acceptodds
Under review as a conference paper at ICLR 2027

A Gärdenforsian Perspective on the Conceptual Geometry of LLMs

Abstract

G\"ardenfors' conceptual spaces offer a geometric language for describing conceptual knowledge. We show that this cognitive theory can be approximately realized in the representation geometry of LLMs. Specifically, we propose the G\"ardenforsian Representation Hypothesis (GRH), which characterizes LLM representations across multiple levels of conceptual abstraction, from quality dimensions (graded attributes) and domains to properties (adjective), category concepts, and prototypes. GRH unifies existing representation hypotheses by showing that they can be understood as instantiations of a common conceptual geometry, each capturing a different level of conceptual abstraction. We test GRH by progressively reconstructing such a conceptual space from LLM representations. Experiments on THINGSplus and Binder show that the recovered conceptual geometry closely aligns with human conceptual organization, providing evidence that LLM representations naturally admit a G\"ardenfors'-style conceptual-space structure.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Conceptual Space, Interpretability, Concept Geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Pb8zOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes the Gärdenforsian Representation Hypothesis (GRH), a layered view of conceptual geometry in LLM representations. The framework organizes conceptual structure into five linked levels: quality dimensions as linear directions, semantic domains as subspaces, properties as convex regions within domains, concepts as regions spanning multiple domains, and prototypes as geometric centers that predict graded typicality. The authors progressively reconstruct these structures from LLM representations and evaluate the five hypotheses using human semantic annotations from THINGSplus and Binder Feature Norms. The analysis is repeated across Llama-3.1-8B, Gemma-7B, and Mistral-7B, with broadly consistent results across models.

Critical strengths
  1. The paper provides a coherent and genuinely informative unification of several previously separate geometric views of concepts. I found the progression from quality directions to domains, property regions, multi-domain concepts, and prototypes particularly elegant. Rather than treating these geometric objects as unrelated observations, the paper organizes them into a single layered account and tests each level empirically. This makes the work substantially more interesting to me than an incremental representation-analysis paper.

  2. The empirical program is broad and reasonably systematic. The authors do not rely on a single model or a single geometric statistic. The full reconstruction pipeline is repeated independently across three LLM backbones, and several stages include meaningful controls such as unseen-category generalization, wrong-domain or random-subspace comparisons, and random domain decompositions. I especially appreciate that the paper attempts to validate the semantic meaning of the recovered geometry rather than only visualizing it.

Critical weaknesses

The main limitation, in my view, is that the evidence is not fully independent of the human conceptual structure that the paper aims to recover. Human ratings are used to identify quality directions, human-defined domains are used as an evaluation target, thresholded human ratings define several properties, and category membership is used to estimate concept regions and prototypes. As a result, I view the experiments as strong evidence that LLM representations are compatible with and can support a Gärdenforsian organization, but somewhat weaker evidence that this complete conceptual organization emerges naturally or independently inside the model. This distinction matters for the strongest interpretation of the paper, although it does not overturn my overall positive assessment.

Questions for authors
  1. How do the authors distinguish “recoverability under human supervision” from “naturally emergent conceptual geometry”? Several stages of A1–A5 are constructed using human semantic ratings or category membership. I would be interested in the authors’ view on which parts of GRH are genuinely discovered from the representation geometry, and which parts should instead be interpreted as showing that the geometry can support a human-specified conceptual organization. A more explicit discussion of this distinction would strengthen the central claim.

  2. How essential is the higher-level GRH geometry relative to simpler alternatives? For A3, the single-direction half-space slightly outperforms the domain-level convex region, which is understandable because the properties are defined by thresholding scalar ratings. For A5, the prototype-typicality correlations are consistently positive but modest, and the raw hidden-space centroid is stronger than the proposed quality-space prototype for Llama. I would therefore appreciate a clearer discussion of which parts of the higher-level geometry are empirically necessary, rather than merely compatible with the data.

Reviewer_w3kLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose GRH, a LLM representation hypothesis. It aims to explain how concepts are organized in LLM representation hidden spaces. From the Gardenforsian theory, 5 levels of human cognition architecture is used to interpret the LLM representations geometry: quality dimensions as linear directions, domains as sub-spaces, properties as convex regions, concepts as multi-domain regions, and prototypes as geometric centers. The authors evaluate the framework using human semantic judgement from THINGSplus and Binder on three LLM backbones. The results show alignment between the conceptual geometry and human judgement.

Critical strengths
  1. This is an interesting topic to interpret the LLM representation. This paper tries to connect the human cognition theory with LLM representation hidden space. Rather than studying individual semantic concepts, it studies how the semantic can be organized into difference conceptual levels.
  2. Systematic evaluation across difference conceptual levels. This paper evaluates five proposed conceptual levels with three LLM backbones. These experiments provide useful evidence that LLM representations contain semantic structures aligned with human judgments.
Critical weaknesses
  1. The experiments are not convincing that the proposed conceptual geometry naturally exists in LLM representations. The authors use human rating to estimate the semantic directions, cluster directions into domains, and construct convex regions and prototypes. However, constructing a geometry that fits human judgement is different from showing this geometry exists in the original LLM representations. For example, convexity of property regions is imposed through ellipsoid fitting rather than independently verified. I am not sure if the experiments are sufficient to validate the proposed hypothesis.
Questions for authors
  1. Does the proposed framework consistently outperform the simpler frameworks? For example in Table 11, the Raw hidden centroid of Llama, axis-scaled quality space of Gemma and mistral are highlighted.
  2. The proposed five-level conceptual structure is constructed using human-labeled quality dimensions, spectral clustering, and fitted convex regions and prototypes. How can the authors distinguish between discovering an intrinsic conceptual structure in LLM representations and imposing this structure through the construction procedure?
Reviewer_Xr9gOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Gardenfors was right (for LLMs)

Critical strengths

lots of nice pictures, i'm a sucker for those. very nice that prototype theory is recovered too!

Critical weaknesses

not a weakness of the methodology nor of the paper, it's something that the convexity consideration of gardenfors creates a bind for: how do you deal with comparatives like "bigger than" or "faster than"?

Questions for authors

[Note: I did the first pass with my own faulty eyes and brain, no AI use on my side yet, I want to have a discussion with you first, and please don't read into whatever my scores are yet.]

  1. Do you recover geometric structure within integral domains? e.g. cycles for colour?
  2. How (if at all) does convexity hold up for n-ary concepts and relations? Did you look at any of those?
  3. Do you have any insights on the representation of metaconcepts, like negation?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank the reviewers for their feedback. We uploaded a revised manuscript with new experiments added in Sections 4.3 and 4.7.

Summary of Common Reviewer Questions & Core Clarifications

  • Intrinsic/Discovered vs. Imposed Geometry (Reviewers w3kL & Pb8z). Both reviewers asked whether the recovered conceptual geometry is intrinsic to the LLM representations or imposed by the construction procedure. We study intrinsic or emergent geometry by following the readout approach used by Park et al. in the Linear Representation Hypothesis [1] and by Xiong in the Lattice Representation Hypothesis [2]. We keep the model representration geometry fixed and use human ratings and category labels to only estimate quality directions, concept regions, and prototypes. We added two additional controls to test whether the same procedure gives comparable results when we change the annotation assignments or the representations. Our controlled analysis further confirms that the recovered geometry is driven by the learned LLM representations instead of the construction procedure itself.

  • Simpler Baselines vs. Controlled Baselines (Reviewers w3kL & Pb8z). Simpler alternatives show what is sufficient for a given task. A single-direction half-space is convex, and a raw hidden-space centroid is a prototype, so their strong performance is consistent with A3 and A5. Controlled baselines test what the specific recovered geometry contributes. As shown in Table 11, replacing the recovered domains with same-rank random subspaces lowers held-out concept-membership AUROC from 0.874–0.879 to 0.631–0.746 across the three backbones. This demonstrates that our recovered domain structure provides essential empirical leverage that cannot be replicated by arbitrary projections.

 

Additional Controlled Experiments in the Revised Manuscript

  • Label permutation and input embedding controls (Section 4.7). We shuffled entire annotation rows across objects in the construction split 199 times while keeping the evaluation annotations fixed. Table 5 shows that all six scores systematically collapsed. We also ran the exact same procedure on Llama-3.1-8B’s input embeddings versus final-layer representations, where the final layer scored substantially higher on all six measures. These controls show that the construction procedure alone almost completely fails (collapsing to chance or near-chance levels) when decoupled from true semantic correspondence, confirming that the recovered geometry is driven by the learned LLM representations instead of the construction procedure itself."

 

Selected results from Table 5. Shuffled annotations and input embedding controls. Shuffled scores are means ± standard deviations over 199 permutations. A2 uses Binder. The other rows use THINGSplus.

MetricOriginal annotationsShuffled annotationsInput embeddings
A1 Held-out ρ0.6740.000 ± 0.0270.388
A1 Category-disjoint ρ0.443−0.001 ± 0.0140.114
A2 Domain NMI0.6600.289 ± 0.0250.483
A3 Property AUROC0.9810.500 ± 0.0150.624
A4 Concept AUROC0.8740.503 ± 0.0160.675
A5 Typicality ρ0.168−0.052 ± 0.047−0.024

 

  • Convexity test without ellipsoid fitting (Section 4.3). We added approximate betweenness and exact convex-hull checks directly on held-out object coordinates and ratings. As shown in Fig. 6, high-rated objects on both datasets cluster substantially more often along segments between high-rated examples and inside their convex hulls within recovered domains than in either random or wrong-domain controls. These results demonstrate that approximate convexity is an inherent spatial property of the LLM representation space, instead of just an artifact imposed by ellipsoid fitting."

 

Selected results from Fig. 6. Betweenness compares the share of high-rated cases using the same number of held-out cases per property in each domain. The hull rows report rating proportions inside the recovered-domain hulls.

CheckCondition or rating groupTHINGSplusBinder
BetweennessRecovered domain60.5%73.5%
Same-rank random domain27.3%29.4%
Incorrect domain28.7%26.6%
Exact convex hullUpper-quartile ratings56.7%65.3%
Lower-quartile ratings2.8%3.2%

 

References

[1] Park, K., Choe, Y. J., and Veitch, V. (2024). The Linear Representation Hypothesis and the Geometry of Large Language Models. ICML 2024.

[2] Xiong, B. (2026). The Lattice Representation Hypothesis of Large Language Models. ICLR 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.