J\. Chai, N\. Schluter, and J\. Tetreault \(Eds\.\),Online,pp\. 5185–5198\.External Links:[Link](https://aclanthology.org/2020.acl-main.463/),[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.463)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1)\.
- A\. M\. Borghi, F\. Binkofski, C\. Castelfranchi, F\. Cimatti, C\. Scorolli, and L\. Tummolini \(2017\)The challenge of abstract concepts\.Psychological Bulletin143\(3\),pp\. 263–292\.External Links:[Document](https://dx.doi.org/10.1037/bul0000089)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1)\.
- T\. Bricken, A\. Templeton, J\. Batson, B\. Chen, A\. Jermyn, T\. Conerly, N\. Turner, C\. Anil, C\. Denison, A\. Askell, R\. Lasenby, Y\. Wu, S\. Kravec, N\. Schiefer, T\. Maxwell, N\. Joseph, Z\. Hatfield\-Dodds, A\. Tamkin, K\. Nguyen, B\. McLean, J\. E\. Burke, T\. Hume, S\. Carter, T\. Henighan, and C\. Olah \(2023\)Towards monosemanticity: decomposing language models with dictionary learning\.Transformer Circuits Thread\.Note:https://transformer\-circuits\.pub/2023/monosemantic\-features/index\.htmlCited by:[§1](https://arxiv.org/html/2605.08837#S1.p4.1),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- R\. A\. Brooks \(1991\)Intelligence without representation\.Artificial Intelligence47\(1\),pp\. 139–159\.External Links:ISSN 0004\-3702,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/0004-3702%2891%2990053-M),[Link](https://www.sciencedirect.com/science/article/pii/000437029190053M)Cited by:[§6](https://arxiv.org/html/2605.08837#S6.p1.1)\.
- M\. Brysbaert, A\. B\. Warriner, and V\. Kuperman \(2014\)Concreteness ratings for 40 thousand generally known English word lemmas\.Behavior Research Methods46\(3\),pp\. 904–911\.External Links:[Document](https://dx.doi.org/10.3758/s13428-013-0403-5)Cited by:[Table 9](https://arxiv.org/html/2605.08837#A8.T9.8.8.4),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1.p1.7)\.
- H\. D\. Critchley, S\. Wiens, P\. Rotshtein, A\. Öhman, and R\. J\. Dolan \(2004\)Neural systems supporting interoceptive awareness\.Nature Neuroscience7\(2\),pp\. 189–195\.External Links:[Document](https://dx.doi.org/10.1038/nn1176)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1)\.
- V\. Diveica, P\. M\. Pexman, and R\. J\. Binney \(2023\)Quantifying social semantics: an inclusive definition of socialness and ratings for 8,388 English words\.Behavior Research Methods55,pp\. 461–473\.External Links:[Document](https://dx.doi.org/10.3758/s13428-022-01810-x)Cited by:[Table 9](https://arxiv.org/html/2605.08837#A8.T9.3.3.4),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1.p1.7)\.
- A\. Ettinger \(2020\)What BERT is not: lessons from a new suite of psycholinguistic diagnostics for language models\.Transactions of the Association for Computational Linguistics8,pp\. 34–48\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00298)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- P\. Fung, Y\. Bachrach, A\. Celikyilmaz, K\. Chaudhuri, D\. Chen, W\. Chung, E\. Dupoux, H\. Gong, H\. Jégou, A\. Lazaric,et al\.\(2025\)Embodied ai agents: modeling the world\.arXiv preprint arXiv:2506\.22355\.Cited by:[§6](https://arxiv.org/html/2605.08837#S6.p1.1)\.
- W\. Gurnee and M\. Tegmark \(2024\)Language models represent space and time\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 2483–2503\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/0a6059857ae5c82ea9726ee9282a7145-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- M\. Harpaintner, N\. M\. Trumpp, and M\. Kiefer \(2018\)The semantic content of abstract concepts: A property listing study of 296 abstract words\.Frontiers in Psychology9,pp\. 1748\.External Links:[Document](https://dx.doi.org/10.3389/fpsyg.2018.01748)Cited by:[§A\.1](https://arxiv.org/html/2605.08837#A1.SS1.SSS0.Px2.p1.5),[Table 2](https://arxiv.org/html/2605.08837#A1.T2),[Appendix B](https://arxiv.org/html/2605.08837#A2.p1.1),[§F\.2](https://arxiv.org/html/2605.08837#A6.SS2.SSS0.Px1.p1.1),[Appendix F](https://arxiv.org/html/2605.08837#A6.p1.9),[Figure 10](https://arxiv.org/html/2605.08837#A7.F10),[Figure 10](https://arxiv.org/html/2605.08837#A7.F10.4.2),[Figure 12](https://arxiv.org/html/2605.08837#A7.F12),[Figure 12](https://arxiv.org/html/2605.08837#A7.F12.5.2),[§G\.1](https://arxiv.org/html/2605.08837#A7.SS1.p1.1),[§1](https://arxiv.org/html/2605.08837#S1.p2.2),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[Figure 1](https://arxiv.org/html/2605.08837#S3.F1),[Figure 1](https://arxiv.org/html/2605.08837#S3.F1.3.2),[Figure 2](https://arxiv.org/html/2605.08837#S3.F2),[Figure 2](https://arxiv.org/html/2605.08837#S3.F2.3.2),[3\(a\)](https://arxiv.org/html/2605.08837#S3.F3.sf1),[3\(a\)](https://arxiv.org/html/2605.08837#S3.F3.sf1.3.2),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px3.p1.4),[§3](https://arxiv.org/html/2605.08837#S3.p2.1)\.
- O\. Hauk, I\. Johnsrude, and F\. Pulvermüller \(2004\)Somatotopic representation of action words in human motor and premotor cortex\.Neuron41\(2\),pp\. 301–307\.External Links:[Document](https://dx.doi.org/10.1016/S0896-6273%2803%2900838-9)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1)\.
- J\. Hoffmann, S\. Borgeaud, A\. Mensch, E\. Buchatskaya, T\. Cai, E\. Rutherford, D\. de Las Casas, L\. A\. Hendricks, J\. Welbl, A\. Clark, T\. Hennigan, E\. Noland, K\. Millican, G\. van den Driessche, B\. Damoc, A\. Guy, S\. Osindero, K\. Simonyan, E\. Elsen, O\. Vinyals, J\. W\. Rae, and L\. Sifre \(2022\)Training compute\-optimal large language models\.InProceedings of the 36th International Conference on Neural Information Processing Systems,NIPS ’22,Red Hook, NY, USA\.External Links:ISBN 9781713871088Cited by:[§6](https://arxiv.org/html/2605.08837#S6.p1.1)\.
- R\. Huben, H\. Cunningham, L\. R\. Smith, A\. Ewart, and L\. Sharkey \(2024\)Sparse autoencoders find highly interpretable features in language models\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=F76bwRSLeK)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p4.1),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- A\. E\. Kelly, Y\. N\. Kenett, J\. D\. Medaglia, J\. J\. Reilly, P\. Dudhat, and E\. G\. Chrysikou \(2024\)Conceptual structure of emotions\.Emotion24\(6\),pp\. 1550–1561\.External Links:[Document](https://dx.doi.org/10.1037/emo0001327)Cited by:[§A\.1](https://arxiv.org/html/2605.08837#A1.SS1.SSS0.Px1.p1.1),[§A\.1](https://arxiv.org/html/2605.08837#A1.SS1.SSS0.Px2.p1.5),[Table 3](https://arxiv.org/html/2605.08837#A1.T3),[Figure 7](https://arxiv.org/html/2605.08837#A3.F7),[Figure 7](https://arxiv.org/html/2605.08837#A3.F7.3.2),[Table 5](https://arxiv.org/html/2605.08837#A3.T5),[Table 5](https://arxiv.org/html/2605.08837#A3.T5.2.1),[Appendix C](https://arxiv.org/html/2605.08837#A3.p1.6),[Appendix F](https://arxiv.org/html/2605.08837#A6.p1.9),[Figure 11](https://arxiv.org/html/2605.08837#A7.F11),[Figure 11](https://arxiv.org/html/2605.08837#A7.F11.4.2),[§G\.2](https://arxiv.org/html/2605.08837#A7.SS2.p1.1),[§1](https://arxiv.org/html/2605.08837#S1.p2.2),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[3\(b\)](https://arxiv.org/html/2605.08837#S3.F3.sf2),[3\(b\)](https://arxiv.org/html/2605.08837#S3.F3.sf2.3.2),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px1.p1.1),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2605.08837#S3.SS1.SSS0.Px3.p1.4),[§3](https://arxiv.org/html/2605.08837#S3.p2.1)\.
- G\. Lakoff and M\. Johnson \(2003\)Metaphors we live by\.University of Chicago Press\.External Links:ISBN 9780226470993,[Link](http://dx.doi.org/10.7208/chicago/9780226470993.001.0001),[Document](https://dx.doi.org/10.7208/chicago/9780226470993.001.0001)Cited by:[§1](https://arxiv.org/html/2605.08837#S1.p1.1)\.
- B\. Li \(2026\)Incompressible knowledge probes: estimating black\-box llm parameter counts via factual capacity\.arXiv preprint arXiv:2604\.24827\.Cited by:[Figure 7](https://arxiv.org/html/2605.08837#A3.F7),[Figure 7](https://arxiv.org/html/2605.08837#A3.F7.3.2),[Figure 2](https://arxiv.org/html/2605.08837#S3.F2),[Figure 2](https://arxiv.org/html/2605.08837#S3.F2.3.2),[Figure 4](https://arxiv.org/html/2605.08837#S4.F4),[Figure 4](https://arxiv.org/html/2605.08837#S4.F4.3.2)\.
- K\. Li, A\. K\. Hopkins, D\. Bau, F\. Viégas, H\. Pfister, and M\. Wattenberg \(2023\)Emergent world representations: exploring a sequence model trained on a synthetic task\.InThe Eleventh International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=DeG07_TcZvT)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- T\. Lieberum, S\. Rajamanoharan, A\. Conmy, L\. Smith, N\. Sonnerat, V\. Varma, J\. Kramar, A\. Dragan, R\. Shah, and N\. Nanda \(2024\)Gemma scope: open sparse autoencoders everywhere all at once on gemma 2\.InProceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP,Y\. Belinkov, N\. Kim, J\. Jumelet, H\. Mohebbi, A\. Mueller, and H\. Chen \(Eds\.\),Miami, Florida, US,pp\. 278–300\.External Links:[Link](https://aclanthology.org/2024.blackboxnlp-1.19/),[Document](https://dx.doi.org/10.18653/v1/2024.blackboxnlp-1.19)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- D\. Lynott, L\. Connell, M\. Brysbaert, J\. Brand, and J\. Carney \(2020\)The Lancaster sensorimotor norms: multidimensional measures of perceptual and action strength for 40,000 English words\.Behavior Research Methods52,pp\. 1271–1291\.External Links:[Document](https://dx.doi.org/10.3758/s13428-019-01316-z)Cited by:[Table 9](https://arxiv.org/html/2605.08837#A8.T9.1.1.3),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1.p1.7)\.
- M\. Mancano and C\. Papagno \(2026\)Emotional and social dimension of abstract concepts meet with interoception in right anterior insula\.Journal of Neuroscience46\(2\),pp\. e0238252025\.External Links:[Document](https://dx.doi.org/10.1523/JNEUROSCI.0238-25.2025)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1)\.
- C\. McDougall, A\. Conmy, J\. Kramár, T\. Lieberum, S\. Rajamanoharan, and N\. Nanda \(2025\)Gemma scope 2 \- technical paper\.Technical reportGoogle DeepMind\.Note:[https://storage\.googleapis\.com/deepmind\-media/DeepMind\.com/Blog/gemma\-scope\-2\-helping\-the\-ai\-safety\-community\-deepen\-understanding\-of\-complex\-language\-model\-behavior/Gemma\_Scope\_2\_Technical\_Paper\.pdf](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/gemma-scope-2-helping-the-ai-safety-community-deepen-understanding-of-complex-language-model-behavior/Gemma_Scope_2_Technical_Paper.pdf)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1),[§5](https://arxiv.org/html/2605.08837#S5.p1.1)\.
- K\. McRae, G\. S\. Cree, M\. S\. Seidenberg, and C\. McNorgan \(2005\)Semantic feature production norms for a large set of living and nonliving things\.Behavior Research Methods37\(4\),pp\. 547–559\.External Links:[Document](https://dx.doi.org/10.3758/BF03192726)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2605.08837#S3.p1.1)\.
- N\. Nanda, A\. Lee, and M\. Wattenberg \(2023\)Emergent linear representations in world models of self\-supervised sequence models\.InProceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP,Y\. Belinkov, S\. Hao, J\. Jumelet, N\. Kim, A\. McCarthy, and H\. Mohebbi \(Eds\.\),Singapore,pp\. 16–30\.External Links:[Link](https://aclanthology.org/2023.blackboxnlp-1.2/),[Document](https://dx.doi.org/10.18653/v1/2023.blackboxnlp-1.2)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- N\. Panickssery, N\. Gabrieli, J\. Schulz, M\. Tong, E\. Hubinger, and A\. Turner \(2024\)Steering LLaMA 2 via contrastive activation addition\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15504–15522\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.828)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- S\. Pezzelle, E\. Takmaz, and R\. Fernández \(2021\)Word representation learning in multimodal pre\-trained transformers: an intrinsic evaluation\.Transactions of the Association for Computational Linguistics \(TACL\)9,pp\. 1563–1579\.External Links:[Link](https://aclanthology.org/2021.tacl-1.93/)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Pulvermüller \(2005\)Brain mechanisms linking language and action\.Nature Reviews Neuroscience6\(7\),pp\. 576–582\.External Links:[Document](https://dx.doi.org/10.1038/nrn1706)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1)\.
- G\. G\. Scott, A\. Keitel, M\. Becirspahic, B\. Yao, and S\. C\. Sereno \(2019\)The Glasgow norms: ratings of 5,500 words on nine scales\.Behavior Research Methods51,pp\. 1258–1270\.External Links:[Document](https://dx.doi.org/10.3758/s13428-018-1099-3)Cited by:[Table 9](https://arxiv.org/html/2605.08837#A8.T9.6.6.5),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1.p1.7)\.
- N\. Sofroniew, I\. Kauvar, W\. Saunders, R\. Chen, T\. Henighan, S\. Hydrie, C\. Citro, A\. Pearce, J\. Tarng, W\. Gurnee, J\. Batson, S\. Zimmerman, K\. Rivoire, K\. Fish, C\. Olah, and J\. Lindsey \(2026\)Emotion concepts and their function in a large language model\.External Links:2604\.07729,[Link](https://arxiv.org/abs/2604.07729)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1),[§5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1.p1.7)\.
- S\. Suresh, K\. Mukherjee, X\. Yu, W\. Huang, L\. Padua, and T\. T\. Rogers \(2023\)Conceptual structure coheres in human cognition but not in large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 722–738\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.47/)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- A\. Templeton, T\. Conerly, J\. Marcus, J\. Lindsey, T\. Bricken, B\. Chen, A\. Pearce, C\. Citro, E\. Ameisen, A\. Jones, H\. Cunningham, N\. L\. Turner, C\. McDougall, M\. MacDiarmid, C\. D\. Freeman, T\. R\. Sumers, E\. Rees, J\. Batson, A\. Jermyn, S\. Carter, C\. Olah, and T\. Henighan \(2024\)Scaling monosemanticity: extracting interpretable features from claude 3 sonnet\.Transformer Circuits Thread\.External Links:[Link](https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html)Cited by:[§H\.2](https://arxiv.org/html/2605.08837#A8.SS2.SSS0.Px2.p1.1),[§1](https://arxiv.org/html/2605.08837#S1.p4.1),[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1),[§5\.3](https://arxiv.org/html/2605.08837#S5.SS3.p1.1)\.
- C\. Tigges, O\. J\. Hollinsworth, A\. Geiger, and N\. Nanda \(2024\)Language models linearly represent sentiment\.InProceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP,Y\. Belinkov, N\. Kim, J\. Jumelet, H\. Mohebbi, A\. Mueller, and H\. Chen \(Eds\.\),Miami, Florida, US,pp\. 58–87\.External Links:[Link](https://aclanthology.org/2024.blackboxnlp-1.5/),[Document](https://dx.doi.org/10.18653/v1/2024.blackboxnlp-1.5)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- J\. Troche, S\. J\. Crutch, and J\. Reilly \(2017\)Defining a conceptual topography of word concreteness: clustering properties of emotion, sensation, and magnitude among 750 English words\.Frontiers in Psychology8,pp\. 1787\.External Links:[Document](https://dx.doi.org/10.3389/fpsyg.2017.01787)Cited by:[Appendix D](https://arxiv.org/html/2605.08837#A4.SS0.SSS0.Px1.p1.11),[§F\.2](https://arxiv.org/html/2605.08837#A6.SS2.SSS0.Px2.p1.9),[Appendix F](https://arxiv.org/html/2605.08837#A6.p1.9),[Figure 13](https://arxiv.org/html/2605.08837#A7.F13),[Figure 13](https://arxiv.org/html/2605.08837#A7.F13.5.2),[§G\.4](https://arxiv.org/html/2605.08837#A7.SS4.p1.1),[§H\.1](https://arxiv.org/html/2605.08837#A8.SS1.p1.1),[§1](https://arxiv.org/html/2605.08837#S1.p3.1),[§3](https://arxiv.org/html/2605.08837#S3.p1.1),[§3](https://arxiv.org/html/2605.08837#S3.p2.1),[Figure 4](https://arxiv.org/html/2605.08837#S4.F4),[Figure 4](https://arxiv.org/html/2605.08837#S4.F4.3.2),[§4](https://arxiv.org/html/2605.08837#S4.SS0.SSS0.Px1.p1.1),[§4](https://arxiv.org/html/2605.08837#S4.p1.1)\.
- S\. Trott \(2024\)Can large language models help augment english psycholinguistic datasets?\.Behavior Research Methods56\(6\),pp\. 6082–6100\.External Links:[Link](http://dx.doi.org/10.3758/s13428-024-02337-z),[Document](https://dx.doi.org/10.3758/s13428-024-02337-z)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- A\. M\. Turner, L\. Thiergart, G\. Leech, D\. Udell, J\. J\. Vazquez, U\. Mini, and M\. MacDiarmid \(2024\)Steering language models with activation engineering\.External Links:2308\.10248,[Link](https://arxiv.org/abs/2308.10248)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- C\. Villani, L\. Lugli, M\. T\. Liuzza, and A\. M\. Borghi \(2019\)Varieties of abstract concepts and their multiple dimensions\.Language and Cognition11\(3\),pp\. 403–430\.External Links:[Document](https://dx.doi.org/10.1017/langcog.2019.23)Cited by:[§3](https://arxiv.org/html/2605.08837#S3.p2.1)\.
- C\. Wang, Y\. Zhang, R\. Yu, Y\. Zheng, L\. Gao, Z\. Song, Z\. Xu, G\. Xia, H\. Zhang, D\. Zhao, and X\. Chen \(2025a\)Do LLMs “feel”? emotion circuits discovery and control\.External Links:2510\.11328,[Link](https://arxiv.org/abs/2510.11328)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- Y\. Wang, D\. Liang, and Y\. Zeng \(2025b\)Cognitive alignment between humans and llms across multimodal domains\.Note:Research Square preprintExternal Links:[Document](https://dx.doi.org/10.21203/rs.3.rs-5736241/v1),[Link](https://www.researchsquare.com/article/rs-5736241/v1)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Wei, Y\. Tay, R\. Bommasani, C\. Raffel, B\. Zoph, S\. Borgeaud, D\. Yogatama, M\. Bosma, D\. Zhou, D\. Metzler, E\. H\. Chi, T\. Hashimoto, O\. Vinyals, P\. Liang, J\. Dean, and W\. Fedus \(2022\)Emergent abilities of large language models\.Transactions on Machine Learning Research\.Note:Survey CertificationExternal Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=yzkSU5zdwD)Cited by:[§6](https://arxiv.org/html/2605.08837#S6.p1.1)\.
- X\. Wu, H\. Wang, Z\. Yan, X\. Tang, P\. Xu, W\. Siok, P\. Li, J\. Gao, B\. Lyu, and L\. Qin \(2025\)AI shares emotion with humans across languages and cultures\.External Links:2506\.13978,[Link](https://arxiv.org/abs/2506.13978)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px3.p1.1)\.
- Q\. Xu, Y\. Peng, S\. A\. Nastase, M\. Chodorow, M\. Wu, and P\. Li \(2025\)Large language models without grounding recover non\-sensorimotor but not sensorimotor features of human concepts\.Nature Human Behaviour9,pp\. 1871–1886\.External Links:[Document](https://dx.doi.org/10.1038/s41562-025-02203-8)Cited by:[§2](https://arxiv.org/html/2605.08837#S2.SS0.SSS0.Px2.p1.1),[§5\.2](https://arxiv.org/html/2605.08837#S5.SS2.p1.2)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing, H\. Zhang, J\. Gonzalez, and I\. Stoica \(2023\)Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.Vol\.36,Curran Associates, Inc\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/91f18a1287b398d378ef22505bf41832-Paper-Datasets_and_Benchmarks.pdf)Cited by:[§A\.1](https://arxiv.org/html/2605.08837#A1.SS1.p1.1)\.

## Appendix ALLMs\-as\-coders for property\-generation experiments

### A\.1Benchmarking LLMs\-as\-coders

In this section, we evaluate the performance of LLMs on the downstream task of automated property coding\. Our methodology builds upon the established “LLM\-as\-a\-judge” paradigm\[Zhenget al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib78)\]\. However, deploying LLMs as coders presents a more straightforward, and arguably more reliable, framework than using them as evaluative judges\. A typical LLM\-judge must typically surpass the candidate model in reasoning capabilities to accurately score its output, while a LLM\-coder merely has to be able to classify the properties properly\. This coding task is fundamentally a simplified variation of the rating task detailed in Section[4](https://arxiv.org/html/2605.08837#S4), a domain where we have already demonstrated robust LLM performance\. Furthermore, because this task involves classification rather than quality evaluation, it avoids the self\-enhancement bias typically observed in LLM\-as\-judge setups\[Zhenget al\.,[2023](https://arxiv.org/html/2605.08837#bib.bib78)\]\. This lack of self\-bias makes it methodologically sound to use the same model to both generate and subsequently code its own responses\.

#### Data collection and benchmark creation\.

To rigorously benchmark the models as coders, we needed a substantial dataset of coded properties\. Fortunately,Kellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]provides a large\-scale public dataset: the complete49,94249,942word\-property pairs produced in their experiment\. For experiment 1, no such word\-property pairs were provided, so an expert annotator labeled 2,077 pairs, which were generated by the older versions of the three frontier LLMs at the time \(Claude\-Opus 4\.6, GPT\-5\.1, Gemini\-3\.0\-Pro\)\. We utilized these two datasets as our primary benchmarks to identify the optimal LLM coder for each experiment\.

#### Validation metrics and human reference\.

For each candidate coder we report percentage agreement with the ground\-truth label and Cohen’sκ\\kappa\. The human reference for Experiment 1 is the76\.8%76\.8\\%two\-coder joint agreement reported byHarpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]on their original validation subset;Harpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]do not report aκ\\kappaceiling, so we leave that cell blank\. For Experiment 2, the human reference is the64\.25%64\.25\\%two\-coder agreement andκ=0\.52\\kappa=0\.52reported byKellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]\.

To ensure deterministic coding results, we use temperatureT=0\.0T=0\.0for all LLM\-coders\.

Table 2:Experiment 1 coder leaderboard\. Six LLM coders predict the 4\-category property label on the 2077\-pair ground truth\. % Agreement is overall percentage agreement,κ\\kappais Cohen’s kappa\. Per\-category columns are recall, i\.e\. the percentage of pairs labeled by the human as that category that the LLM also labeled as that category\.†Human ceiling is the 76\.8% two\-coder joint agreement reported byHarpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]\.Gemini\-2\.5\-Flash\-Litewas selected as the coder for all experiments with this dataset, unless otherwise specified\.Table 3:Experiment 2 coder leaderboard\. % Agreement andκ\\kappaare computed against the49,59249,592gold human\-coded labels\.†Human ceiling is the 64\.25% two\-coder agreement andκ=0\.52\\kappa=0\.52reported byKellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]\.Gemini\-2\.5\-Flashwas selected as the coder for all experiments with this dataset, unless otherwise specified\.
#### Error margins for Gemini\-2\.5\-Flash\-Lite as a coder\.

The Experiment 1 coder benchmark in Table[2](https://arxiv.org/html/2605.08837#A1.T2)measures property\-level agreement, but the Experiment 1 leaderboard actually consumes the per\-word category\-frequency vector and its Pearsonrragainst the human norms\. To estimate the error coder substitution introduces at that leaderboard level, we re\-coded the single hand\-coded run of three reference generation models \(Claude Opus4\.64\.6, Gemini3\.13\.1Pro, GPT\-5\.45\.4\) with both Flash\-Lite and the human ground truth on the exact same \(word, property\) pairs and bootstrapped the difference over words \(10001000paired iterations, same resampled words for both coders within an iteration so thatΔ=\(Flash\-Lite−Human\)\\Delta=\(\\text\{Flash\-Lite\}\-\\text\{Human\}\)is a paired statistic\)\. The Meanrra human coder would have produced has a cross\-model average difference from the LLM\-coder values ofΔ=\+0\.014±0\.019\\Delta=\+0\.014\\pm 0\.019, with every model95%95\\%CI containing zero\. On the same property pairs the cross\-model average frequencyΔ\\Delta’s are\+2\.3\+2\.3% Sensorimotor,−0\.1\-0\.1% Internal,−0\.5\-0\.5% Social,−1\.7\-1\.7% Verbal Association, with every per\-model95%95\\%CI containing zero\. So the per\-category attribution rates a human coder would have produced are within roughly±2\\pm 2percentage points of the Flash\-Lite values\.

## Appendix BExtended results for Experiment 1

Here we present extended quantitative results for the first experiment \(Section[3\.2](https://arxiv.org/html/2605.08837#S3.SS2.SSS0.Px1)\) based on theHarpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]dataset\.

### B\.1Detailed model setup

We evaluate2121LLMs spanning closed\-weight frontier and open\-weight families\. All generations use each provider’s default sampling temperature\. Coding is performed atT=0\.0T=0\.0for determinism \(Section[A\.1](https://arxiv.org/html/2605.08837#A1.SS1)\)\.

The closed\-weight models are accessed through their official APIs:Anthropic\(Claude Opus 4\.6, Claude Sonnet 4\.6, Claude Haiku 4\.5\),OpenAI\(GPT\-5\.4\), andGoogle\(Gemini 3\.1 Pro, Gemini 3 Flash, Gemini 3\.1 Flash\-Lite, Gemini 2\.5 Flash, Gemini 2\.5 Flash\-Lite\)\. The open\-weight models are run on 8x Quadro RTX 6000 GPUs, with model\-parallel sharding for the7070B and120120B systems:Meta\(Llama 3\.1 70B Instruct, Llama 3\.1 8B Instruct\),OpenAI open\-weights\(GPT\-OSS 120B, GPT\-OSS 20B\),Google open\-weights\(Gemma\-3\-27B\-IT, Gemma\-3\-12B\-IT, Gemma\-3\-4B\-IT\), andAlibaba Qwen\(Qwen3 8B, Qwen3 4B, Qwen3\-VL 30B, Qwen3\-VL 8B, Qwen3\-VL 4B\)\.

Each model produces1010independent runs over the full293293\-word Harpaintner stimulus set\. Per\-word category\-frequency vectors are averaged across runs before correlating with the human norms;±\\pmvalues reported in the leaderboard are bootstrap standard deviations over10001000resamples of the runs \(Section[E\.2](https://arxiv.org/html/2605.08837#A5.SS2)\)\.

### B\.2Detailed results for all 21 models

Figure[6](https://arxiv.org/html/2605.08837#A2.F6.4)presents the frequency profiles for all models\. Table[4](https://arxiv.org/html/2605.08837#A2.T4)details the performance of all 21 models evaluated in our panel\. Smaller models occasionally outperform substantially larger ones, suggesting that simply increasing parameter count does not organically close the grounding gap\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x7.png)

![Refer to caption](https://arxiv.org/html/2605.08837v1/x8.png)

![Refer to caption](https://arxiv.org/html/2605.08837v1/x9.png)

![Refer to caption](https://arxiv.org/html/2605.08837v1/x10.png)

Figure 6:Property\-frequency boxplots for all 21 LLMs, sorted by Mean r\.Table 4:Full experiment 1 leaderboard, all 21 models in the panel, sorted by Meanrr\. Ten generation runs per model \(Claude Haiku 4\.5: nine\)\. Bold marks the column max\.

## Appendix CExtended results for Experiment 2

Here we present extended quantitative results for the second experiment \(Section[3\.2](https://arxiv.org/html/2605.08837#S3.SS2.SSS0.Px2)\) based on theKellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]dataset\. The Kelly stimulus set partitions the357357target words into three matched conditions:118118abstract emotion words,118118abstract non\-emotion words, and119119concrete words\. The main text reports alignment on the combined236236\-word abstract subset \(emotion plus non\-emotion\)\. For completeness, in this appendix we also report the leaderboard restricted to the119119\-word concrete subset, which serves as a control condition: concrete nouns are the easy case for property generation, since they admit clearly perceptual properties that all candidate models recover well\.

Properties are classified by Gemini\-2\.5\-Flash into the four superordinate Kelly categories:Taxonomic,Entity,Situation, andIntrospective\. Each model produces1010runs over the full357357\-word stimulus set; per\-concept category\-frequency vectors are averaged across runs before correlating with the human reference distribution\. The primary metric is Pearsonrrbetween the per\-concept distributions of the model and the human aggregate, computed separately on the abstract and concrete subsets\.

Figure[7](https://arxiv.org/html/2605.08837#A3.F7)replicates the scaling plot for the Experiment 2 setup: even the strongest models reach only Meanr≈0\.36r\\approx 0\.36on the abstract subset, far below the human\-to\-human ceiling ofr≈0\.91r\\approx 0\.91, and there is no monotonic trend in parameter count\. Table[5](https://arxiv.org/html/2605.08837#A3.T5)lists every model on both subsets sorted by abstract Pearsonrr\. The pattern mirrors Experiment 1: the abstract\-subset gap remains roughly flat across model sizes and providers, while the same models reach much higher alignment \(r∈\[0\.29,0\.53\]r\\in\[0\.29,0\.53\]\) on the concrete subset\. This contrast is consistent with the central claim of the paper: the gap is specific to abstract concepts, not a general property\-generation deficit\. Figure[8](https://arxiv.org/html/2605.08837#A3.F8)confirms this from a different angle: on the concrete subset the human\-to\-model and model\-to\-model correlations sit in similar ranges, whereas on the abstract subset \(Fig\.[3](https://arxiv.org/html/2605.08837#S3.F3)in the main text\) models are much more correlated with each other than with humans\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x11.png)Figure 7:Mean r for all LLMs vs\. the human ceiling on Experiment 2\[Kellyet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]\. Parameter count for closed models is based on estimations fromLi \[[2026](https://arxiv.org/html/2605.08837#bib.bib74)\]\.Table 5:Extended leaderboard on experiment 2Kellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\], for both the abstract and concrete subsets\. Models are sorted by Pearsonrron the abstract subset\.![Refer to caption](https://arxiv.org/html/2605.08837v1/x12.png)Figure 8:Experiment 2 properties correlation heatmap for the concrete word subset\. The correlation between LLMs–humans is closer to LLMs\-LLMs in this case, compared to Fig\.[3](https://arxiv.org/html/2605.08837#S3.F3)\.
## Appendix DRating experiment additional results

#### Task\.

We replicate the rating study ofTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]using751751English nouns rated on1414Abstract Conceptual Grounding \(ACG\) dimensions under the uniform framing*“I relate this word to \[X\]”*, with responses on a11to77Likert scale \(11= Strongly Disagree,77= Strongly Agree\)\. Each model produces ten rating seeds; per\-word ratings are averaged across seeds before correlating with the human norms\. The primary metric is the Pearsonrraveraged over the1414dimensions\. Following the original study, the1414dimensions are organised into three components:Sensory\(Color, Taste/Smell, Tactile, Visual Form, Auditory\),Internal\(Emotion, Polarity, Social, Morality, Thought, Self\-Motion\), andMagnitude\(Space, Quantity, Time\)\. Table[6](https://arxiv.org/html/2605.08837#A4.T6)reports the per\-component Pearson correlations alongside the overall Meanrrfor every model out of the2121\.

#### Results\.

Three patterns stand out in Table[6](https://arxiv.org/html/2605.08837#A4.T6), and each one provides a different angle on the grounding gap reported in the main text\.

The first observation is that the strongest models are essentially at the human ceiling on this task\. Claude Sonnet 4\.6 reaches Meanr=0\.760r=0\.760against a cross\-dimension human ceiling of approximately0\.780\.78\(Appendix[F](https://arxiv.org/html/2605.08837#A6)\), and the next six models in Table[6](https://arxiv.org/html/2605.08837#A4.T6)all sit within0\.040\.04of that mark\. The same models top out atr≈0\.37r\\approx 0\.37on the property generation task of Experiment 1 against a ceiling of0\.970\.97\(Table[4](https://arxiv.org/html/2605.08837#A2.T4)\), an order\-of\-magnitude larger gap to the human reference\. We take the contrast as direct evidence for the central claim of the paper: when a grounding dimension is named explicitly in the prompt, current LLMs can recover it; the gap is therefore specific to the spontaneous property generation setting and is not a general inability to register sensorimotor, internal, or social content\.

The second observation is that scaling helps recognition, but only on a subset of the components\. Within every open\-weight family we tested, the Sensory component improves dramatically with model size: Gemma scales from−0\.086\-0\.086at44B to0\.7040\.704at2727B, Llama scales from−0\.161\-0\.161at88B to0\.7290\.729at7070B, and Qwen3 scales from0\.3210\.321at44B to0\.8070\.807at88B \(Table[6](https://arxiv.org/html/2605.08837#A4.T6)\)\. The Magnitude component shifts up by a comparable margin within each family\. The Internal and Social components, by contrast, are essentially flat across model sizes: every model in Table[6](https://arxiv.org/html/2605.08837#A4.T6)sits in the range\[0\.58,0\.84\]\[0\.58,0\.84\]on these two components, regardless of parameter count or provider\. Recognition of affective and social content is therefore not a capability bottleneck at any scale we tested\. Perceptual grounding, on the other hand, is recoverable only at sufficient scale, and once that scale is reached the model can rate perceptual content as well as humans do\.

The third observation is that the Motor sub\-dimension is a persistent weakness that scaling does not close\. As shown in Table[6](https://arxiv.org/html/2605.08837#A4.T6), the strongest closed\-weight models all score weakly on Motor, with Gemini 3\.1 Pro at0\.1190\.119, Gemini 3 Flash at0\.1890\.189, Claude Opus 4\.6 at0\.3930\.393, and Gemini 3\.1 Flash\-Lite even falling below zero at−0\.058\-0\.058\. More importantly, the within\-family scaling that lifts Sensory does not lift Motor: Gemma improves only modestly from0\.4910\.491at44B to0\.6140\.614at2727B \(with1212B sitting higher than2727B at0\.6930\.693\), Llama actually drops from0\.5280\.528at88B to0\.4950\.495at7070B, and Qwen3 drops from0\.5400\.540at44B to0\.3750\.375at88B\.

Table 6:Rating experiment results for all 21 models\. Meanrris the average of the 14 per\-dimension Pearson correlations against the human ratings\.

## Appendix EVariance and uncertainty in Experiment 1

The Experiment 1 leaderboard reports each model’s Mean Pearsonrras the point estimate from1010independent generation runs, plus a bootstrap standard deviation\. Two questions a reader is right to ask: \(i\) is a1010\-run point estimate close to the asymptote that the same model would reach with infinitely many runs, and \(ii\) what does the bootstrap±\\pmbar actually quantify? We answer both empirically by runningGemma\-3\-4B\-IT100 times on the full 293\-word Harpaintner pool, treating that aggregate as the ground truth, and comparing it to what we would have estimated from random1010\-run subsets of the same data\.

### E\.1Convergence: how accurate is the 10\-run estimate?

Figure[9](https://arxiv.org/html/2605.08837#A5.F9)plots Meanrras a function of the number of runs averaged, with shaded±1\\pm 1std bands over5050random subsets at each subset size\. The estimate climbs monotonically toward the 100\-run asymptote ofr=\+0\.322r=\+0\.322\. At the Experiment 1 leaderboard cadence of1010runs, the average estimate isr=\+0\.309r=\+0\.309, lower than the asymptote byΔ​r=\+0\.013\\Delta r=\+0\.013\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x13.png)Figure 9:Mean Pearsonrrvs\. number of runs averaged, on Experiment 1, forGemma\-3\-4B\-ITcoded by Gemini\-2\.5\-Flash\-Lite\. Solid line: mean over5050random subsets at each subset size; band:±1\\pm 1std across subsets; dashed: 100\-run asymptote; dotted: the1010\-run cadence used by the Experiment 1 leaderboard\.The curve has two practical implications\.*For absoluterrvalues on a single model*, a1010\-run point estimate sits about0\.010\.01below the asymptote it would converge to, so quotedrrvalues are slightly conservative\.*For ranking models against each other*, this bias is approximately constant across models, so the relative ordering on the1010\-run Experiment 1 leaderboard is stable\.

### E\.2Bootstrap calibration

The Experiment 1 leaderboard’s±\\pmvalue next to each model’s Meanrris a bootstrap standard deviation computed fromNboot=1000N\_\{\\text\{boot\}\}=1000resamples of the1010runs with replacement\. For each resample we re\-aggregate the per\-word frequencies, recompute Meanrragainst the human norms, and take the standard deviation of the resulting distribution\.

## Appendix FEstimation of human correlation ceilings

Each experiment reports a per\-category Pearsonrrhuman ceiling, defined as the correlation between two independent panels ofNNraters scored on the same stimuli\. This is the largestrra noiseless model could expect to reach against the published norms, which are themselves panel means\. We obtain it in one of two ways depending on what each original study released\.Kellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]provides participant\-level responses, so we compute the ceiling by direct resampling\.Harpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]andTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]release only per\-word means and SDs, so we estimate the ceiling from those aggregates using the standard ICC\+\+Spearman–Brown chain:ICC1\\mathrm\{ICC\}\_\{1\}recovers single\-rater reliability from the ratio of between\-word to total variance, and Spearman–Brown projects single\-rater reliability up to theNN\-vs\-NNscale viarfull=2​rsplit/\(1\+rsplit\)r\_\{\\text\{full\}\}=2r\_\{\\text\{split\}\}/\(1\+r\_\{\\text\{split\}\}\)\. We always reportrfullr\_\{\\text\{full\}\}, because the human norm against which models are correlated is a full\-panel mean rather than a single rater\.

### F\.1Experiment 2: empirical split\-half

We draw10001000random half\-splits of the participants per concept \(medianN≈30N\{\\approx\}30raters across357357concepts\), compute per\-concept property\-category percentages on each half, and average Pearsonrracross splits\. Spearman–Brown then projects the half\-panel correlation torfull=0\.909r\_\{\\text\{full\}\}=0\.909on the 236\-word abstract subset used in the main text andrfull=0\.968r\_\{\\text\{full\}\}=0\.968on the 119\-word concrete subset used in the appendix\.

### F\.2Validation of the ceiling estimation

Experiment 2 is the only dataset where both methods are available on the same data, since the participant\-level release also yields per\-word means and SDs\. Applying the ICC\+\+SB chain we use for the other experiments to the Experiment 2 category\-frequency matrix gives a within\-dataset cross\-check on the estimate\.

Table 7:ICC\+\+Spearman–Brown analytical estimate vs\.10001000\-split empirical ceiling on Experiment 2, both atrfullr\_\{\\text\{full\}\}scale\.The two methods agree to within0\.0070\.007on both slices, with the analytical estimate sitting fractionally above the empirical ceiling\. We therefore consider the ICC\+\+SB ceilings reported below as accurate to within∼0\.01\{\\sim\}0\.01\.

#### Experiment 1 ceilings\.

Harpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]publish per\-word frequencies in four aggregate categories\. The resultingrfullr\_\{\\text\{full\}\}ceilings are:

Sensorimotor:​0\.966,Internal:​0\.964,Social:​0\.971,Verbal:​0\.992\(r¯full=0\.974\)\.\\text\{Sensorimotor: \}0\.966,\\quad\\text\{Internal: \}0\.964,\\quad\\text\{Social: \}0\.971,\\quad\\text\{Verbal: \}0\.992\\quad\(\\overline\{r\}\_\{\\text\{full\}\}=0\.974\)\.

#### Rating\-experiment ceilings\.

Trocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]publish per\-word means and SDs on each of the 14 grounding dimensions\. We apply ICC\+\+SB dimension\-by\-dimension withσbetween2\\sigma^\{2\}\_\{\\text\{between\}\}from the cross\-stimulus variance of per\-word means,σwithin2\\sigma^\{2\}\_\{\\text\{within\}\}from the mean squared SD, andNNfrom the median raters per word \(≈24\\approx 24\)\. The 14rfullr\_\{\\text\{full\}\}ceilings range from0\.620\.62on Polarity to0\.910\.91on Visual Form \(cross\-dimension mean0\.780\.78\)\.

## Appendix GPrompting setup

This section presents verbatim every prompt used on the three experiments\. The literal token\{word\}is substituted by each target concept on each call, and\{properties\}\(in the coding prompts\) by the comma\-separated list of generated properties for that concept\.

### G\.1Property Generation Prompt \(Experiment 1\)

The generation prompt for Experiment 1 is shown below in Figure[10](https://arxiv.org/html/2605.08837#A7.F10)\. It asks the model for exactly four spontaneous properties per noun, with one in\-context example \(sympathy\)\. It follows the original property listing instructions ofHarpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\]\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x14.png)Figure 10:Property generation prompt for Experiment 1Harpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\], reproduced verbatim\. The placeholder\{word\}is replaced by each of the target nouns at inference time\.
### G\.2Property Generation Prompt \(Experiment 2\)

The generation prompt for Experiment 2 is shown below in Figure[11](https://arxiv.org/html/2605.08837#A7.F11)\. It asks the model for exactly five spontaneous clues per noun, with two in\-context examples \(doganddejected\)\. It follows the human rater instructions in Appendix A ofKellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\]\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x15.png)Figure 11:Property generation prompt used for Experiment 2Kellyet al\.\[[2024](https://arxiv.org/html/2605.08837#bib.bib9)\], reproduced verbatim\. The placeholder\{word\}is replaced by each target concept at inference time\.
### G\.3Coding Prompt \(Experiment 1\)

Each generated property is assigned one of five mutually exclusive categories:Sensorimotor feature\(a feature experienced by the senses\);Social constellation\(the coexistence or interaction of persons\);Internal state and emotion\(internal cognitive processes such as motivation, emotion, volition\);Association\(thematically or symbolically related but not descriptive\); andOther abstract concept\(an abstract feature that describes the concept\)\. After coding, the two abstract categories:AssociationandOther abstract conceptare merged to form the final abstract categoryVerbal association\. The prompt provides a single in\-context example \(thesympathyproperties from the generation prompt, one per category\) and asks the coder to emit one labelled line per property\. The full coding prompt is shown in Figure[12](https://arxiv.org/html/2605.08837#A7.F12)\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x16.png)Figure 12:Coding prompt for Experiment 1\. Categories are listed verbatim fromHarpaintneret al\.\[[2018](https://arxiv.org/html/2605.08837#bib.bib8)\];\{word\}and\{properties\}are replaced at inference time by the concept and the comma\-separated generated property list\.
### G\.4Rating Experiment Prompt

The rating experimentTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]uses 14 affective conceptual grounding dimensions:color, taste/smell, tactile, visual form, auditory, emotion, polarity, social, morality, thought, self\-motion, space, quantity, time\. Every dimension uses the same instruction skeleton, with only the dimension specific statement substituted in\. The full prompt for theemotiondimension is shown verbatim in Figure[13](https://arxiv.org/html/2605.08837#A7.F13); the other thirteen dimensions are produced by replacing the sentence “I relate this word with human emotion\.” with the analogous sentence for the target dimension, taken verbatim fromTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]\. The output is a list of\(’word’, INT\_SCORE\)tuples on a 1–7 Likert scale\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x17.png)Figure 13:Rating prompt for the Troche benchmark, shown for theemotiondimension\. The same instruction skeleton is reused for all 14 dimensions; only the target sentence \(“I relate this word withX\.”\) changes between dimensions, with the wording for each dimension taken verbatim fromTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]\.

## Appendix HExtended mechanistic analysis results

### H\.1Feature identification dataset construction

This appendix documents how we built the426426\-noun balanced corpus used as the candidate pool for SAE feature identification \(Section[5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1)\)\. Published norm sets cover individual dimensions of abstract\-concept structure but no single resource gives us four balanced experiential categories on the same vocabulary, as well asTrocheet al\.\[[2017](https://arxiv.org/html/2605.08837#bib.bib2)\]does\. But we did not want to use our test\-set to discover features, so we had to look at other norms\. Additionally, we wanted to use only the very highest words per category, and this meant that we neeeded a larger dataset to subsample only the highest ones\. We therefore built our own noun corpus by combining four complementary datasets\. Each category is anchored to one primary norm and a single thresholded metric as shown in Table[9](https://arxiv.org/html/2605.08837#A8.T9)below, so that only the highest\-scoring words per category are selected\. A word that is not a noun or that satisfies thresholds for more than one grounded category is dropped from both\. The final corpus comprises of 110 sensorimotor, 102 internal state, 102 social and 112 abstract nouns, for a total of 426 nouns\.

Table 8:Norm sources and the metric extracted from each\.Table 9:Feature\-identification corpus composition\. Sensorimotor is further partitioned into1111Lancaster sub\-modalities at1010words each \(Visual, Auditory, Olfactory, Gustatory, Haptic, Interoception, Mouth, Head, Hand\_arm, Foot\_leg, Torso\), enabling per\-submodality stratification at the splitting stage\.#### Sentence generation

Each of the426426nouns is paired with 10 context sentences that elicit the noun as the next token\. Sentences are produced with Gemini\-3\.13\.1\-pro\-preview at temperature0\. For example, forw=w\{=\}*surprise*one of the generated sentences is

> “If you are feeling sad because you think everyone forgot your birthday, but then they jump out of the dark to make you smile, they have planned a \_\_\_”\.

The blank is the position where the target noun was originally written\.

The prompt instructs Gemini to write1010numbered sentences ending with the target word, framed as a one\-line riddle that makes the missing word easy to recover, and themed by the noun’s category:

> Write1010numbered sentences about everyday life that end with the word ‘<WORD\>’\. Your goal in writing these sentences is to make sure that it is obvious that the sentence ends with this word, like a very simple and easy riddle to learn the word\. You must also make sure that the sentence is aligned with the theme of ‘<THEME\>’\. Start your answer with ‘11\.<the first sentence\>’\.

The string for<THEME\>is filled per subcategory \(for Sensorimotor: Visual→\\to“visual perception”, Auditory→\\to“auditory perception”, etc\.\) or per category \(Internal, Social\)\. For Abstract nouns the theme clause is dropped: there is no grounded modality to anchor on, so we let Gemini choose any framing\.

A regex post\-processor then strips the target noun from the end of each generated line and verifies it is the same as the intended target\. The final3,5213\{,\}521successful rows generated for the426426training nouns are released publicly\.

#### Last\-word\-target probing: per\-noun activation signatures

Once the dataset is built, for each sentence inS​\(w\)S\(w\), we remove the target nounwwand feed the resulting prefix to the model\. The SAE activation is read at the last token of the prefix, i\.e\. the position where the model is about to emitwwgiven everything before it\. Per \(layer, feature\) we take the median across the1010sentences\. Any feature that has a non\-zero median for five or more nouns from a single category is considered as a candidate for that category\.

#### Why this construction\.

The setup ensures \(i\) The SAE sees a natural mid\-sentence residual and therefore stays inside its training distribution, unlike isolated\-token probes where many SAE features go silent\. \(ii\) The residual at positionn−1n\{\-\}1is a read\-out of the pre\-next\-token distribution conditioned on the preceding context, which is target\-word\-specific without requiring the model to actually emit the target\. \(iii\) The aggregation is median over1010diverse context sentences, which suppresses any single sentence’s effect\.

#### Identified features per layer\.

In Figure[14](https://arxiv.org/html/2605.08837#A8.F14)we see the features that have at least one non\-zero median noun activation in our dataset and the ones that have at least five activations from a single category: the candidates\. The features that have only a single activation are of course substantially more: either they fired randomly, or they are very specific to that noun\. We observe an interesting phenomenon: until layers 15\-17 the model seems to be encoding mostly and understanding the context of the previous and current tokens, so it does not have many features that correspond to the next token content which is our interest\. From layer 17 \(where it has about 1500 features\) and after we see that the number of specific features increases rapidly and at layer 24 it reaches 3500 features and stays constant, indicating that the model starts making decisions about the next word\. On the other hand, the candidate features we selected exhibit a different behavior: they increase slightly but noticeably after layer 15 and then they remain constant, until a sudden step increase in the last layer\. We hypothesize that this is because these features are mostly responsible for category\-level decisions that have to do with the theme and context of the prefix, as intended by their selection process, and not so much for specific decisions dictated by linguistic or reasoning rules\.

![Refer to caption](https://arxiv.org/html/2605.08837v1/x18.png)Figure 14:Number of SAE features per layer identified with our detection algorithm in Gemma\-3\-4B\.

### H\.2Feature validation

#### Feature interpretation\.

Determining the precise functionality of the features highly correlated with grounding dimensions remains challenging, as most lack an obvious semantic meaning\. To facilitate interpretation and visualization, we provide links to Neuronpedia for representative features within each category\. Additionally, we report automatic interpretations derived from our dataset activations using Claude\-Opus\-4\.6\. Note that the highest\-correlated features discovered in our analysis are often not hosted on Neuronpedia, as the platform currently only supports a subset of layers \(i\.e\., layers 9, 16, 22, and 29\)\. Consequently, we present the best\-performing features currently available\.

Importantly, while Neuronpedia typically displays the highest positive and negative logits for each feature, our analysis focuses on the highest positive*next\-token*logits \(i\.e\., the logits predicted for the token immediately following the active feature\)\. This distinction stems from our dataset’s construction and aligns more accurately with our primary task of property\-generation\.

Representative features from each category available on Neuronpedia:

- •
- •
- •
- •
- •

![Refer to caption](https://arxiv.org/html/2605.08837v1/x19.png)

![Refer to caption](https://arxiv.org/html/2605.08837v1/x20.png)

Figure 15:Boxplot for steered model with sensory \(top\) and internal \(bottom\) features\.
#### Feature steering\.

For steering, we use clamping\[Templetonet al\.,[2024](https://arxiv.org/html/2605.08837#bib.bib39)\]and we set the activation to double the median of the highest\-activating noun from our identification dataset \(Section[5\.1](https://arxiv.org/html/2605.08837#S5.SS1.SSS0.Px1)\)\. While this heuristic ensures a clear signal, it yields inconsistent intervention strengths\. Specifically for the internal\-feature, we found it had unususally high max\-activation and we kept it at 1x, because otherwise the model did not follow instructions properly\.

In Figure[15](https://arxiv.org/html/2605.08837#A8.F15)we present the boxplots for the Sensory and Internal features from above, compared against the baseline model\. Both features lead to substantial shifts of the distribution of generated properties towards their target category, while the other categories decline or remain almost constant\.