{"id":"129f48f5-1cb4-4e32-a8e0-b9d0198aaa7f","arxiv_id":"2506.04373","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Sentence embeddings can be decomposed into sparse, interpretable atoms via supervised dictionary learning, and mean pooling preserves mainly atoms aligned with the sentence direction.","lead":"This paper applies dictionary learning to token embeddings from sentence encoders to break sentence representations into interpretable features, then tracks how mean pooling preserves or dilutes each feature. It is a proof-of-concept for making sentence embeddings more transparent and could support interpretability tools for retrieval and similarity systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'mechanistic decomposition' claim is unsupported by the current evidence: the supervised dictionary is a correlational, label-aligned fit, and no causal or sentence-level validation shows the atoms are the encoder's actual mechanisms.","rationale":"The reader identified the sparse-linear model as the weakest assumption; my concern is closely related but shifts the focus to the strength of the claim. Even if X≈DZ holds at token level, the dictionary could be one of many equivalent bases, and the paper's evidence does not show that the atoms are the encoder's mechanisms. The supervision makes the label-alignment partly self-fulfilling, so the interpretability results cannot independently confirm linear encodability beyond the probe experiments. This matters because the abstract and introduction promise 'mechanistic' and 'causal' insight; the reported analyses are descriptive. I recommend keeping the reader's CONDITIONAL verdict: the paper is a useful proof-of-concept, but the central claim should be accepted only after the sentence-level reconstruction and causal ablation check are run (or the claim is softened). I do not see grounds for rejection, since the probes and reconstruction numbers are legitimate evidence for a weaker claim.","tokens_in":17010,"tokens_out":9492,"duration_ms":99031,"concrete_test":"On a held-out set of 1,000 Brown sentences, compute the official sentence embedding s (model encode() with mean pooling), token codes z_t, and dictionary reconstruction D\\bar{z}; report cosine/R². Then causally intervene: zero the top-attribution atoms from §5.1, re-pool to s', and compare downstream similarity/probe changes against zeroing the same number of random or lowest-attribution atoms. Poor reconstruction or failure of top-atom ablations to shift behavior as predicted would weaken the 'mechanistic' claim to a descriptive supervised decomposition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, as stated in the abstract, is that the method 'mechanistically decompose[s] sentence embeddings into interpretable components' and that 'many semantic and syntactic aspects are linearly encoded in the embeddings.' For that claim to hold, the learned dictionary atoms must correspond to features the encoder actually uses, not merely to a basis that fits the activations well. The evidence in §4.2 and §5 consists of token-level reconstruction loss, POS/DEP prediction F1, and correlational attribution scores a_k = \\bar{z}_k ⟨d_k, s⟩. A sparse dictionary fit is not unique, and because the training objective includes the same POS/DEP labels later used to interpret the atoms, label-alignment of atoms is partly a consequence of supervision rather than an independent discovery about the model. No intervention or behavioral test shows that manipulating top-attribution atoms changes the encoder's output in the predicted direction, and no comparison against a dense low-rank baseline (e.g., PCA) establishes that the sparse-linear model is the correct description rather than a convenient approximation. The 'mechanistic' wording therefore overstates the current evidence; the method is so far a descriptive, label-aligned decomposition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a supervised dictionary-learning framework for token-level representations of sentence encoders and uses mean pooling of the learned sparse codes to attribute content to sentence embeddings. Experiments are run on the Brown corpus with three sentence-transformers (multilingual-e5-large, all-mpnet-base-v2, all-MiniLM-L6-v2). The paper reports that linear probes recover part-of-speech and dependency information, that supervised dictionary learning yields POS/DEP-aligned atoms, and that a contribution score ranks atoms and linguistic classes by their role in the pooled representation. The central claims are that many semantic and syntactic aspects are linearly encoded and that the method provides a mechanistic decomposition of sentence embeddings.","tokens_in":17297,"tokens_out":5894,"duration_ms":61172,"significance":"If the central claims held, the paper would offer a useful bridge between token-level dictionary-learning interpretability and sentence-level embedding analysis, with potential applications in transparency and controllable representation. Strengths of the manuscript include the use of three publicly available models, a reproducible hyperparameter sweep reported in full in the appendix, and the explicit probing baselines in Section 4.1. However, the evidence currently supports a descriptive, label-aligned decomposition rather than a mechanistic account: the dictionary is trained with the same POS/DEP labels later used to interpret it, no causal interventions are performed, and the quantitative results are reported without variance and on a single corpus. These gaps directly affect the strength of the abstract's 'mechanistically decompose' claim.","major_comments":[{"comment":"The supervised dictionary-learning objective in §3.1 includes the term λ L_sup(f_θ(z_i), y_i), and §4.2 states that the objective uses cross-entropy losses for POS and dependency prediction. The interpretability evidence in Figures 5, 6, and 14 then evaluates atom alignment with these very POS and DEP labels. This means that the apparent interpretability is at least partly a consequence of the supervision signal, rather than an independent discovery about the encoder's internal structure. An unsupervised dictionary-learning baseline, or an evaluation on held-out labels not used during training, is needed to support the claim that the atoms capture genuine latent features of the representations.","section":"§4.2, Figures 5/6/14"},{"comment":"The contribution score is defined as a_k = \\bar{z}_k ⟨d_k, s⟩. Since s = D\\bar{z} = Σ_j \\bar{z}_j d_j, and Figure 12 shows that the dictionary atoms are nearly orthogonal, ⟨d_k, s⟩ ≈ \\bar{z}_k ||d_k||^2, so a_k is approximately proportional to \\bar{z}_k^2. Under the reported near-orthogonality, the 'directional alignment' factor mostly restates the atom's usage, and the rankings in Figures 8–10 become largely rankings by squared mean activation. The paper should use an alignment measure that excludes the atom's own contribution to s, or explicitly qualify that a_k is a usage-weighted self-alignment score rather than an independent directional measure.","section":"§3.2, Eq. (3)"},{"comment":"The formal objective in §3.1 is min_{D,Z,θ} Σ_i ||x_i − Dz_i||² + λ L_sup(f_θ(z_i), y_i) subject to ||z_i||₀ ≤ ε, with no static-embedding reconstruction term and no explicit sparsity penalty. Section 4.2, however, states that the objective combines four terms: contextual reconstruction, POS/DEP cross-entropy, static-embedding reconstruction, and a sparsity penalty. Appendix C further lists hyperparameters such as alpha-pos, alpha-dep, alpha-static, and alpha-sparse (Figures 14b/14c) that do not appear in the formal equation. As written, these inconsistencies prevent the method from being reproduced from the equations alone, and the relationship among λ, the alpha weights, and the l1 coefficients should be specified precisely.","section":"§3.1 and §4.2"},{"comment":"The central claim that the method 'mechanistically decompose[s]' sentence embeddings is not supported by the present experiments. The dictionary is fit to token activations, and the contribution metric is descriptive; no intervention shows that removing, scaling, or otherwise manipulating an atom's contribution changes the encoder's output in a predicted direction, and no comparison against a dense low-rank baseline (e.g., PCA) establishes that the sparse-linear model describes the encoder's actual computation. In addition, the paper selects the best hyperparameter run from Tables 1–3 and reports all interpretability results for that single configuration, without variance across seeds or a held-out corpus. At minimum, the title, abstract, and Section 6 should be reframed to 'descriptive decomposition,' or causal tests and multi-run evaluation should be added.","section":"§5, Tables 1–3"}],"minor_comments":[{"comment":"The text contains a typo: 'Meal pooling compresses a sequence of token codes' should read 'Mean pooling compresses a sequence of token codes.'","section":"§5"},{"comment":"The Limitations section begins with 'This work is a bluerprint'; 'bluerprint' should be 'blueprint.'","section":"Limitations"},{"comment":"The word 'interpreatbility' appears in 'input-saliency interpreatbility methods'; it should be 'interpretability.'","section":"§2.2"},{"comment":"There is an inconsistency in the description of atom 58: Figure 6's caption says 'atom 58 detects adjectival modifiers,' while §5.1 says 'atom 58, which activates for adverbs describing actions.' Please clarify which category is intended.","section":"Figure 6 and §5.1"},{"comment":"The caption says 'POS-aligned heatmaps for four sentence embedding models,' but the figure contains only three panels: all-MiniLM-L6-v2, all-mpnet-base-v2, and multilingual-e5-base. Either a fourth panel is missing or the caption should be corrected.","section":"Figure 14 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible descriptive analysis of sentence-embedding structure, but the 'mechanistic' framing is currently stronger than the evidence supports. The label-supervised dictionary learning and the post-hoc selection of the best run are the two most important weaknesses; both are addressable within the manuscript's scope, but they require either substantially more evidence (unsupervised baselines, causal tests, variance reporting) or a deliberate reframing of the claims. I would not recommend rejection if the authors are willing to make that change, but the current version overstates its contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: this is a useful proof-of-concept, but the abstract's 'mechanistic' language oversells it. What's actually new is the combination: supervised dictionary learning on token embeddings of sentence encoders, followed by mean-pooling attribution to see which atoms survive into the sentence vector. That combination is new, and the probing results (linear POS probes around 0.89, DEP slightly nonlinear) are clean and support the weaker claim that syntactic information is largely linearly encoded. The atom-level POS/DEP alignment in Figures 5 and 6 and the pooling attribution patterns (nouns, verbs, and root dominating) are plausible and match prior work. If the code is actually released as stated, reproducibility is feasible.\n\nSoft spots, in order. First, circularity: the dictionary is trained with POS/DEP cross-entropy losses (Section 4.2), then the same labels are used to show the atoms are interpretable (Figures 5, 6, 14). That atoms align with the supervision is partly enforced, not discovered. Second, the attribution score a_k = zbar_k <d_k, s> is constructed so frequent, high-alignment atoms dominate; the finding that nouns and verbs dominate is partly baked in. Third, no error bars, one corpus, and the analysis picks the best hyperparameter configuration post-hoc. Fourth, no comparison to a dense low-rank baseline like PCA, so we don't know that the sparse-linear model is the right description rather than a convenient approximation. Fifth, and most important for the 'mechanistic' label: there's no intervention test. We never see whether perturbing a top atom changes the encoder's output in the predicted direction. So the strong claim is unsupported. The paper itself calls it a blueprint and lists limitations, which is honest, but the abstract still oversells.\n\nWho this is for: people working on sentence embedding interpretability or sparse dictionary methods. It is a reasonable starting point, not a finished mechanistic account. I would send it to review; a good referee can push the authors to re-frame the claims, add variance and a PCA baseline, and ideally one causal or behavioral check. That would make the contribution solid.","headline":"A useful proof-of-concept for decomposing sentence embeddings, but the 'mechanistic' label oversells a descriptive, label-aligned dictionary method.","tokens_in":17776,"tokens_out":1929,"would_cite":true,"duration_ms":25088,"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":"The paper claims that sentence embeddings can be decomposed into sparse, interpretable atoms via supervised dictionary learning on token representations, and that mean pooling preserves the atoms best aligned with the sentence direction.","keywords":["sentence embeddings","dictionary learning","mechanistic interpretability","mean pooling","sparse coding","probing","part-of-speech tagging","linear encoding"],"falsifier":"Take a sentence, identify its dominant atom by the attribution score, then intervene by setting that atom's activation to zero (or to a large value) in the decomposed code and feed the reconstructed embedding into a downstream similarity or classification task. If the decomposition is mechanistic, the change in output should be predictable from the atom's labeled feature; if the output barely moves or changes erratically, the atoms are only a low-rank approximation with no causal force.","tokens_in":16819,"feed_emoji":"🧩","tokens_out":11263,"duration_ms":84275,"temperature":0.7,"pith_summary":"Sentence embeddings are central to modern language AI, but their content is normally opaque: a sentence becomes a dense vector that can be compared but not read. This paper argues that the content can be recovered by decomposing each token's embedding into a sparse combination of learned dictionary atoms, guided by linguistic labels such as part of speech and dependency relations. Because the final sentence vector is obtained by mean pooling the tokens, the pooled representation inherits a weighted average of those atoms, and the authors show that each atom's contribution can be scored by how often it is used and how well it aligns with the sentence direction. The central finding is that many semantic and syntactic properties are linearly encoded and survive pooling, which makes sentence embeddings inspectable and opens a path to controlling what they represent.","feed_headline":"Sentence embeddings decompose into sparse, interpretable atoms","feed_subtitle":"The method exposes which linguistic features survive pooling, making sentence embeddings inspectable.","key_machinery":"The central object is the sparse dictionary model $X \\approx DZ$, in which each token embedding $x_t$ is written as a linear combination of $k$ dictionary atoms with a code $z_t$ satisfying $\\|z_t\\|_0 \\le 5$. The dictionary is trained with a supervised objective combining reconstruction error, cross-entropy losses for part-of-speech and dependency prediction, reconstruction of the static word embedding, and a sparsity penalty. The pooling analysis then rests on the identity $s = D\\bar{z}$, where $\\bar{z}$ is the mean of the token codes; this makes the sentence vector a weighted combination of the same atoms. Each atom's contribution is scored as $a_k = \\bar{z}_k \\langle d_k, s\\rangle$ — the product of how frequently the atom is used and how strongly it aligns with the overall sentence direction — which lets the authors rank atoms and linguistic classes by their impact on the final embedding.","core_discovery":"The paper's central claim is that a sentence embedding is not an opaque mixture but a weighted sum of a small number of interpretable directions. Concretely, the authors model token representations as $x_t = D z_t$, where $D$ is a learned dictionary of atoms and each sparse code $z_t$ has at most five nonzero entries, and they learn $D$ together with a classifier so that atoms line up with syntactic roles. Applying this to three sentence encoders on a Brown-corpus sample, they find that linear probes nearly match nonlinear probes for part-of-speech prediction (up to 0.89 on MiniLM), that individual atoms specialize in features such as numerals or adjectival modifiers, and that after mean pooling the dominant contributors are semantically central categories like nouns, verbs, and roots. The authors take this as evidence that much semantic and syntactic information is linearly encoded in the embeddings, with the pooled vector inheriting a weighted average of interpretable atoms.","pith_inferences":["If the sparse-linear hypothesis holds broadly, interpretability tools learned on one encoder or language could transfer to others, since the underlying dictionaries may share structural similarities.","The contribution score $a_k = \\bar{z}_k \\langle d_k, s\\rangle$ implies a testable prediction: sentences whose dominant atoms belong to a given semantic class should cluster along those atoms' directions, which contrastive similarity experiments could verify.","Comparing the same contribution score under max or attention pooling would quantify which pooling strategy best preserves rare or spike-activating atoms, a natural extension the paper leaves open.","If rare tags such as interjections really live in their own subspaces, targeted manipulation of those directions could enable few-shot adaptation or bias control without retraining."],"forward_implications":["A sentence embedding can be read as a small set of weighted interpretable atoms, giving a natural way to explain similarity, retrieval, or classification decisions in terms of specific linguistic features.","Mean pooling is transparent rather than lossy in this view: it linearly aggregates sparse codes, so the pooled vector's content is expressible as an average of atom activations.","The diagnostic transfers to any mean-pooled sentence encoder, offering a scalable route to feature attribution without retraining the encoder.","The dominance of nouns, verbs, and syntactic roots in the attribution scores gives empirical weight to linguistic theories that place core semantic roles near the root of a sentence.","Rare categories such as interjections appear to occupy dedicated singular vectors in the probe weights, suggesting low-frequency features have their own linear subspaces even when they contribute little to the pooled representation."],"supporting_citations":[{"why":"Supplies the supervised dictionary-learning objective the paper adapts to token embeddings.","marker":"Mairal et al., 2008"},{"why":"Establishes the dictionary-learning-on-neural-activations paradigm for monosemantic features that this work extends to sentence level.","marker":"Bricken et al., 2023"},{"why":"Shows sparse autoencoders find interpretable token-level features, motivating the decomposition approach.","marker":"Cunningham et al., 2023"},{"why":"Provides the structural-probe technique used to test whether syntax is linearly recoverable.","marker":"Hewitt and Manning, 2019"},{"why":"Documents that sentence embeddings encode linguistic properties, the claim this work makes inspectable.","marker":"Conneau et al., 2018"},{"why":"Supplies the sentence-transformers encoders and mean-pooling setup on which all experiments run.","marker":"Reimers and Gurevych, 2019"},{"why":"Justifies the subword-aggregation step used to obtain token-level representations before probing and dictionary learning.","marker":"Ács et al., 2021"},{"why":"Prior explainable-feature decomposition of sentence embeddings that this mechanistic approach complements.","marker":"Opitz and Frank, 2022"}],"fun_headline_variants":["Sparse atoms expose semantic linearity in sentence embeddings","Dictionary learning splits sentence embeddings into interpretable pieces","Sentence vectors decode as weighted sums of semantic atoms","Pooled embeddings inherit sparse, interpretable features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes that token embeddings are well approximated as sparse linear combinations of a fixed dictionary, with at most five nonzero coefficients per token; if that sparse-linear model is wrong, the learned atoms and the pooling attributions do not describe the encoder's actual computation.","fun_headline_variants_meta":{"raw":{"variants":["Sparse atoms expose semantic linearity in sentence embeddings","Dictionary learning splits sentence embeddings into interpretable pieces","Sentence vectors decode as weighted sums of semantic atoms","Pooled embeddings inherit sparse, interpretable features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1313,"prompt_tokens":875,"completion_tokens":438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":390}},"tokens_in":491,"tokens_out":438,"duration_ms":4777,"temperature":1.0,"reasoning_tokens":390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:42:57.936770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sentence, identify its dominant atom by the attribution score, then intervene by setting that atom's activation to zero (or to a large value) in the decomposed code and feed the reconstructed embedding into a downstream similarity or classification task. If the decomposition is mechanistic, the change in output should be predictable from the atom's labeled feature; if the output barely moves or changes erratically, the atoms are only a low-rank approximation with no causal force.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the supervised dictionary-learning objective the paper adapts to token embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the dictionary-learning-on-neural-activations paradigm for monosemantic features that this work extends to sentence level."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior explainable-feature decomposition of sentence embeddings that this mechanistic approach complements."}],"review_version":1}