{"id":"c9aa50da-8523-42af-b5ad-918fa8e6038a","arxiv_id":"2506.11485","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Relational schemata in BERT are not emergent from pretraining: high classification accuracy coexists with unstructured embeddings, and only fine-tuning induces clustering by relation type.","lead":"This study asks whether BERT's success at classifying semantic relations reflects true conceptual structure or merely surface patterns. Using accuracy and representational similarity analysis, it finds that pretrained BERT can classify relations accurately but only organizes its internal representations by relation type after fine-tuning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pretrained RSA near zero may reflect the coarse three-category target matrix, not absence of relational structure.","rationale":"I read the paper as claiming a dissociation between decodable relation signals and structured relational representation: pretrained BERT classifies relations at 79% while its [CLS] geometry shows essentially no category-level structure, and only fine-tuning produces such structure. For that claim to stand, the RSA result must be a faithful indicator of whether relational schemata exist. The least secure link is not the RSA method per se but the target matrix it is tested against: nine ConceptNet labels are collapsed into three broad buckets, and the matrix treats every same-bucket pair as similar and every cross-bucket pair as dissimilar. Since relations like causes, used-for, and capable-of are grouped together as 'functional,' a model that distinguishes those fine-grained relations would score low against the 3-way matrix while still possessing relational structure. This concern is internal to the paper's own design rather than a disagreement with external consensus, and it is directly testable with the reported embeddings. The reader's weakest assumption about [CLS] and cosine geometry is related but not identical; I partially agree. If the 9-way check fails to reveal structure, a tempered version of the dissociation may survive, but the current categorical conclusion ('not emergent from pretraining') is stronger than the evidence supports. I would keep the reader's conditional verdict, with the added explicit condition that the author report the 9-way RSA or an equivalent probe-derived target matrix before the central claim is accepted.","tokens_in":10517,"tokens_out":11225,"duration_ms":109567,"concrete_test":"Recompute the RSA of Table 2 on the same pretrained [CLS] embeddings, keeping the same cosine dissimilarity and concept pairs, but replace the 3-way target RDM with a 9-way target RDM that assigns dissimilarity 0 only to pairs sharing the same original ConceptNet relation label (is-a, defined-as, distinct-from, part-of, has-a, made-of, used-for, causes, capable-of). If the Spearman rho rises substantially above 0.04 (e.g., above 0.2), or if within-subtype mean cosine similarity clearly exceeds between-subtype similarity inside a broad category, then pretrained BERT encodes fine-grained relational structure and the 'not emergent' claim fails. If the 9-way rho stays near zero, the coarse-target objection is weakened, but nonlinear or low-dimensional structure would still need separate tests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central dissociation in Table 2 is that pretrained BERT shows a Spearman correlation of about 0.04 between the [CLS] embedding RDM and a binary ground-truth RDM, which is read as the absence of relational schemata. The ground-truth RDM is constructed from only three coarse categories into which nine ConceptNet relation labels are collapsed (is-a/defined-as/distinct-from, part-of/has-a/made-of, used-for/causes/capable-of; Sections 4.2 and 4.4.2). A near-zero correlation with that matrix only shows that pairwise [CLS] cosine similarities do not reproduce the paper's three-way taxonomy. It does not rule out structure at the level of the nine fine-grained relations, and the same embeddings already support 79% accuracy under a linear probe (Section 5.1), so category-discriminative geometry exists in [CLS] space. The central claim therefore equates relational competence with coarse-category cosine clustering in the [CLS] token, an operationalization the paper motivates but does not validate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":10690,"tokens_out":3150,"duration_ms":34060,"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":[{"comment":"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.","section":"Section 4.4.2, Table 2"},{"comment":"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":"Sections 4.2, 4.3, 5.1"},{"comment":"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":"Section 5.2, Figure 1"},{"comment":"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.","section":"Section 5.3, Figure 2"}],"minor_comments":[{"comment":"The paragraph beginning 'To measure performance, I assessed both the pretrained and fine-tuned BERT models' is duplicated verbatim within the same subsection.","section":"Section 4.4.1"},{"comment":"There is a typo: 'These quanitative results' should read 'These quantitative results'.","section":"Section 5.1"},{"comment":"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":"Section 4.2"},{"comment":"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":"Section 2.4"},{"comment":"Table 2's caption reads 'Spearman correlation (ρ) the model's embedding similarity matrix and conceptual ground truth'; the word 'between' is missing.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely question, but its central claim currently hinges on a single, coarse RSA target and on qualitative visualizations. The missing methodological details and lack of error bars are more than presentation issues, because the 0.04 versus 0.77 contrast is the entire empirical basis for the 'not emergent' conclusion. I believe the authors can address this with additional analyses rather than a complete rework, hence major_revision rather than reject. I would also encourage the editor to check that the scope fits a computational linguistics venue given the cognitive-science framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nThe 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.\n\nWho 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.\n\nRecommendation: 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.","headline":"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.","tokens_in":11186,"tokens_out":2727,"would_cite":false,"duration_ms":26035,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["relational schemata","BERT","representational similarity analysis","performance vs competence","relation classification","fine-tuning","emergent vs inducible structure","embedding geometry"],"falsifier":"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.","tokens_in":10315,"feed_emoji":"🧠","tokens_out":9751,"duration_ms":85700,"temperature":0.7,"pith_summary":"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.","feed_headline":"Before fine-tuning, BERT labels relations; structure appears only after","feed_subtitle":"Accuracy stays high before fine-tuning; training lifts embedding-relation correlation from 0.04 to 0.77","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines BERT, the model whose pretrained and fine-tuned representations are compared.","marker":"Devlin et al., 2019"},{"why":"Supplies ConceptNet, the source of the 3,600 labeled concept pairs used in the classification and fine-tuning tasks.","marker":"Speer et al., 2017"},{"why":"Introduces representational similarity analysis, the method used to compare embedding geometry with the ground-truth relation structure.","marker":"Kriegeskorte et al., 2008"},{"why":"Inspires the layer-wise RSA analysis that tracks where relational structure emerges across BERT's 13 layers.","marker":"Gandikota et al. (2024)"},{"why":"Provides prior evidence that BERT's layers capture different levels of abstraction, used to interpret the layer-wise pattern.","marker":"Liu et al., 2019"},{"why":"Defines relational schemata and structure mapping, the theoretical notion of abstraction the study tests.","marker":"Gentner, 1983"},{"why":"Exemplifies emergent relational structure in word embeddings, the contrast case for emergent versus inducible structure.","marker":"Mikolov et al., 2013"},{"why":"Frames the 'right for the wrong reasons' concern that benchmark success need not reflect underlying competence.","marker":"McCoy et al., 2019"}],"fun_headline_variants":["Relational schema? Only after fine-tuning, not from pretraining","BERT's relational structure is trained in, not born with","High accuracy, low structure: BERT's relations aren't emergent","Fine-tuning induces BERT's relational schemata; pretraining doesn't"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Relational schema? Only after fine-tuning, not from pretraining","BERT's relational structure is trained in, not born with","High accuracy, low structure: BERT's relations aren't emergent","Fine-tuning induces BERT's relational schemata; pretraining doesn't"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000919,"raw_usage":{"total_tokens":3903,"prompt_tokens":868,"completion_tokens":3035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2969}},"tokens_in":484,"tokens_out":3035,"duration_ms":20889,"temperature":1.0,"reasoning_tokens":2969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:34.790976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}