{"id":"6e2e5faa-d58a-4819-9e8b-0695f86d854c","arxiv_id":"2504.14766","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Embedding Dimension Importance (EDI) ranks embedding dimensions by how strongly they encode individual linguistic properties, and a handful of top-ranked dimensions can recover most of a full classifier's accuracy.","lead":"This paper introduces a new metric and a new dataset for locating where individual BERT, GPT-2, and MPNet embedding dimensions store linguistic properties such as negation and tense. A generalist reader might care because it is a step toward making large language model representations more inspectable and auditable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mean pooling over unequal-length pairs makes the top EDI dimensions for negation, factuality, and intensifier indistinguishable from a token-insertion artifact; the control pairs do not control for this.","rationale":"The reader's weakest assumption identifies exactly the mean-pooling insertion confound, and I agree it is the most load-bearing threat to the central claim. The paper's own data show the pattern: negation builds every pair by appending 'not', factuality by appending 'could', and intensifier by appending 'surprisingly' (Table 2 and Appendix B). With mean pooling, the difference vector changes systematically with token count and with the inserted token's embedding, so the three statistical signals in EDI can all be driven by lexical-surface evidence rather than by a dimension-level encoding of the linguistic property. The unrelated control pairs do not have matched length or matched lexical insertion, so they cannot test this. The conditional verdict is appropriate because the concern is testable and because properties such as polarity, tense, voice, and definiteness, which do not add a token, are not subject to this particular confound; a weaker version of the claim about those properties could survive. The additional report that all control dimensions have equivalent Wilcoxon p-values (Section 6.1, Figure 5) should be checked as a possible implementation bug, since it would affect the 0.6-weighted term across all LPs, but the primary required action remains the neutral-insertion control.","tokens_in":18615,"tokens_out":9998,"duration_ms":100174,"concrete_test":"Build a matched-length/insertion control set with the same structure as the LDSP pairs: for each negation, factuality, and intensifier pair, create a control pair by inserting a semantically neutral filler word at the same position in the second sentence (e.g., 'The project is successful' vs 'The project is definitely successful'; 'The car is red' vs 'The car is visibly red'), keeping token counts and sentence order fully parallel. Run the full EDI pipeline (Wilcoxon, MI, RFE, EDI ranking) and Evaluation 1 on these controls, and compare the top-ranked dimensions to those reported for the real LPs. If the top-12 overlap (e.g., Jaccard index > 0.5) or the high-EDI accuracy curves match those in Figures 20/18/11, the top EDI dimensions are insertion artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 computes sentence embeddings by mean pooling over token vectors. For the three LPs whose defining operation is inserting an extra token ('not' for negation, 'could' for factuality, 'surprisingly' for intensifier), the second sentence is systematically one token longer, so every coordinate of the pooled vector shifts by a term proportional to (embedding_of_inserted_token - mean_of_other_tokens)/(n+1). The Wilcoxon p-values, mutual information, and RFE weights in the EDI score can therefore rank dimensions by this insertion signal alone, without any evidence that the dimension encodes the linguistic property as a generalizable feature. The paper's control LDSPs are completely unrelated sentences of uncontrolled length, so they do not rule out this confound. Because negation and factuality are among the paper's strongest and most emphasized results (EDI maxima 0.9987 and 0.9740; 11 and 4 dimensions reaching 95% baseline accuracy), the central claim that specific dimensions robustly encode these properties is not yet supported. A separate red flag is the report in Section 6.1 and Figure 5 that for control all dimensions had equivalent Wilcoxon p-values; this is only consistent with a bug in the per-dimension Wilcoxon computation, and it raises further doubt about the p-value component that carries weight 0.6 in EDI.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the LDSP-10 dataset, which consists of 1,000 sentence pairs per linguistic property (control, synonym, quantity, tense, intensifier, voice, definiteness, factuality, polarity, negation), and proposes the Embedding Dimension Impact (EDI) score that combines a Wilcoxon signed-rank p-value, mutual information, and logistic-regression RFE weights to rank embedding dimensions by their relevance to each property. The evaluation trains logistic classifiers on top-EDI, low-EDI, and cross-property dimension subsets for BERT, GPT-2, and MPNet embeddings, reporting that for most properties fewer than 12 top-ranked dimensions reach 95% of full-dimensional accuracy, while low-EDI dimensions are near chance. The paper concludes that properties such as negation and polarity are robustly encoded in specific dimensions, while synonymy shows more complex patterns.","tokens_in":18845,"tokens_out":6263,"duration_ms":55402,"significance":"If the central claim were established, the paper would provide a practical, model-agnostic framework for identifying dimension-level linguistic encodings, with potential applications in interpretability and bias mitigation. The dataset and the cross-model evaluation (BERT, GPT-2, MPNet) are potentially useful resources, and the broad evaluation design (held-out splits, low-EDI negative control, cross-property transfer) is a step in the right direction. However, the current evidence is not yet sufficient: the mean-pooling artifact for unequal-length pairs, the anomalous Wilcoxon report for the control condition, and the unclear split used when computing EDI scores all undermine the paper's strongest conclusions. These issues are fixable with additional controls and clarifications, and the framework itself retains value, but the manuscript currently overstates the robustness of its findings.","major_comments":[{"comment":"The mean-pooling procedure creates a systematic confound for negation, factuality, and intensifier. For these properties the second sentence is, by construction, longer by one token ('not', 'could', 'surprisingly'). Under mean pooling over token vectors, every coordinate of the sentence embedding shifts by a term proportional to (embedding_of_inserted_token − mean_of_other_tokens)/(n+1). The Wilcoxon, MI, and RFE components of the EDI score can therefore rank dimensions based on this insertion signal alone, without any evidence that the dimension encodes a generalizable linguistic property. The control LDSPs are completely unrelated sentences of uncontrolled length, so they do not rule out this artifact. Because negation (EDI max 0.9987) and factuality (EDI max 0.9740) are among the paper's most emphasized results, the claim that specific dimensions robustly encode these properties is not yet supported. The authors should add a control condition in which a neutral word (e.g., 'actually' or 'very') is inserted with the same structural pattern, and show that the EDI scores for this insertion control are substantially lower than for the target property.","section":"§4, Table 2"},{"comment":"The statement that for control 'all dimensions had equivalent Wilcoxon p-values' is statistically implausible unless the Wilcoxon computation contains a bug. For non-identical real-valued data across 768 dimensions, the probability that all p-values are equal is zero. Since the p-value component carries weight 0.6 in the EDI score, this observation raises doubts about the validity of the entire EDI ranking. If the p-values are in fact merely all above some threshold, that should be reported as such; if they are literally identical, the Wilcoxon implementation must be corrected. The authors should report the full distribution of p-values for the control condition and, if the anomaly is a coding error, rerun the analysis for all properties.","section":"§6.1, Figure 5"},{"comment":"The paper does not specify whether the EDI scores (including the RFE feature selection and the Wilcoxon/MI statistics) are computed on the training split only or on the full LDSP-10 dataset. The evaluation algorithms in Appendix C train and test classifiers, but the EDI ranking is not explicitly restricted to the training portion. If EDI is computed on the full dataset, then the test labels are used during feature selection, and the reported high-EDI accuracies (e.g., 11 dimensions for negation, 4 for factuality) reflect a form of label leakage. Please clarify exactly how the 80/20 split is used in all four evaluation stages (baseline, Evaluation 1, Evaluation 2, Evaluation 3), and if necessary, recompute the EDI scores on the training split only.","section":"§5.2, Appendix C"}],"minor_comments":[{"comment":"The definition of the Wilcoxon statistic is imprecise: 'The test ranks the absolute values of the nonzero differences... The Wilcoxon test statistic W is computed as the sum of ranks of positive |d_i|' mixes notation. Standard W is the sum of ranks of the positive differences (i.e., over d_i > 0), not over 'positive |d_i|'; consider rephrasing for clarity.","section":"§4.1"},{"comment":"The number of linguistic properties is inconsistent: the abstract says 'ten key linguistic features', Section 3 says 'nine LPs', and the list of properties in Section 3 omits 'voice' although Table 1 includes it. Please correct the count and ensure the property list matches the dataset.","section":"§3 and Table 1"},{"comment":"Both appendices begin with 'This section will contain the visualizations of the results for GPT-2/MPNet embeddings,' which suggests the manuscript is not yet final, especially since the figures are actually present. This text should be revised.","section":"Appendix E and F"},{"comment":"The captions for the evaluation plots do not explain all line styles and markers (e.g., what the red dashed line represents in Figure 9, and what the red dashed line in Figure 11 represents). Please add legends or fully descriptive captions.","section":"Figure 9/10/11"},{"comment":"The EDI weights (0.6/0.2/0.2), the MI 10-bin quantization, and the 0.8 EDI threshold are introduced without sensitivity analysis; the manuscript says only that the weights were chosen after 'our testing' (Section 4.4). Reporting how the main conclusions change under alternative weights or bin counts would substantially strengthen the robustness of the proposed metric.","section":"§4.4 and §6.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is currently undermined by the mean-pooling insertion confound and the anomalous Wilcoxon report for the control condition. The framework is salvageable, and the dataset and cross-model comparisons are useful, but the authors should be required to add a neutral-insertion control, verify the Wilcoxon implementation, and clarify the data split used for computing EDI scores. If those checks fail, the claimed dimension-specific encodings for negation and factuality would not be supported by the current evidence. I also note the appendix text in E/F ('This section will contain...') suggests the manuscript was not finalized, which should be addressed before any further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth reading for its LDSP-10 dataset and the EDI composite metric, but you should not trust the headline results for negation and factuality as evidence of localized linguistic encoding. The mean pooling step creates a straightforward length artifact for any property whose defining operation is inserting a token, and the control condition has a statistical anomaly that smells like a bug.\n\nWhat's genuinely new: LDSP-10, a set of LLM-generated minimal pairs for nine LPs plus a control, with manual validation reported at >99%. The EDI score—a weighted combination of Wilcoxon p-values, mutual information, and RFE weights—is a practical way to rank dimensions, and the paper evaluates it across BERT, GPT-2, and MPNet with several sensible checks: high-EDI subsets recover baseline accuracy with few dimensions, low-EDI dimensions are near chance, and cross-property transfer is weaker and semantically interpretable (negation helps polarity). That is a solid experimental skeleton.\n\nThe soft spots are not all equal. The most serious is the token-insertion confound. For negation, factuality, and intensifier, the two sentences differ in length, so the mean-pooled difference vector is exactly (embedding_of_inserted_token − sentence_mean)/(n+1). The top EDI dimensions for these properties may therefore simply reflect the lexical token, not a generalizable linguistic feature. The control pairs—unrelated sentences of uncontrolled length—cannot rule this out. This directly affects the paper's most emphasized results.\n\nSecond, the paper reports that in the control condition all dimensions had equivalent Wilcoxon p-values. That is statistically implausible for per-dimension tests on real embedding data, and it suggests the p-value computation (weight 0.6 in EDI) is buggy. This needs a concrete explanation before any dimension-level claim can be trusted.\n\nOther issues are smaller: the EDI weights (0.6/0.2/0.2) and the 10-bin MI discretization are justified by vague 'testing'; there are no error bars or repeated splits; and the abstract's bias-mitigation framing is not supported by the experiments. Also, if EDI is computed on the full dataset before the 80-20 split, feature selection leaks test labels—the paper doesn't say it isn't.\n\nWho gets value: people working on probing or feature-selection benchmarks will want to know about LDSP-10, and the paper is a good teaching example of why length-controlled controls matter. It deserves a serious referee, but the revision bar is high: fix the Wilcoxon bug, add equal-length controls (e.g., insert neutral words), report variance across splits, and temper the claims.","headline":"Useful dataset and composite metric, but the key results for negation and factuality are confounded by mean pooling over unequal-length pairs, and the control p-values point to a likely Wilcoxon bug.","tokens_in":19426,"tokens_out":7152,"would_cite":false,"duration_ms":63194,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a weighted score combining Wilcoxon p-values, mutual information, and recursive feature elimination ranks BERT embedding dimensions so that most linguistic properties reach near-full classifier accuracy with fewer…","keywords":["embedding interpretability","BERT","sentence embeddings","linguistic properties","feature selection","EDI score","probing","LDSP-10"],"falsifier":"Construct control sentence pairs identical in structure but with a neutral word of the same length inserted in place of the property word, then compute EDI scores for those controls; if the top-scoring dimensions for the real pairs and the filler pairs largely coincide, the localization finding is an artifact of length rather than of the linguistic property.","tokens_in":18306,"feed_emoji":"🧠","tokens_out":8960,"duration_ms":71790,"temperature":0.7,"pith_summary":"The paper tries to show that individual coordinates of sentence embeddings from pretrained language models carry identifiable linguistic information, not just statistical noise. To test this, it introduces LDSP-10, a dataset of one thousand minimally paired sentences per linguistic property, and an Embedding Dimension Importance (EDI) score that combines a signed-rank p-value, mutual information, and a logistic-regression weight from recursive feature elimination. Across BERT, GPT-2, and MPNet, the highest-ranked EDI dimensions reproduce at least 95% of full-embedding classification accuracy for most properties, while the lowest-ranked 100 dimensions perform near chance. The paper concludes that negation and polarity are localized in a small number of embedding coordinates, whereas synonymy is distributed across many.","feed_headline":"A dozen embedding dimensions carry most of BERT's language signal","feed_subtitle":"New EDI score ranks embedding coordinates; negation and polarity localize while synonymy spreads out.","key_machinery":"The load-bearing object is the Embedding Dimension Importance (EDI) score, defined for each dimension $d$ and linguistic property $lp$ as $EDI_{d,lp}=0.6(-\\log p_{d,lp})+0.2\\,M_{d,lp}+0.2\\,R_{d,lp}$, where $p$ is the signed-rank p-value over paired sentence embeddings, $M$ is the mutual information between the quantile-binned dimension and the sentence-position label, and $R$ is the absolute logistic-regression weight after recursive feature elimination, set to zero if the dimension is eliminated. The score ranks all dimensions per property, and the evaluation trains logistic classifiers on the top-ranked or bottom-ranked slices to test whether EDI actually separates informative coordinates from uninformative ones. The second load-bearing object is the LDSP-10 dataset: one thousand sentence pairs per property constructed so that each pair differs in exactly one linguistic feature, which is what makes dimension-wise attribution possible. Sentence embeddings are produced by mean-pooling the final hidden layer over tokens.","core_discovery":"The paper's central discovery is that the 768-dimensional embedding space is not uniformly opaque: coordinates can be ranked by how much they encode a given linguistic property, and the ranking predicts classifier behavior. For negation, dimension 544 of BERT's mean-pooled final layer reaches an EDI score of 0.9987, and a logistic classifier using the top 11 dimensions recovers 95% of the accuracy of the full-dimensional baseline. Polarity shows a similar pattern, with dimension 431 scoring 0.9947. Synonymy behaves differently: its maximum EDI score is 0.8751 with a steep drop-off, and 392 top-ranked dimensions are needed to reach the same 95% threshold. The paper reads this contrast as evidence that linguistic properties are encoded heterogeneously, with some features concentrated in focal dimensions and others spread across the representation.","pith_inferences":["Editorial inference: because insertion-based properties (negation, intensifier, factuality) change sentence length before mean pooling, control pairs that insert a neutral word of the same length would reveal whether the top EDI dimensions are encoding the property or merely the length shift.","Editorial inference: the EDI ranking could double as an intervention map: if mean-ablating high-EDI dimensions removes property-specific accuracy in a downstream task, the score would identify coordinates worth editing rather than just worth reading.","Editorial inference: applying the same minimal-pair construction to demographic contrasts (for example, gender-marked names or role nouns) could turn the framework into a bias-localization tool, although the paper only floats this as future work."],"forward_implications":["A logistic classifier built on fewer than twelve top EDI dimensions reaches at least 95% of full-embedding accuracy for most of the ten properties, so property classification can be run on a small coordinate subset.","The 100 lowest-ranked dimensions perform near chance for every property, indicating those coordinates contribute little property-specific signal and are candidates for pruning or masking.","Negation and polarity are the most localized properties: each has a single dimension with EDI near 1.0 and broad agreement among the three statistical tests.","Synonymy is the most distributed property, requiring 392 top-ranked dimensions to hit the 95% threshold, so word-substitution equivalence is not pinned to a few coordinates.","Cross-property transfer is partial and semantically structured: negation's top dimensions give the strongest transfer to polarity, while unrelated properties transfer poorly."],"supporting_citations":[{"why":"Supplies BERT, the primary model whose final-layer embeddings are analyzed dimension by dimension.","marker":"Devlin et al., 2019"},{"why":"Motivates the mutual-information term in the EDI score as an information-theoretic probe of linguistic structure.","marker":"Pimentel et al. (2020)"},{"why":"Supplies the recursive feature elimination procedure whose logistic-regression weights form the third EDI term.","marker":"Zeng et al., 2009"},{"why":"Defines the probing-task paradigm for sentence embeddings that the paper adapts to dimension-wise attribution.","marker":"Conneau et al. (2018)"},{"why":"Introduces intrinsic probing through dimension selection, the prior approach the EDI score extends.","marker":"Torroba Hennigen et al. (2020)"},{"why":"Identifies salient neurons via correlation analysis, providing the comparison point for focal versus distributed encoding.","marker":"Durrani et al. (2024)"},{"why":"Warns against conflating representational importance with functional utility, which the paper's classifier evaluations are designed to address.","marker":"Antverg and Belinkov (2022)"}],"fun_headline_variants":["Most of BERT's signal lives in a few targeted dimensions","EDI score ranks BERT coordinates: negation tight, synonymy diffuse","Dimension 544 drives negation; synonymy spans hundreds of coordinates","BERT's meaning splits: localized negation vs spread-out synonymy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the assumption that when a sentence pair differs by adding a word such as 'not' or 'surprisingly', the shift in the averaged word-vector representation is caused by the linguistic property itself and not merely by the sentence being longer or the inserted word's position.","fun_headline_variants_meta":{"raw":{"variants":["Most of BERT's signal lives in a few targeted dimensions","EDI score ranks BERT coordinates: negation tight, synonymy diffuse","Dimension 544 drives negation; synonymy spans hundreds of coordinates","BERT's meaning splits: localized negation vs spread-out synonymy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1528,"prompt_tokens":911,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":527,"tokens_out":617,"duration_ms":6167,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:42:03.931069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct control sentence pairs identical in structure but with a neutral word of the same length inserted in place of the property word, then compute EDI scores for those controls; if the top-scoring dimensions for the real pairs and the filler pairs largely coincide, the localization finding is an artifact of length rather than of the linguistic property.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the recursive feature elimination procedure whose logistic-regression weights form the third EDI term."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces intrinsic probing through dimension selection, the prior approach the EDI score extends."},{"cited_title":"Discovering Salient Neurons in Deep NLP Models","cited_arxiv_id":"2206.13288","evidence_quote":"Identifies salient neurons via correlation analysis, providing the comparison point for focal versus distributed encoding."}],"review_version":1}