acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Lattice: A Graded Representation Hypothesis for LLM Concept Hierarchies

Abstract

The Lattice Representation Hypothesis holds that large language models (LLMs) encode concept hierarchies as intersections of attribute half-spaces. Testing it on SNOMED CT, an ontology whose defining attributes are human-curated, we find that half-spaces are recovered reliably, but the lattice's inclusion score measures attribute overlap rather than subsumption: it orients parent-child pairs at only 59.5% and barely separates siblings from parents. Much of the directional information is lost in the inclusion aggregation, but a single linear direction in the embedding recovers it. In this paper, we propose the Graded Representation Hypothesis: hierarchy is realized by a relatedness subspace (the lattice) together with a specificity direction whose projection assigns each concept a grade, a real-valued grading that approximately preserves the order of the is-a poset. Experiments on SNOMED CT and WordNet show that the learned grade correctly orients 94% of held-out is-a edges, generalizes to unseen subtrees, transfers across ontologies without retraining, and, combined with one relatedness score, accounts for most of the linearly decodable subsumption signal with 68 parameters. These results suggest that LLMs encode not only which concepts are related but how abstract each concept is, as a single linear coordinate. Code will be available.

Primary area:
interpretability and explainable AI
Keywords:
language model interpretability, representation geometry, concept hierarchies, linear probing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qpWqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether LLM embeddings contain the hierarchy between concepts, such as knowing that Bus is a kind of Vehicle. Using SNOMED CT, it studies whether concept attributes can be recovered from embeddings, whether one common direction can separate child concepts from their parents, and whether this direction together with concept similarity can predict subsumption relations. It also tests the same idea across other LLMs and WordNet.

Critical strengths
  1. The hierarchy-direction result is empirically interesting and is observed via fairly strong tests. A low-capacity principal component (PC) subspace correctly orients 94.2% of held-out child-parent edges, and still reaches 90.3% when entire branches are held out. Shuffled labels fall to chance, while a random direction and PC1 perform much worse. Similar results are obtained with the other tested encoders, and the direction also transfers between SNOMED and WordNet without retraining on the target ontology. This is useful evidence that frozen LLM embeddings contain a shared directional signal related to hierarchy .... i.e., the result is not just memorization of individual edges. I think this is a real empirical finding, even though I do not think it is sufficient to establish an ontology or lattice embedding.

  2. The paper gives a useful diagnosis of why flat attribute inclusion fails on SNOMED. The ExactFlat experiment is particularly informative. Under the flat attribute representation, 65.3% of true hierarchy edges appear incomparable. Once the hierarchy among destination values is restored, this falls to 0% and oracle F1 rises to 0.996. This clearly shows that the hierarchy of attribute values matters and that a flat FCA representation can lose important subsumption information. I found this diagnostic useful even though it also raises concerns about using the flat FCA construction to make broader claims about the ontology.

Critical weaknesses
  1. What I could get from the paper is that the strongest result is a supervised child-parent prediction. This is not the same as what the paper claims -- i.e., an evidence for a hierarchy structure inside the LLM. PCA first finds the main alignment patterns across the sampled concept embeddings. For a child-parent pair, the difference vector δ=xC−xP\delta=x_C-x_P then tells us how that change aligns with each of these PCA patterns. Logistic regression uses known child-parent labels to decide how much each PCA pattern should count, and combines them into the final direction uu. In other words, uu is not itself a direction discovered directly from the LLM geometry. It is a supervised mapping that makes child-parent differences point in the desired direction. It can downweight a dominant PCA pattern and strongly amplify a weaker one if that helps prediction. The PCA basis itself also depends on which concept embeddings are used to estimate it, and the paper does not test what happens to uu when only that concept set is perturbed.

    Even if I give the paper the most favorable case, that is, suppose the PCA space is stable, uu closely follows the natural embedding geometry, and the child-parent scores have a healthy range rather than tiny margins, the main claim still does not follow. Consider Vehicle with two children, Bus and Car. A scalar can give h(Vehicle)=0h(\text{Vehicle})=0, h(Bus)=1h(\text{Bus})=1, and h(Car)=2h(\text{Car})=2. Both true child-parent relations are perfectly oriented, but the scalar has also invented an order between Bus and Car although they are incomparable. Therefore, even a perfect pairwise subsumption prediction would show that hierarchy information can be read from the embeddings. It would not show that the ontology, or its lattice structure, is actually embedded in the LLM.

  2. Before testing whether the LLM contains a lattice (which itself is restricted to only two LLM families of not much consequential difference with unclear justification of selection), the paper first replaces the SNOMED ontology (best represented via a Description Logics fragment) by a much simpler FCA object. The paper does not establish that this replacement preserves the structure being studied. For example, suppose a parent has causative agent = dust and its child has causative agent = inorganic dust. SNOMED knows that inorganic dust is a specialization of dust. In the flat FCA representation, these initially become two different (relation type, destination) attributes. The paper also ignores role-group identity. The paper's own results show how much this representation choice matters. With the flat attributes, 65.3% of true hierarchy edges appear incomparable. After the hierarchy among destination values is added back, the incomparable fraction falls to 0% and oracle F1 rises to 0.996. So a large part of the apparent lattice failure can be created or removed simply by changing how the Description-Logic ontology is converted into the FCA table. This makes the starting point of the structural claim unclear. Before asking whether the LLM represents the SNOMED lattice, the paper needs to show that the FCA object being tested is a faithful representation of the relevant SNOMED semantics.

Questions for authors

Clarification requests

  1. How stable is the learned hierarchy direction uu to the particular concepts used to fit PCA? The paper reports stability across folds, but those folds change both the PCA population and the hierarchy-supervision data. Could you keep the hierarchy-training edges fixed, perturb only the concept embeddings used to fit PCA, and report how much the resulting PCA subspace and uu change? A stable uu under this test would make the interpretation of uu as a property of the frozen representation.

  2. How much does the learned uu rely on the dominant PCA directions vs. weak PCA directions? Please report how the weight of uu is distributed over PCs together with their eigenvalues.

  3. After destination-value closure, how much of the original SNOMED semantics is still lost by the FCA construction? In particular, role-group identity is ignored. Can you quantify whether preserving the omitted SNOMED structure changes the containment/comparability results? This would affect how much I can interpret the FCA results as statements about the SNOMED ontology itself.

Suggestions for future work

  • Testing more genuinely different model families and scales would help establish how general the observed phenomenon is.
  • A full order-recovery evaluation, including incomparability, branching, multiple inheritance, and meet/join structure, would be useful if the goal is to argue for an ontology or lattice embedding rather than pairwise subsumption decodability.
11 comments

Thank you for the detailed and technically precise review. Your W2 prompted an additional experiment whose outcome changes part of our paper, and we report it in full below together with the three checks you requested. In summary:

  • The grade result stands, and is independent of the attribute construction. Fitting one linear direction per split using only embeddings and is-a supervision yields 90.3% mean orientation accuracy across held-out branches.
  • Our criticism of the lattice hypothesis was too broad, and we narrow it. With value-closed attribute labels, the embedding-based inclusion score orients 77.9% of direct edges whose child was held out from attribute fitting, up from 55.8%. As you suspected, the result depends on how the attributes are constructed.
  • We do not claim an embedding of the ontology or its lattice. The claim is approximate order preservation along true edges.

What our results establish

Our results establish a supervised linear readout of hierarchical direction from frozen concept representations. They do not establish an embedding of the full SNOMED CT ontology or its lattice, and we agree that even perfect pairwise orientation would not establish recovery of the full partial order. The submitted wording blurred this distinction. We will revise the abstract, introduction, and Section 2.4 to state the empirical claim as approximate order preservation along true edges and to distinguish it from an order embedding.

W2, Q3: The attribute construction

You were right that the construction matters, and it matters on the embedding side too. The submission showed this for the full-attribute oracle: destination-value closure reduced the proportion of direct edges with incomparable attribute profiles from 65.3% to 0% and raised oracle macro-F1 to 0.996. We have now run the corresponding test on the embeddings. We used the same embeddings, training concepts, and released-code Incl rule in both conditions. We used the same 104 high-support attribute columns, excluding one of the original 105 because all training concepts became positive for it after closure. We fitted attribute directions separately with exact-value and value-closed labels; the latter also mark in-vocabulary ancestors of a recorded value as positive within the same relation type. The 7,153 matched direct edges are the subset of the 7,352 internal direct edges for which both sibling and unrelated candidates were available.

Embedding-based inclusion scoreExact-value labelsValue-closed labels
Orientation accuracy, edges whose child was held out from attribute fitting (747)55.76%77.91%
Orientation accuracy, all matched edges (7,153; descriptive)59.46%82.04%
Forward-versus-reverse AUROC, all matched edges (descriptive)0.5560.808

The weak orientation reported in the submission depended on how the attribute labels were constructed: incorporating the hierarchy of destination values improved orientation while keeping the attribute columns and scoring rule fixed. It should not be treated as a general limitation of LRH inclusion. The improvement was not limited to orienting known edges: on the same 747 held-out-child cases, value closure increased the rate at which Incl ranked the true parent above a matched sibling from 62.58% to 79.52%, counting ties as half.

What we will change. The revision will report both constructions side by side on the oracle and embedding sides, and will withdraw the general statement that inclusion scores cannot orient subsumption. The first half of the paper will be rewritten as a diagnosis, which is the reading your Strength 2 already gives it: exact matching omits coverage implied by the destination hierarchy. Restoring that coverage reduced incomparability in the oracle (65.3% → 0%) and improved orientation in the embeddings (55.8% → 77.9%).

What does not change. The grade does not use the attribute table in any form. It is fitted from embeddings and is-a edges alone, so none of the results in the next two sections depend on how attributes are constructed. We refit the grade on the same concept split, using only direct is-a edges whose endpoints were both in the training set, and evaluated all three readouts on the same 747 direct edges with a held-out child.

ReadoutWhat it requiresPaired orientation accuracy (747 edges)
Inclusion score, exact-value labels104 attribute directions55.76%
Inclusion score, value-closed labels104 attribute directions and the value hierarchy77.91%
Gradeone direction learned from embeddings with is-a supervision94.38%

Value closure substantially improved Incl, yet the single linear grade was 16.47 percentage points more accurate at orienting these known parent–child pairs, without requiring an attribute table.

Two readouts of hierarchical direction. We read the two results as convergent. In the full-attribute oracle, value closure made the child's attribute set a superset of its parent's on every tested direct edge. The grade recovers an ordering of the same kind directly from the embeddings, without being given attributes or their hierarchy. On this reading the lattice and the grade are not rival accounts. The contribution of the grade is that hierarchical direction is available as a single linear coordinate, and transfers across branches, disease domains and ontologies.

The algebraic decomposition is unchanged by value closure wherever the attribute gate is nonempty. What changes empirically is how well the resulting inclusion scores distinguish the two orientations of a known edge.

Why a binary context. LRH is stated over a binary object–attribute formal context. To evaluate it with human-curated relations, we represented each recorded non-is-a relation type and destination as a binary attribute. Treating these pairs as atomic and flattening role groups were choices in our construction, not requirements of LRH, and a flat binary table is not a lossless rendering of SNOMED CT's description-logic semantics. We will say so explicitly.

Q3: Role groups after value closure. To preserve role grouping, we required each parent role group to be covered by a single child role group, rather than allowing its relations to be matched across groups. The cohort is the respiratory hierarchy: 7,352 direct edges and 41,532 sibling pairs that are incomparable in the reference graph.

Attribute containment after value closureFlattened role groupsPreserved role groups
Direct edges: forward containment (including bidirectional)7,352 / 7,3527,352 / 7,352
Direct edges: bidirectional containment596589
Sibling pairs: containment in at least one direction (out of 41,532)5,784 (13.93%)5,732 (13.80%)

Preserving role groups retained forward containment on all 7,352 direct edges and removed reverse containment on seven. It removed attribute comparability for 52 of the 5,784 sibling pairs deemed comparable after value closure (0.90%). Even with groups preserved, 13.80% of sibling pairs that are incomparable in the reference graph still satisfy attribute containment in at least one direction.

W1 (a): The Vehicle / Bus / Car example

We agree with the example: Bus and Car can receive different grades without either subsuming the other. An ideal monotone grade satisfies h(x)>h(y)h(x)>h(y) when x⊏yx\sqsubset y, but the converse does not hold. The learned grade approximates this ordering on true edges; a grade difference alone does not establish comparability.

Siblings are included as hard negatives in our relation-prediction evaluations, which is the Bus/Car case. The table reports mean AUROC across five held-out SNOMED CT disease domains. Both rows use the same positives: direct edges and sampled two-hop ancestor pairs.

Test negativesCosineGrade differenceCosine + grade
All negative types0.6300.8510.905
Siblings only0.4500.8580.844

Cosine alone does not reliably rank the positives above siblings in this comparison. Grade differences tend to be larger for the positives, but this statistical separation does not decide comparability for an individual pair. Under strict branch holdout, the Llama-3.1-8B combined scorer also reaches 0.843 AUROC against sibling negatives.

W1 (b), Q1, Q2: Supervised readout and PCA controls

We estimate uu with a supervised linear probe trained on known child–parent edges, and we will say so wherever it is introduced. Three controls bound what the supervision contributes. With shuffled training labels, the same pipeline falls to chance. The learned direction reaches 90.3% mean orientation accuracy under complete-branch holdout, and a direction fitted on either ontology retains an orientation signal on the other without refitting. And a direction with no reweighting at all, the mean child-minus-parent vector in the same PCA-64 subspace, already reaches 83.79% orientation accuracy and 0.898 AUROC on held-out branches, against 90.34% and 0.964 for logistic regression (cosine between the two directions: 0.661). The signal is present in the mean displacement; logistic regression refines it.

Q1: Perturbing only the PCA fit set. We refitted PCA on different concept sets, keeping the supervised training edges, test edges, regularization and scaling fixed, and then refitted the probe in each perturbed subspace. The first row is the full-fit baseline; the 80% and 50% rows use five draws of training-edge endpoints per held-out branch. The last row uses three draws per branch from a pooled set of four other disease domains after excluding all respiratory concepts. Values are means over branches and draws, compared with the full-fit baseline of the same branch.

PCA fit setPCA-64 subspace overlapCosine of uu to full-fit uu (original space)Orientation accuracyOrientation AUROC
100% of training endpoints1.0001.00090.34%0.9640
80%0.9730.985 (range 0.972–0.993)90.35%0.9633
50%0.9210.960 (range 0.931–0.977)90.35%0.9632
Out-of-domain concepts0.4790.68190.50%0.9645

Subspace overlap is the mean squared cosine of the principal angles. Under subsampling, the subspace and the direction are both stable. Under a change of population, the retained subspace changes substantially and the fitted vector moves with it, yet mean held-out performance remains close to the baseline after refitting. The orientation signal is therefore robust to the tested changes in the PCA fit set, while the particular vector uu is specific to the subspace it is fitted in. We will describe uu as one representative readout and not as a unique axis.

Reviewer_2t4sOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors submitted a paper titled “Beyond the Lattice: A Graded Representation Hypothesis for LLM Concept Hierarchies”

The submission includes a main PDF and a Supplementary Material zip file (the zip was not downloadable from the form, but it looks like it is not a problem caused by the authors)

The paper is organized in a standard structure. The AI USE STATEMENT and REPRODUCIBILITY statements are both included. A long appendix is also included.

The paper addresses a problem related to the Lattice Representation Hypothesis and proposes a methodology called the Graded Representation Hypothesis.

The authors experimented on SNOMED CT and WordNet, showing that GRH can correctly orient 94% of is-a hedges.

Critical strengths

The paper addresses an interesting topic for the community. The approach is well formalized.

Critical weaknesses

I can't see any major Weaknesses apart from the experimental part. The experimental setting includes two datasets: one domain-specific SNOMED CT and the other a lexical database of English. I believe that more insights can be found by:

  1. including more datasets (small-scale, large-scale, domain-specific, and of common knowledge).
  2. datasets and measures can be adopted from the different existing SemEval tasks, in particular, measures capable of detecting the ability to reconstruct the hierarchy globally
  3. An error analysis can both help find the cause of the wrong orientation of edges and let the reader understand the overall behavior of the proposed approach.
Questions for authors

I don't have questions, but suggestions about minor aspects.

  1. The following two papers can be referenced with the proper published citation: 1.a - The paper " MedConcept: Unsupervised Concept Discovery for Interpretability in Medical VLMs." can be cited with the following reference: Md Rakibul Haque, K. M. Arefeen Sultan, Tushar Kataria e Shireen Y. Elhabian. MedConcept: Unsupervised Concept Discovery for Interpretability in Medical VLMs. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026 (Atti della conferenza internazionale, Strasburgo, Francia). 1.b - The paper "Emergent Linear Representations in World Models of Self-Supervised Sequence Models", can be cited with this reference: @inproceedings{nanda-etal-2023-emergent, title = "Emergent Linear Representations in World Models of Self-Supervised Sequence Models", author = "Nanda, Neel and Lee, Andrew and Wattenberg, Martin", editor = "Belinkov, Yonatan and Hao, Sophie and Jumelet, Jaap and Kim, Najoung and McCarthy, Arya and Mohebbi, Hosein", booktitle = "Proceedings of the 6th {B}lackbox{NLP} Workshop: Analyzing and Interpreting Neural Networks for {NLP}", month = dec, year = "2023", address = "Singapore", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2023.blackboxnlp-1.2/", doi = "10.18653/v1/2023.blackboxnlp-1.2", pages = "16--30", abstract = "How do sequence models represent their decision-making process? Prior work suggests that Othello-playing neural network learned nonlinear models of the board state (Li et al., 2023a). In this work, we provide evidence of a closely related linear representation of the board. In particular, we show that probing for my colour'' vs. opponent{'}s colour'' may be a simple yet powerful way to interpret the model{'}s internal state. This precise understanding of the internal representations allows us to control the model{'}s behaviour with simple vector arithmetic. Linear representations enable significant interpretability progress, which we demonstrate with further exploration of how the world model is computed." }
4 comments

Thank you for the positive assessment and for three concrete suggestions: broader data, a graph-level SemEval evaluation, and an error analysis. We carried out all three. In short: the grade direction generalizes across five SNOMED CT disease domains (90.8% orientation accuracy on fully held-out domains), correlates with human ratings it was never trained on, and quadruples the number of reference edges recovered in SemEval taxonomy reconstruction relative to cosine alone. The error analysis shows that errors concentrate where the child's name is shorter than the parent's.

W1: More datasets

We broadened the evaluation beyond the respiratory SNOMED CT and WordNet noun hierarchies studied in the submission. The table summarizes the evaluation settings across the submission and this response.

Evaluation settingSizeRole
SNOMED CT, expanded to five disease domains34,320 distinct concepts; 60,762 distinct within-domain direct is-a edgesLeave-one-domain-out generalization
WordNet noun hierarchies (original submission)Four held-out subtreesBranch holdout and cross-ontology transfer
SemEval-2016 Task 13, Food (WordNet-derived) [1]1,486 terms; 1,533 reference edgesZero-shot taxonomy reconstruction (W2)
Human specificity ratings [2]5,276 rated English nounsComparison with human judgments

SNOMED CT beyond respiratory disorders. We held out one disease domain at a time and excluded all of its concepts from training. Both evaluations used the same concept split and grade direction in each fold. Orientation tests whether the child scores higher than the parent on a known direct edge. Two-hop relation prediction distinguishes direct parents and sampled two-hop ancestors from reversed pairs, siblings, unrelated concepts, and depth-matched non-ancestors.

Held-out domainTest direct edgesOrientation accuracyOrientation AUROCTwo-hop AUROC: cosineTwo-hop AUROC: grade differenceTwo-hop AUROC: cosine + grade
Respiratory7,35290.67%0.9670.6410.8460.906
Cardiovascular13,63991.47%0.9720.6190.8580.909
Neurological15,12090.98%0.9710.6260.8570.909
Genitourinary10,35691.44%0.9720.6400.8590.912
Skin15,28289.20%0.9500.6240.8380.889
Equal-weight mean—90.75%0.9670.6300.8510.905

Results are stable across domains: a direction fitted on four domains orients about 91% of edges in the fifth, and adding the grade raises two-hop relation AUROC from 0.63 to 0.91. The five test sets contain 61,749 direct-edge records in total, corresponding to 60,762 distinct edges because some edges belong to more than one domain.

Human specificity ratings. Our primary analysis applies a direction fitted only on direct edges from respiratory SNOMED CT to the 1,965 nouns with a unique WordNet noun synset match. We also report results for all 5,276 rated nouns, including those in the primary analysis. The WordNet direction was fitted on the full WordNet noun graph and is included as an exploratory reference: some evaluated nouns may map to synsets used in fitting it. Neither direction was fitted to the human ratings [2].

Evaluation setNounsSpearman ρ, SNOMED CT directionSpearman ρ, WordNet direction (exploratory)
Nouns with a unique WordNet noun synset1,9650.3080.364
All rated nouns5,2760.4030.484

This tests the grade against human judgments rather than another ontology's edges.

W2: SemEval and global hierarchy reconstruction

We evaluated full taxonomy reconstruction on the English Food WordNet panel of SemEval-2016 Task 13 [1]. The shared-task guidelines exclude WordNet hypernym relations as a system resource. We therefore used a two-factor scorer fitted on the five SNOMED CT disease domains, with only direct-parent pairs labeled positive among the 322,797 training pairs. No Food reference edge was used for fitting. This is a zero-shot transfer from a medical ontology, intended as a controlled comparison of scorers and not as a competitive TExEval system: the scorer uses no task-specific corpus, lexical patterns, or in-domain supervision.

Every scorer is passed through the same decoder: each non-root term selects its highest-scoring parent, the lowest-scoring edge of each cycle is removed, and unattached components are linked to the root food. All scorers therefore output exactly 1,485 edges. We report direct-edge F1 and, as the global measure you asked for, F1 over the ancestor closure of the predicted graph against that of the reference graph, excluding pairs whose ancestor is the root. "Cosine + character length" replaces the grade with name length; it tests whether the gain could come from the surface cue identified in the error analysis below.

ScorerReference edges recovered (of 1,533)Direct-edge F1Non-root ancestor pairs recovered (of 5,470)Ancestor-closure F1
Cosine190.013320.008
Cosine + character length490.033700.018
Cosine + grade780.0521590.042

With the decoder and edge budget fixed, adding the grade recovers four times as many reference edges and five times as many ancestor relations as cosine alone, and clearly more than adding name length. The improvement extends to ancestor relations after excluding links to the shared root, rather than being limited to direct edges.

The limits are equally clear. Absolute F1 is low, and the single-parent decoder cannot recover the multiple inheritance in the reference graph. These results show what the grade adds under a fixed decoding rule, not that it suffices for taxonomy reconstruction.

W3: Error analysis

Errors concentrate where the child's name is shorter than the parent's. We examined reversed edges, where the parent receives a higher grade than the child, across four held-out WordNet subtrees. Length is the average token count of the lemmas used to encode a concept. The right-hand column excludes an edge if any child lemma is a prefix phrase followed by a parent lemma.

Child's name relative to parent'sReversed / test edgesAfter excluding modifier + parent names
Shorter1,820 / 4,568 (39.8%)1,717 / 4,269 (40.2%)
Equal length613 / 3,109 (19.7%)533 / 2,715 (19.6%)
Longer907 / 15,527 (5.8%)608 / 9,592 (6.3%)

Is the grade then just name length? We fitted a linear regression on training concepts to predict the grade from average token count, average character length, and a word-frequency proxy. We then compared orientation AUROC on the same held-out edges for the original grade, the lexical prediction, and the residual (grade minus lexical prediction).

DatasetOriginal gradeLexical predictionResidual
SNOMED CT (respiratory)0.9640.8580.837
WordNet0.9290.8360.775

Subtracting the lexical prediction lowered AUROC in both datasets, but the residual still carried orientation signal. The three lexical features are informative, but their linear prediction does not account for all of the signal. In the separate SemEval comparison, replacing the grade with character length recovered 49 reference edges instead of 78.

A limitation of name-only input. WordNet distinguishes two senses whose sole lemma is plant. Their encoded inputs are identical, so they receive the same embedding and the same grade. Errors of this kind cannot be fixed by the probe and would require encoding glosses or context.

Q1: Citation updates

Thank you for the publication details. We will replace the arXiv entries for MedConcept and Emergent Linear Representations in World Models of Self-Supervised Sequence Models with the published MICCAI and BlackboxNLP references you provided.

[1] Bordea, G., Lefever, E., and Buitelaar, P. SemEval-2016 Task 13: Taxonomy Extraction Evaluation (TExEval-2). SemEval 2016.

[2] Muraki, E. J. and Pexman, P. M. (2026). Distinguishing abstraction from abstractness: Specificity norms for 8,500 English words. Behavior Research Methods, 58, Article 60. doi:10.3758/s13428-026-02949-7.

Thank you very much for your comment.

Thank you very much for your response and time.

Reviewer_uSdmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper tests the Lattice Representation Hypothesis on SNOMED CT, a medical ontology with human-curated attributes. It finds the inclusion score from LRH mostly measures relatedness, not actual subsumption — it can barely tell a parent from a child or a sibling. Then the authors put forward the Graded Representation Hypothesis: concept hierarchies in LLMs are made of two parts, a relatedness lattice plus one single linear “specificity direction” that gives each concept a scalar grade. They verify this with linear probing, show the grade generalizes across held-out subtrees and even transfers between SNOMED CT and WordNet.

Critical strengths
  1. The critique of the lattice hypothesis is well grounded. Unlike prior work that built attributes from another LLM, this paper uses human-curated attributes as ground truth, which makes the test much fairer. The decomposition showing why inclusion scores lose directional information is convincing, and corrects a common misunderstanding in this line of research.
  2. Experimental controls are more thorough than average probing papers. Using siblings as hard negatives instead of random unrelated pairs, plus complete-branch holdout and cross-ontology transfer, makes the conclusions way more credible than works that just test on random train/test splits.
Critical weaknesses
  1. There is basically zero practical application value. The entire paper is about a hypothesis and probing observations — there is no downstream task shown at all. You can’t use this grade signal to improve hypernym detection, ontology completion, RAG or anything real. For a main conference like ICLR, a pure mechanism hypothesis with no tangible use case feels too thin.
  2. The classic limitation of all probing work applies here: being able to linearly decode a grade from embeddings does not mean the LLM actually uses this direction for reasoning. The hypothesis describes a pattern you can read out, but does not prove it is part of the model’s internal computation. This limits how much we actually learn about how LLMs really represent hierarchy.
Questions for authors

Have you tested whether using this grade signal can improve performance on any real downstream task like hierarchical classification or hypernym ranking? A positive result would substantially change my assessment.

4 comments

Thank you for the constructive review, and in particular for naming a concrete test of practical utility. We ran it. In short: adding the grade to cosine similarity raises ancestor-ranking MAP from 0.114 to 0.387 on SNOMED CT and from 0.014 to 0.055 on WordNet, and the gain transfers zero-shot to a public SemEval taxonomy benchmark.

W1 / Q1: Does the grade signal improve a downstream task?

Hypernym ranking on held-out subtrees. We evaluated on five held-out SNOMED CT respiratory subtrees and four held-out WordNet noun subtrees. For each query concept, we ranked all other concepts in its held-out subtree. In the primary setting we exclude the subtree root, so that always retrieving the shared root earns no credit. This leaves 126–2,053 candidates per query in SNOMED CT and 3,917–10,503 in WordNet. We compared cosine similarity, grade alone (the query's grade minus the candidate's grade), and the two-factor scorer. The grade direction and the scorer were fitted outside the held-out subtree. The scorer was trained with direct parents and two-hop ancestors as positives, but the ranking evaluation credited all reachable ancestors in the candidate pool. The query's known ancestors were used only to judge relevance, not to select candidates.

The table below reports ancestor ranking. Metrics are averaged over queries within each subtree, and subtrees are then weighted equally. Both root settings use the same queries: those with at least one relevant ancestor remaining after root exclusion.

DatasetScorerMAP, w/o rootMRR, w/o rootRecall@10, w/o rootMAP, w. root
SNOMED CTCosine0.1140.2980.1880.099
SNOMED CTGrade only0.2500.4760.3530.348
SNOMED CTCosine + grade0.3870.7090.5080.441
WordNetCosine0.0140.0490.0240.013
WordNetGrade only0.0400.0960.0360.057
WordNetCosine + grade0.0550.1260.0980.102

Cosine + grade improved ancestor MAP over cosine in all nine subtrees, and over grade alone in eight of nine. On SNOMED CT, Recall@10 rose from 18.8% to 50.8%. Absolute values on WordNet are low because each query is ranked against its entire subtree (expected MAP under uniform random ranking of the same candidates: 0.0027); the relative gain over cosine is about fourfold.

Zero-shot transfer to SemEval-2016 Task 13 (Food) [1]. To test the signal on a public benchmark from an unrelated domain, we fitted both two-factor scorers on the expanded five-domain SNOMED CT dataset, using the same fixed grade direction. The scorers differ in whether only direct parents or all sampled forward ancestors are labeled positive during training. No Food reference edge was used for fitting. For each query, we ranked the other terms in the 1,486-term English Food WordNet vocabulary, excluding the shared root, food. The table reports results for the 1,470 queries with a relevant answer remaining after root exclusion.

ScorerDirect-parent MAPDirect-parent Recall@10Ancestor MAPAncestor Recall@10
Cosine0.0400.0780.0220.032
Cosine + grade (direct-parent supervision)0.0930.1750.0740.127
Cosine + grade (all-ancestor supervision)0.0590.1430.0660.145

With direct-parent supervision, the scorer trained on SNOMED CT more than doubles direct-parent MAP and more than triples ancestor MAP on Food terms.

Ancestor ranking versus direct-parent retrieval. On the nine held-out subtrees, the two-hop scorer improved ancestor ranking but did not beat cosine for direct-parent retrieval on average in either ontology. Keeping the grade direction, training pairs, and test queries fixed, we refitted the scorer with only direct parents labeled positive. No-root direct-parent MAP rose from 0.1630 to 0.2593 on SNOMED CT, above cosine (0.2144), and from 0.0196 to 0.0253 on WordNet, still below cosine (0.0396). Although the Food benchmark is derived from WordNet, it uses a scorer trained on SNOMED CT and a different candidate pool from the WordNet subtree evaluation. In the matched subtree comparison, direct-parent supervision improved retrieval over two-hop supervision, but did not consistently outperform cosine.

Full taxonomy reconstruction on Food. We transferred the SNOMED CT scorer trained with direct-parent supervision and applied the same decoder to every scorer. Each non-root term selected its highest-scoring parent; we removed the lowest-scoring edge in each cycle and linked unattached components to the root, yielding 1,485 edges per scorer.

Cosine + grade recovered 78 of 1,533 reference edges, versus 19 for cosine alone and 49 for cosine + character length. After excluding pairs with the root as ancestor, it recovered 159 of 5,470 reference ancestor pairs, versus 32 and 70, respectively. Absolute F1 remained low (0.052 for direct edges and 0.042 for non-root ancestor pairs), and the single-parent decoder cannot recover multiple inheritance.

W2: Decodability does not imply that the model uses the direction

We agree, and we do not claim causal use. An observational probe shows that information is available in a representation, not that the model relies on it. Our claim concerns representation geometry: an approximate hierarchical ordering can be read out linearly from frozen, pooled final-layer representations of concept names.

We estimate the direction with a supervised linear probe, but the controls below show that the readout is not merely a product of probe capacity or memorization of the training edges. With shuffled labels the same probe falls to chance. The direction orients 90.3% of edges when entire branches are held out. And it retains an orientation signal when transferred between SNOMED CT and WordNet without refitting on the target ontology. The new ranking experiments add that an external system can exploit the signal; they do not show that the LLM itself does.

In the revision, we will state this distinction where the hypothesis is first introduced, rather than only in the limitations, and we will replace wording that could be read as a claim about the model's internal computation. Layer-wise analysis and targeted interventions along the direction are the natural next test of causal use, and we leave them to future work.

Presentation

We will clarify two passages in the revision. In the presentation of the GRH, we will clarify that relatedness identifies plausible candidates but does not establish taxonomic comparability, while the grade indicates direction. In “Candidates and tasks,” we will separate candidate definitions from the positive and negative labels used in each task.

[1] Bordea, G., Lefever, E., and Buitelaar, P. SemEval-2016 Task 13: Taxonomy Extraction Evaluation (TExEval-2). SemEval 2016.

AuthorsOpenReview ↗

More positive results

Thank you again for your review. Since our first response, the discussion with the other reviewers has produced several further results. We summarize those most relevant to your two concerns, so that you do not need to read the other threads.

On practical utility (W1, Q1). In addition to the hypernym-ranking results in our first response (ancestor MAP 0.114 → 0.387 on SNOMED CT; direct-parent MAP 0.040 → 0.093 zero-shot on SemEval-2016 Food):

  • Comparison with an established supervised method. Under matched conditions (same frozen embeddings, same training edges, no refitting on the target), the single grade direction transfers from SNOMED CT to WordNet as well as BiRRE, a pairwise model with multiple projections and a nonlinear classifier: 76.16% vs. 75.62% ± 0.94 orientation accuracy on 75,831 WordNet edges.
  • Broader coverage. Across five SNOMED CT disease domains, each held out in turn, the direction orients 90.8% of edges, and adding the grade raises relation-prediction AUROC from 0.63 (cosine) to 0.91.

On what probing tells us (W2). We still do not claim that the model uses this direction. Two new results do, however, indicate that it reflects the representation and not the probe:

  • It needs almost no fitting. The unweighted mean child-minus-parent vector already orients 83.8% of edges in held-out branches (90.3% with logistic regression).
  • It tracks human judgments it was never fitted to. A direction fitted only on respiratory SNOMED CT edges correlates with human specificity ratings of general English nouns (Spearman ρ = 0.31–0.40).

One correction, for transparency. A check requested by Reviewer 3 showed that the weak orientation of the LRH inclusion score is specific to our flat attribute construction; with value-closed attributes it improves substantially. We will narrow that criticism in the revision. The grade results above do not use the attribute table and are unaffected. The same analysis showed that the directional information in the corrected lattice is itself carried by one scalar per concept, which is consistent with the grade.

You wrote that a positive downstream result would substantially change your assessment. We would be grateful to know whether these results address your question, and we are glad to clarify or add anything that would help in the time remaining.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading and for a substantive discussion. We are encouraged that the reviewers agree on the core of the work. Reviewer uSdm found the experimental controls "more thorough than average probing papers", noting the use of human-curated attributes, sibling hard negatives, complete-branch holdout and cross-ontology transfer. Reviewer 2t4s found the topic interesting for the community and the approach well formalized. Reviewer qpWq described the hierarchy-direction result as "a real empirical finding" observed under strong tests, and after the additional experiments acknowledged that the learned direction "is robust and generalizes well" and that the requested checks were satisfactorily answered.

The revised paper is now uploaded, with changes marked in blue. It addresses the following points raised in the reviews:

  • Practical utility (Reviewer uSdm): downstream hypernym-ranking experiments, including zero-shot transfer to a public SemEval benchmark.
  • Broader evaluation and error analysis (Reviewer 2t4s): additional SNOMED CT disease domains, a graph-level SemEval evaluation, agreement with human specificity ratings, and an analysis of orientation errors.
  • Scope of the hypothesis (Reviewers uSdm and qpWq): Hypothesis 1 is now stated as a readout claim. We no longer say that a hierarchy is "realized by" linear structures, and we do not claim that the model uses the direction.
  • Attribute construction and motivation (Reviewer qpWq): the paper now reports both the exact-value and the value-closed attribute encodings, narrows its criticism of lattice inclusion accordingly, and motivates the grade as a readout that does not depend on how an attribute table is built.
  • Supervised readout and PCA (Reviewer qpWq): PCA fit-set perturbation, the distribution of the direction over principal components, and an unweighted mean-difference direction.
  • Related work and novelty (Reviewer qpWq): a revised positioning against supervised hypernymy methods, with a matched comparison.

The original submission argued that the lattice fails and a grade is needed instead. The revised paper makes a more careful and, we believe, more useful claim: the directional part of subsumption can be read from frozen LLM representations as a single concept-level coordinate, without an attribute table. This coordinate generalizes to unseen branches and disease domains, transfers across ontologies without refitting, and tracks human judgments of specificity it was never fitted to. None of the grade results depend on the attribute table, so they are unaffected by the correction to the lattice analysis. We are grateful to the reviewers for helping us sharpen the contribution, and we are glad to answer any remaining questions in the last period.

AuthorsOpenReview ↗

General Response to AC

Dear Area Chairs,

Thank you for handling our submission. As the discussion period closes soon, we would like to summarize where it stands, since two of the three threads did not reach a conclusion.

What we have done during the rebuttal. All experiments requested in the reviews have been completed and are reported in our responses. The revised paper has been uploaded with changes marked in blue, and a general response summarizes the changes.

Reviewer uSdm wrote that a positive downstream result "would substantially change my assessment". We reported such results: adding the grade raises ancestor-ranking MAP from 0.114 to 0.387 on SNOMED CT, and the gain transfers zero-shot to a public SemEval benchmark. We have not received a reply to our response or to our follow-up.

Reviewer 2t4s asked for broader data, a graph-level evaluation and an error analysis. All three were provided, and the reviewer has replied to our response.

Reviewer qpWq engaged in detail, and the paper is better for it. The reviewer has acknowledged that the requested checks (Q1, Q2) are "satisfactorily answered" and that the learned direction "is robust and generalizes well". The remaining concerns were that Hypothesis 1 and the motivation needed to change, and that novelty should be assessed against supervised hypernymy methods. We agreed with the first and revised both: the hypothesis is now a readout claim, and the criticism of lattice inclusion is narrowed to the attribute encoding under which it was observed. We addressed the second with a matched comparison against BiRRE, in which the single grade direction transfers across ontologies as well as the pairwise model (76.16% vs. 75.62% ± 0.94). We have not received a reply to that response or to the revision.

The revision makes a narrower claim than the submission, and we have been explicit about what was corrected. The main result is unchanged and does not depend on the corrected part: a single linear direction in frozen LLM representations orders concepts in unseen branches and disease domains, transfers across ontologies without refitting, and tracks human specificity judgments it was never fitted to. None of the reviewers disputes this result after the rebuttal.

We would be grateful if, during the final discussion, you could invite the reviewers, especially (Reviewer uSdm and Reviewer qpWq), to take another look at the responses and the revised paper, so that they can be considered together in the decision. Thank you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.