Relational Schemata in BERT Are Inducible, Not Emergent: A Study of Performance vs. Competence in Language Models

Cole Gawin

REVIEW 4 major objections 5 minor 2 references

Relational Schemata in BERT Are Inducible, Not Emergent: A Study of Performance vs. Competence in Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Pretrained BERT can classify semantic relations accurately while its [CLS] embedding geometry shows almost no relation structure; structured relational schemata appear only after fine-tuning.

desk verdict A clean, interesting dissociation between classifier accuracy and RSA structure in BERT, but the conclusion rests on a coarse three-category ground truth and the paper lacks the reproducibility details to carry it. read the letter →

arxiv 2506.11485 v1 pith:QD6LPVDE submitted 2025-06-13 cs.CL cs.AI

Cole Gawin This is my paper · ORCID

classification cs.CLcs.AI

keywords relationalschemataBERTrepresentationalsimilarityanalysisperformancevscompetencerelationclassificationfine-tuningemergentinduciblestructureembeddinggeometry

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether BERT has internalized abstract relation types—taxonomic (is-a), mereological (part-whole), and functional (used-for)—as structured schemas, or merely uses surface cues to sort concept pairs into those classes. It reports a dissociation: a probe on frozen pretrained embeddings reaches 79% accuracy on relation classification, while the geometry of those embeddings correlates only $\rho \approx 0.04$ with the true relation structure. After fine-tuning on the same task, accuracy reaches 90% and the correlation jumps to $\rho \approx 0.77$, with concept pairs of the same relation type clustering in the embedding space. The paper concludes that relational schemata are inducible, not emergent: pretraining leaves latent, decodable relational signals, but structured relational abstraction appears only with task scaffolding. The upshot is that behavioral performance on semantic tasks should not be equated with conceptual competence in language models.

What carries the argument

The central object is the representational dissimilarity matrix (RDM) computed over [CLS] token embeddings of concept pairs. Each entry is one minus the cosine similarity between two concept-pair embeddings; this is compared against a ground-truth RDM that assigns 0 to pairs sharing a relation type and 1 otherwise. The Spearman rank correlation between the two upper triangles quantifies whether the embedding geometry reflects the relation categories. Layer-wise RSA over BERT's 13 layers and UMAP projections provide the supporting visualization of where and how structure appears.

What would settle it

Recompute the representational dissimilarity matrices from concatenated or difference embeddings of the two concepts, or from a linear projection of the hidden states, instead of the [CLS] token alone. If any of these yield a high Spearman correlation with the ground-truth relation matrix in the pretrained model, the claim that pretrained BERT lacks relational schemata is falsified.

Watch

Extended reading notes

Core claim

The central claim is that relational schemata are not an emergent property of BERT's pretraining; they are inducible through supervised fine-tuning. The evidence is a dissociation: a logistic-regression probe on frozen [CLS] embeddings reaches 79% accuracy on a three-way relation classification task, yet the Spearman correlation between the model's representational dissimilarity matrix and a ground-truth relation matrix is only $\rho \approx 0.04$. After fine-tuning BERT on the task, accuracy rises to 90% and the correlation rises to $\rho \approx 0.77$, with UMAP projections showing same-relation pairs clustering and layer-wise RSA showing structure concentrated in middle-to-deep layers. The paper interprets this as behavior without competence: the pretrained model can decode relations without organizing them into abstract, generalizable schemas.

Load-bearing premise

The load-bearing premise is that relational schemata would be visible as global similarity structure in the [CLS] embedding of a concept pair, so if the structure lived in a lower-dimensional subspace, in the interaction between the two concept tokens, or in a nonlinear code, the low RSA correlation would not rule it out.

Editorial extensions

If this is right

High classification accuracy on a relation task does not by itself show that the model represents relation types as abstract categories.
Fine-tuning reorganizes BERT's [CLS] embedding space so that pairs sharing a relation type become geometrically close, with the strongest effect in middle-to-deep layers.
Pretrained BERT's relational knowledge is real but unstructured, so probes can decode it even though the model has no schema-level organization.
Models intended for relational or analogical reasoning will likely need explicit structural supervision, not just more pretraining data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the dissociation holds for other architectures, representational similarity analysis before and after fine-tuning could serve as a more stringent competence test than linear-probe accuracy alone.
Applying the same before/after RSA design to causal or temporal relations, or to decoder-only models, would test whether inducibility is specific to BERT's masked-objective pretraining.
The split of the functional category into two clusters after fine-tuning hints that the coarse three-way taxonomy may hide finer relational structure; labeling functional subclasses could reveal distinct schemata.
A direct extension would be to fine-tune on one relation taxonomy and then measure RSA on a held-out taxonomy: if schemata transfer, they are genuinely abstract; if not, fine-tuning may simply fit task-specific geometry.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Cole Gawin's preprint makes a claim that is worth taking seriously: pretrained BERT can classify semantic relations at 79% accuracy while its [CLS] embedding geometry shows almost no correlation (Spearman rho=0.04) with the three-category relation structure, and only after fine-tuning does the RSA alignment jump to rho=0.77. The performance/competence framing is apt, and the control experiment with permuted labels is a nice touch that rules out some trivial artifacts. The layer-wise RSA is also a sensible way to localize the effect of fine-tuning. If the result holds up, it's a concrete demonstration that benchmark success can coexist with unstructured internal organization, which matters for interpretability and for how we credit models with understanding.

The soft spots are real, though. The ground-truth RDM is built from only three coarse categories (taxonomic, mereological, functional), collapsing nine ConceptNet relations into a binary same/different matrix. A near-zero correlation with that matrix only shows that pairwise cosine similarities in [CLS] don't reproduce the paper's three-way taxonomy; it doesn't rule out structure at the level of the nine fine-grained relations. The same embeddings already support 79% accuracy under a linear probe, so category-discriminative geometry exists somewhere in [CLS] space. The paper interprets that as 'distributed associative cues' rather than structure, but the operationalization doesn't actually distinguish those alternatives. The fine-tuning half is also partly circular: supervised training on relation labels will naturally align representations with those labels.

The paper also lacks the usual empirical hygiene: no seeds, no error bars, no significance tests, and the dataset construction ("selected a balanced subset") is underspecified. Fine-tuning hyperparameters are absent. For a single dissociative result that rests on one correlation difference, that's a lot of missing support.

Who is this for? People working on interpretability and on the competence/performance distinction in LLMs will get a clear, provocative case study, but as a proof it's not there yet. The central claim is plausible, but currently it's a demonstration of a possible dissociation, not an established one.

Recommendation: send it to peer review, because the question is important and the core experiment is the right kind of evidence. The reviewers should push for a more granular ground truth, tests of alternative representational geometries, and full reproducibility. With those additions, the paper could be genuinely valuable.

Referee Report

4 major / 5 minor

Summary. The paper asks whether BERT's high accuracy on semantic relation classification reflects structured relational schemata or merely surface-level, decodable signals. Using ConceptNet-derived concept pairs collapsed into three coarse relation categories (taxonomic, mereological, functional), the author compares (i) classification accuracy via logistic regression on frozen [CLS] embeddings versus fine-tuned BERT, and (ii) representational structure via RSA between embedding RDMs and a binary ground-truth RDM. The central reported finding is a dissociation: pretrained BERT achieves 79% accuracy but an RSA of only 0.04, whereas fine-tuned BERT reaches 90% accuracy and an RSA of 0.77. The paper concludes that relational schemata are not emergent from pretraining but can be induced through task-specific fine-tuning, and discusses implications for the performance/competence distinction.

Significance. If the central dissociation is valid, the paper makes a useful contribution to the growing literature on whether LLM performance implies structured conceptual knowledge. The use of RSA and a label-permutation control are appropriate and potentially transferable. The paper also usefully separates behavior from representation and shows a layer-wise emergence of relational structure after fine-tuning. However, the main claim rests on a specific and under-validated operationalization of 'competence' as coarse-category cosine clustering in the [CLS] token, and the methodological reporting is too sparse to assess robustness. The question is important enough to merit revision, but the current evidence does not yet rule out the possibility that pretrained BERT encodes relational structure in a form not captured by the chosen analysis.

major comments (4)

[Section 4.4.2, Table 2] The ground-truth RDM is binary and derived from collapsing nine ConceptNet relation labels into three coarse categories (Section 4.2). A near-zero Spearman correlation between the [CLS] embedding RDM and this coarse binary matrix shows only that the embeddings do not reproduce the paper's three-way taxonomy; it does not rule out structure at the level of the nine fine-grained relations. Indeed, the same embeddings support 79% accuracy under a linear probe (Table 3), so category-relevant geometry is present. The conclusion that pretrained BERT lacks relational schemata is therefore load-bearing on an unvalidated assumption. I ask the authors to add analyses with a nine-level ground-truth RDM, within-category versus between-category distance comparisons, or a permutation baseline over the binary matrix, to test whether the low pretrained RSA is an artifact of the coarse target.
[Sections 4.2, 4.3, 5.1] The experimental reporting is underspecified in ways that affect the reproducibility of the central comparison. The dataset section says 'selected a balanced subset of 400 samples' but does not describe the sampling procedure from ConceptNet, the exact split construction, or how concept pairs were chosen to avoid leakage. Fine-tuning hyperparameters (learning rate, number of epochs, batch size, optimizer, random seed) are not reported. All accuracy and RSA values are point estimates without confidence intervals, multiple seeds, or significance tests. For a claim about a near-zero versus high correlation, the authors should report variability across seeds or bootstrap resampling and a statistical comparison (e.g., a permutation test) between pretrained and fine-tuned RSA distributions.
[Section 5.2, Figure 1] The layer-wise RSA is only reported for the fine-tuned model, with the text noting correlation rises sharply in deeper layers. Since a key part of the argument is that relational structure is absent in pretraining, the absence of layer-wise RSA for the pretrained model leaves open the possibility that the [CLS] pooling (or the final-layer representation) masks structure present in intermediate layers. Please report layer-wise RSA for the pretrained model as well, or explicitly justify why only the final [CLS] representation is the relevant measure of competence.
[Section 5.3, Figure 2] The UMAP visualizations are used as qualitative evidence that 'no discernible clustering' exists in the pretrained model. UMAP projections can compress or distort global structure, and the absence of visible clusters is not a quantitative measure of representational organization. The paper should supplement UMAP with a quantitative clustering or separability index (e.g., silhouette score, k-NN consistency, or cluster purity) computed on the original high-dimensional embeddings, with appropriate baselines, before drawing conclusions from the projection.

minor comments (5)

[Section 4.4.1] The paragraph beginning 'To measure performance, I assessed both the pretrained and fine-tuned BERT models' is duplicated verbatim within the same subsection.
[Section 5.1] There is a typo: 'These quanitative results' should read 'These quantitative results'.
[Section 4.2] The phrase 'selected a balanced subset of 400 samples' is ambiguous; given the subsequent counts (3600 total pairs, 1200 per relation type), it appears to mean 400 samples per relation label across nine labels. Please clarify the wording.
[Section 2.4] The citation 'Tenney et al., 2019' is described as showing that BERT's representations 'become more linguistically aligned after fine-tuning,' but the cited paper (BERT rediscovers the classical NLP pipeline) does not directly report fine-tuning-induced alignment in the way described; please verify the citation or rephrase.
[Section 5.2] Table 2's caption reads 'Spearman correlation (ρ) the model's embedding similarity matrix and conceptual ground truth'; the word 'between' is missing.

Circularity Check

1 steps flagged · score 6.0 of 10

The 'inducible' claim is a fitted consequence of training on the same labels used to define the RSA target; the pretrained 'not emergent' claim is independent.

fitted input called prediction [Sections 4.3, 4.4.2, and 5.2]
"Fine-tuning was conducted on the relation classification task, in which the model was trained to predict one of three high-level relation categories. ... optimizing the model’s parameters to minimize the cross-entropy loss between predicted and true labels. ... A corresponding ground-truth RDM was constructed by assigning a dissimilarity of 0 to concept pairs sharing the same relation type and 1 to all others. ..."

The ground-truth RSA matrix is built from the same three relation labels that serve as the fine-tuning targets. The model is trained with cross-entropy to map [CLS] embeddings to those labels, so the post-fine-tuning alignment between the embedding RDM and the label-derived RDM is a fitted consequence of the training objective, not an independent confirmation that relational schemata are 'induced.' The conclusion that relational schemata can be induced via task scaffolding thus reduces to: after supervised training on relation labels, the representations correlate with those relation labels.

full rationale

The pretrained half of the dissociation is self-contained: frozen [CLS] embeddings are compared to a label-derived target matrix via Spearman RSA with no fitting of the embeddings to those labels, and the 79% linear-probe accuracy shows the low RSA is non-vacuous. The circularity is confined to the fine-tuning half: because the model is trained to predict the same three relation categories that define the RSA ground truth, the post-fine-tuning ρ ≈ 0.77 is an expected fitted consequence rather than an independent observation. This yields partial circularity: the 'not emergent' claim has independent content, while the 'inducible' claim is close to definitional. No load-bearing self-citation or imported-uniqueness issue was found; the self-citation to Gawin et al. (2025) is motivational only.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its free parameters are the unspecified dataset sampling choices and fine-tuning hyperparameters, which are not fixed by any reported procedure. The key axiomatic load is carried by the assumption that RSA on [CLS] embeddings is the right measure of relational competence.

free parameters (2)

Sample selection subset = unspecified
The composition of the dataset directly determines both accuracy and RSA; the selection is described only as 'balanced', which is a free choice that could bias results (Section 4.2).
Fine-tuning hyperparameters = unspecified
Learning rate, number of epochs, batch size, and optimizer settings are not reported, and these could affect the degree of representational reorganization and hence the RSA scores (Section 4.3).

assumptions (4)

domain assumption RSA on [CLS] embeddings is a valid measure of relational competence
The paper defines competence as geometric organization by relation type, operationalized via Spearman correlation of [CLS] embedding dissimilarities with a binary label-derived matrix (Section 4.4.2). This equates one representational measure with conceptual understanding.
domain assumption The three relation categories are mutually exclusive and exhaustively labeled
Concept pairs are taken from ConceptNet labels that 'map cleanly' into the categories (Section 4.2), assuming the mapping is unambiguous and the categories do not overlap.
standard math Logistic regression on frozen embeddings captures latent relational signals
A linear probe is a standard method, but the paper interprets its accuracy as evidence that relational information is present, which is a reasonable assumption for a probe, though not the only one.
domain assumption ConceptNet is a reliable source of ground-truth relations
The dataset is built from ConceptNet without independent validation; label noise would affect both accuracy and RSA, and the paper relies on this external resource as ground truth.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Relational Schemata in BERT Are Inducible, Not Emergent: A Study of Performance vs. Competence in Language Models." pith.science (2026). https://pith.science/paper/QD6LPVDE

@misc{pith2026250611485,
  author       = {Pith},
  title        = {Pith review of: Relational Schemata in BERT Are Inducible, Not Emergent: A Study of Performance vs. Competence in Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QD6LPVDE}},
  note         = {Machine review of arXiv:2506.11485}
}

read the original abstract

While large language models like BERT demonstrate strong empirical performance on semantic tasks, whether this reflects true conceptual competence or surface-level statistical association remains unclear. I investigate whether BERT encodes abstract relational schemata by examining internal representations of concept pairs across taxonomic, mereological, and functional relations. I compare BERT's relational classification performance with representational structure in [CLS] token embeddings. Results reveal that pretrained BERT enables high classification accuracy, indicating latent relational signals. However, concept pairs organize by relation type in high-dimensional embedding space only after fine-tuning on supervised relation classification tasks. This indicates relational schemata are not emergent from pretraining alone but can be induced via task scaffolding. These findings demonstrate that behavioral performance does not necessarily imply structured conceptual understanding, though models can acquire inductive biases for grounded relational abstraction through appropriate training.

Figures

Figures reproduced from arXiv: 2506.11485 by the authors.

**Figure 1.** Spearman correlation (ρ) between model layer representations and relational structure as measured by RSA, for each layer of the fine-tuned BERT model. RSA was also performed layer-wise for each of the fine-tuned model’s 13 layers, including the embedding layer. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

**Figure 2.** UMAP projections of [CLS] token embeddings from pretrained (left) and fine-tuned (right) BERT models. In the pretrained model, no discernible clustering by relation category was observed. Concept pairs appeared interspersed, with substantial overlap across taxonomic, mereological, and functional classes. In the fine-tuned model, a significant degree of relational clustering emerged. Concept pairs of the same relatio… view at source ↗

**Figure 4.** Confusion matrices for relation classification using pretrained BERT with logistic regres [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 linked inside Pith

[1]

Artstein, R., & Poesio, M. (2008). Inter-coder agreement for computational linguistics.Computa- tional Linguistics, 34(4), 555–596. https://doi.org/10.1162/coli.07-034-r2 Bender, E. M., & Koller, A. (2020, July). Climbing towards NLU: On meaning, form, and under- standing in the age of data. In D. Jurafsky, J. Chai, N. Schluter, & J. Tetreault (Eds.), Pro...

arXiv 2008
[72]

https://doi.org/10.2307/1412159 Speer, R., Chin, J., & Havasi, C. (2017). Conceptnet 5.5: An open multilingual graph of general knowledge. Proceedings of the AAAI conference on artificial intelligence, 31(1). Tenney, I., Das, D., & Pavlick, E. (2019, July). BERT rediscovers the classical NLP pipeline. In A. Korhonen, D. Traum, & L. Màrquez (Eds.),Proceedi...

arXiv doi:10.48550/arxiv.2302.13971 2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Artstein, R., & Poesio, M. (2008). Inter-coder agreement for computational linguistics.Computa- tional Linguistics, 34(4), 555–596. https://doi.org/10.1162/coli.07-034-r2 Bender, E. M., & Koller, A. (2020, July). Climbing towards NLU: On meaning, form, and under- standing in the age of data. In D. Jurafsky, J. Chai, N. Schluter, & J. Tetreault (Eds.), Pro...

arXiv 2008

[2] [72]

https://doi.org/10.2307/1412159 Speer, R., Chin, J., & Havasi, C. (2017). Conceptnet 5.5: An open multilingual graph of general knowledge. Proceedings of the AAAI conference on artificial intelligence, 31(1). Tenney, I., Das, D., & Pavlick, E. (2019, July). BERT rediscovers the classical NLP pipeline. In A. Korhonen, D. Traum, & L. Màrquez (Eds.),Proceedi...

arXiv doi:10.48550/arxiv.2302.13971 2017