{"id":"062953d3-95a7-4554-963b-9eb288207a84","arxiv_id":"2508.20322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SLiCS learns group-structured non-negative dictionaries that disentangle dense image embeddings into concept components, improving concept-filtered retrieval and enabling image-to-prompt generation.","lead":"The authors introduce SLiCS, a dictionary-learning method that decomposes CLIP, TiTok, and DINOv2 image embeddings into separate concept-specific subspaces, allowing search and generation to focus on one concept at a time. This makes concept-filtered image retrieval more precise and offers a way to interpret dense embeddings at a concept level.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main retrieval experiments use partial cone projections (Eq. 3), not the joint disentangled components from Eq. (2), so the reported gains do not test the additive decomposition claim.","rationale":"Reader's concern about the linear additive model is real, but I sharpen it: the evidence does not even exercise the additive model, because the evaluation uses partial projections rather than the joint decomposition from Eq. (2). Moreover, Fig. 7's result—joint v*_j underperforms partial projection—is direct internal evidence that the Eq. (1) decomposition is not behaving as a disentangler, since a successful disentangler should make the isolated component useful for retrieving that concept. This is not an external consensus objection; it is an internal consistency issue between the claimed decomposition and the evaluated quantity. The paper has real strengths: consistent gains across CLIP, DINOv2, and TiTok embeddings, an internal SVD baseline, and deterministic supervised results on MIRFlickr25K. The conditional verdict remains appropriate: the method is plausible and empirically useful, but the core 'disentanglement' claim needs a direct test before full acceptance.","tokens_in":19677,"tokens_out":6979,"duration_ms":73375,"concrete_test":"Using the held-out validation set, compute for every query the joint decomposition v*_j = B_j α_j from Eq. (2) with α obtained by NNLS on the ground-truth active label set, and the partial projection ᵥ̃*_j from Eq. (3). Then run two checks: (i) support recovery: estimate active concepts per image from the coefficients (e.g., by group OMP with non-negativity or by thresholding NNLS coefficients) and report F1 against the ground-truth label vector; (ii) concept isolation: for each concept j, use the CLIP text embeddings of all S concept names to compute pairwise cosine similarity between v*_j (and ᵥ̃*_j) and each concept text embedding, and report the rank of concept j and the off-target mass.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CLIP embeddings decompose as x* ≈ Σ_j B_j α_j with α_j ≥ 0 and α_j = 0 when concept j is absent (Eq. 1). The concept-filtered retrieval results in Table 2, however, rank candidates by cos(ᵥ̃*_j, x), where ᵥ̃*_j is the partial projection onto the single cone V_j (Eq. 3), not the joint component v*_j = B_j α_j obtained from Eq. (2). Because the cones are not orthogonal and can overlap, ᵥ̃*_j is the point in V_j closest to the whole embedding and can carry information about other concepts; it is not the concept-isolated part of x*. The paper's own Fig. 7 shows that this partial projection outperforms the joint component for retrieval, which is what one would expect if the joint additive model is not actually isolating concepts: the projection retains extra information that helps match the query. Thus the mAP improvements are consistent with a much weaker statement, namely that supervised per-concept positive-cone projections are useful retrieval filters. They do not establish support recovery (α_j = 0 iff concept absent), reconstruction fidelity of Eq. (1), or cross-concept isolation of v*_j. The convergence guarantee in §3.3.2 is also conditional on an error-reduction check that the authors say they skip in practice, so the 'guaranteed convergence' claim does not apply to the algorithm actually run.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SLiCS (Sparse Linear Concept Subspaces), a supervised dictionary-learning method that decomposes a dense image embedding into a sum of concept-specific components, each a non-negative combination of atoms associated with one concept. The method is applied to CLIP, TiTok, and DINOv2 embeddings, and is evaluated primarily through concept-filtered image retrieval (mAP@20), qualitative word-caption interpretation of learned subspaces, and image-to-prompt generation. An unsupervised variant uses CLIP zero-shot classification to generate pseudo-labels. The central claims are that the learned decomposition is semantically meaningful, that concept-filtered retrieval improves over unfiltered CLIP and SpLiCE baselines, and that the proposed alternating optimization has guaranteed convergence.","tokens_in":19906,"tokens_out":3126,"duration_ms":31840,"significance":"If the linear additive model of Eq. (1) is valid, SLiCS would provide a practical tool for interpreting dense visual embeddings and for concept-focused retrieval without retraining the backbone. The paper's strengths are its breadth of empirical evaluation across three embedding families (CLIP, TiTok, DINOv2), consistent mAP@20 improvements over UF-CLIP, SpLiCE, and F-SpLiCE, the inclusion of a rank-1 sign-optimality proof in §3.3.1, and the practical extension to product-quantized candidate pools for TiTok. However, the current experiments mostly validate a weaker statement: per-concept positive-cone projections are useful retrieval filters. The full additive-decomposition claim, including support recovery and cross-concept isolation, is not directly tested, and the convergence guarantee in §3.3.2 does not apply to the algorithm variant that is actually run.","major_comments":[{"comment":"The headline retrieval results are computed with the partial projection ṽ*_j from Eq. (3), not with the joint additive component v*_j = B_j α_j from Eq. (2). Because the concept cones can overlap, ṽ*_j is the point in one cone closest to the whole embedding and can carry information about concepts other than j; Fig. 7 confirms that this partial projection outperforms the joint component for retrieval. Consequently, the mAP gains reported in Tables 2–4 are consistent with the much weaker claim that supervised per-concept positive-cone projections are effective retrieval filters, and they do not establish Eq. (1) as a faithful decomposition, nor do they establish that α_j = 0 iff concept j is absent. Please report retrieval using v*_j from Eq. (2), report reconstruction error of Eq. (1) on held-out data, and evaluate support recovery (active/inactive concept classification) against ground-truth labels.","section":"§3.2 and §3.5, Table 2"},{"comment":"The claimed convergence guarantee applies only to the cyclic-minimizer variant that includes the error-reduction check in Eq. (13), but the text immediately states: 'In practice, we let the algorithm proceed with an atom update without checking if it will increase the reconstruction error.' Therefore the algorithm actually run, and for which all experimental results are reported, has no proven convergence guarantee. The abstract's statement that the dictionary is 'optimized through a novel alternating optimization with guaranteed convergence' is thus an overstatement. Please either run the version with the check and report the effect on accuracy, or revise the convergence claim to describe the checked variant and provide empirical convergence evidence for the unchecked variant.","section":"§3.3.2 and Abstract"},{"comment":"The load-bearing premise of the paper is the linear additive synthesis model x* ≈ Σ_j B_j α_j with non-negative group-sparse coefficients. This assumption is not validated directly anywhere in the manuscript. No quantitative reconstruction error for Eq. (1) is reported for any embedding, no comparison is made against unconstrained linear subspaces or a simple per-concept non-negative projection baseline of matched capacity, and no test is provided showing that the estimated α_j is zero exactly when the ground-truth label is absent. Without such evidence, the 'disentanglement' interpretation of the learned components remains an artifact hypothesis. Please add these direct measurements and ablations, or substantially soften the disentanglement claims.","section":"§3.1–§3.2 and §4.4"}],"minor_comments":[{"comment":"In the atom-update step, the support set is defined as L_m ← {l ∈ {1,...,M} : A_ml ≠ 0}; this should be over training samples, i.e., l ∈ {1,...,N}, not over atoms.","section":"Algorithm 1"},{"comment":"The dataset name is misspelled as 'MIRFlikr25K' in the conclusion; it should be 'MIRFlickr25K'.","section":"§5"},{"comment":"The caption of Fig. 4 lists panels (a), (c), and (d) with d0 = 2, d0 = 5, and d0 = 20, but there is no panel (b) described, and the text refers to 'Fig. 4(b)' in §4.2. Please fix the panel labels and the corresponding in-text references.","section":"Figure 4"},{"comment":"The sentence '6 concepts have have finer-grained labels' contains a duplicated word and should read '6 concepts have finer-grained labels.'","section":"§4.1"},{"comment":"The 'unsupervised' variant requires the set of concept words and the expected number of active concepts S̃ to be known in advance, and it uses CLIP text embeddings to generate pseudo-labels. This is more accurately described as zero-shot-supervised or pseudo-supervised; the current terminology may overstate the degree of unsupervised learning.","section":"§3.4"},{"comment":"Some unsupervised word captions are clearly not semantically consistent with their subspace (e.g., 'aircraft', 'plane', and 'airplanes' assigned to 'electronic' and 'appliance'). The text mentions this briefly but could analyze it more explicitly as evidence about pseudo-label noise and its effects on disentanglement quality.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for a vision or representation-learning venue. The main risk is that the central 'disentanglement' claim outruns the evidence: the retrieval experiments use partial cone projections rather than the joint additive decomposition, and the convergence guarantee is for a variant that is not run. I would encourage the editor to treat the revision as conditional on direct validation of Eq. (1) or on a revised, more limited claim. Novelty relative to SpLiCE is sufficient, since group-structured supervised dictionary learning and the multi-embedding evaluation are not present in that baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nBottom line: this is a decent subfield contribution on concept-filtered retrieval from CLIP-style embeddings, but the core claim—that the method truly disentangles embeddings additively—doesn't have the evidence the abstract suggests. The reader's conditional verdict and the stress-test note both land correctly.\n\nWhat's new: a group-structured dictionary over positive cones for concepts, learned with a K-SVD-style alternating scheme and non-negative coefficients. The sign-optimality proof for the atom update is a neat theoretical touch. Experiments are broad: they apply it to CLIP (ResNet-50 and ViT-B/32), TiTok, and DINOv2, and the mAP@20 gains over unfiltered retrieval and SpLiCE are consistent and sometimes large. That part is real and useful.\n\nWhere it falls short: the retrieval experiments use the partial cone projection (Eq. 3), not the joint disentangled component from Eq. (2). The stress-test is right—v~*_j is the closest vector in concept j's cone to the whole embedding, so it can carry information about other concepts. Fig. 7 shows this projection actually works better for retrieval than the joint component. So the reported numbers validate supervised per-concept cone projections as retrieval filters, not the additive disentanglement model in Eq. (1). Support recovery, reconstruction fidelity, and cross-concept isolation are never directly tested. The paper honestly admits the partial projection is better, but the title and abstract still sell the decomposition.\n\nSecond: the convergence guarantee is for the version with an error-reduction check that they say they skip in practice. So \"guaranteed convergence\" doesn't apply to the algorithm actually run. Not fatal, but it oversells the theory.\n\nThird: no code release. For a method paper, that's a real barrier to checking the details.\n\nThe unsupervised variant uses CLIP text embeddings to generate pseudo-labels, which is mildly circular in defining concepts, but the retrieval evaluation against ground-truth labels keeps the main numbers clean.\n\nOverall: worth engaging with. Send it to peer review. A referee should push for direct validation of the additive model—synthetic mixtures or support recovery metrics—or a reframing as a projection-based retrieval method. The method itself looks sound and reproducible in principle; the claims need tightening.\n\nBest.","headline":"Useful retrieval technique, but the stated disentanglement claim is not supported by the headline experiments.","tokens_in":20508,"tokens_out":3299,"would_cite":true,"duration_ms":30740,"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 a dense image embedding can be approximated as a sum of concept-specific positive-cone components, and that retrieving with one such component is more precise than retrieving with the whole embedding.","keywords":["concept disentanglement","dictionary learning","image retrieval","CLIP embeddings","DINOv2","TiTok","non-negative least squares","zero-shot labels"],"falsifier":"Build paired queries that are the same scene with and without one added concept (for example, a person in an otherwise identical room), then compare the difference of the two SLiCS decompositions to the person concept's component; if adding the person substantially changes the other concepts' components or the difference is not close to the person cone, the additive model fails.","tokens_in":19402,"feed_emoji":"🧩","tokens_out":10691,"duration_ms":87958,"temperature":0.7,"pith_summary":"SLiCS argues that a dense vision-language embedding is not an unanalyzable whole: it can be written, approximately, as the sum of one component per concept, each component lying in a positive cone spanned by atoms reserved for that concept. The paper shows how to learn those concept cones from multi-label supervision, or from pseudo-labels produced by zero-shot classification of concept text embeddings, using a dictionary-learning routine that alternates non-negative least-squares coefficient estimation with rank-1 singular-value atom updates. If the claim holds, a user can decide which aspects of a query image should drive retrieval, because only the component for the chosen concept is used for ranking. The paper reports that concept-filtered retrieval sharply improves mAP@20 over whole-embedding retrieval on CLIP, DINOv2, and TiTok embeddings, and that the learned cones carry consistent word descriptions linking the decomposition back to text space.","feed_headline":"Slicing embeddings by concept sharpens focused retrieval","feed_subtitle":"A sparse dictionary of concept cones lifts filtered retrieval to 0.929 mAP@20 on MS COCO, beating unfiltered CLIP.","key_machinery":"The machinery is a group-structured semi-non-negative dictionary: a global dictionary $B$ is partitioned into blocks $B_j$, each block spanning the positive cone $\\{B_j a : a \\ge 0\\}$ for one concept, and coefficients are constrained to be non-negative and group-sparse according to the label vector. The optimization alternates a non-negative least-squares step for the coefficients with a K-SVD-style atom update in which each atom is refreshed from the rank-1 truncated SVD of the residual restricted to samples where that atom is active; a majority-sign rule for the atom's polarity is proved optimal, and thresholding restores the non-negativity of the coefficients. The component used for retrieval is the partial projection $\\tilde{v}^\\star_j = B_j \\alpha_j$ obtained by solving the non-negative least-squares problem $\\min_{a \\ge 0} \\|x^\\star - B_j a\\|^2$, which keeps the component close to the query while masking out other concepts.","core_discovery":"The paper's central claim is the linear synthesis model $x^\\star \\approx \\sum_{j=1}^S B_j \\alpha_j$, in which an embedding is decomposed into $S$ concept components, with $B_j$ the dictionary of atoms for concept $j$ and $\\alpha_j$ a non-negative coefficient vector that is zero when the concept is absent. The discovery is that this group-structured, non-negative decomposition can be learned and is useful: the proposed SLiCS algorithm initializes each concept's atoms by truncated SVD on embeddings that contain that concept, then alternates non-negative least-squares coefficient updates and K-SVD-style atom updates, using a provably optimal sign choice followed by thresholding to keep coefficients non-negative. Given the learned dictionaries, approximating a query embedding with only one concept's cone yields a component that isolates that concept, and retrieval ranked by cosine similarity to this component outperforms retrieval with the whole embedding for concept-specific queries.","pith_inferences":["Editorial inference: the additive cone model implies a compositional test: if concept components truly sum, removing one concept's component from an embedding should erase that concept from retrieval while leaving other concepts intact, a property that is implied but not directly measured in the paper.","Editorial inference: because the non-negative cone can only add presence, concept-filtered retrieval for negated queries (images without a concept) would need an explicit complement subspace, which is a natural extension not addressed by SLiCS as described.","Editorial inference: applying SLiCS to multiple encoders for the same concept could expose whether different vision models build aligned or incompatible concept geometry, turning the method into a comparative probe of embedding spaces.","Editorial inference: the near-orthogonality of atoms within a concept group and the group-structured co-occurrence pattern suggest the learned cones behave like simplex codes; if that structure is consistent, concept arithmetic in the embedding space should support editing operations such as swapping one concept for another."],"forward_implications":["Concept-filtered retrieval becomes a practical query mode: ranking candidates by cosine similarity to the query's concept component rather than to the whole embedding, with reported mAP@20 gains such as 0.895 versus 0.728 on MIRFlickr25K and 0.929 versus 0.791 on MS COCO for CLIP ViT-B/32.","Training labels can be replaced by pseudo-labels: zero-shot classification of concept text embeddings against training images supplies the group-sparsity pattern, so the same dictionary learning runs without manual annotation beyond a list of concept names.","The same decomposition applies across embedding families, including contrastive CLIP, self-supervised DINOv2, and highly compressed TiTok token embeddings, with DINOv2 yielding the highest filtered retrieval in the reported experiments.","Each concept cone can be assigned human-readable words: the concept atoms that best reconstruct text embeddings of frequent words yield subspace captions consistent across supervised and unsupervised training.","Disentangled concept components can drive conditional generation by converting a component into a prompt for a latent diffusion model after an orthogonal Procrustes alignment of the image and text spaces."],"supporting_citations":[{"why":"This citation supplies the CLIP image-text co-embedding space that SLiCS decomposes into concept cones.","marker":"[Radford et al., 2021]"},{"why":"This citation supplies the K-SVD dictionary-learning update that SLiCS adapts to group structure and non-negative coefficients.","marker":"[Aharon et al., 2006]"},{"why":"This citation defines the SpLiCE sparse-coding baseline and the mean-centering preprocessing used before CLIP dictionary learning.","marker":"[Bhalla et al., 2024]"},{"why":"This citation establishes that rank-1 semi-non-negative matrix factorization is NP-hard, motivating the paper's heuristic SVD-plus-threshold atom update.","marker":"[Gillis and Kumar, 2015]"},{"why":"This citation documents the modality gap whose mean is subtracted from CLIP image embeddings before training.","marker":"[Liang et al., 2022]"},{"why":"This citation supplies the TiTok-L-32 tokens and codebook used to test SLiCS on compressed autoencoder embeddings.","marker":"[Yu et al., 2024]"},{"why":"This citation supplies the DINOv2 ViT-B/14 embeddings used to test SLiCS on self-supervised features.","marker":"[Jose et al., 2024]"},{"why":"This citation provides the image-to-prompt conversion used to visualize concept components through conditional generation.","marker":"[Ding et al., 2023]"}],"fun_headline_variants":["SLiCS sparse concept subspaces boost retrieval precision","Disentangling embeddings into concept cones sharpens retrieval","Concept-filtered retrieval hits 0.929 mAP@20 with SLiCS","Sparse concept subspaces separate scene info for precise retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a dense image embedding is well approximated by a non-negative sum of concept-specific components; if concepts compose non-additively, interact during encoding, or leave negative semantic traces, the learned components are fitting artifacts rather than true concept separation.","fun_headline_variants_meta":{"raw":{"variants":["SLiCS sparse concept subspaces boost retrieval precision","Disentangling embeddings into concept cones sharpens retrieval","Concept-filtered retrieval hits 0.929 mAP@20 with SLiCS","Sparse concept subspaces separate scene info for precise retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1414,"prompt_tokens":1012,"completion_tokens":402,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":628,"tokens_out":402,"duration_ms":3911,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:46:31.240231+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build paired queries that are the same scene with and without one added concept (for example, a person in an otherwise identical room), then compare the difference of the two SLiCS decompositions to the person concept's component; if adding the person substantially changes the other concepts' components or the difference is not close to the person cone, the additive model fails.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"This citation supplies the CLIP image-text co-embedding space that SLiCS decomposes into concept cones."},{"cited_title":"Calmon, and Himabindu Lakkaraju","cited_arxiv_id":null,"evidence_quote":"This citation defines the SpLiCE sparse-coding baseline and the mean-centering preprocessing used before CLIP dictionary learning."},{"cited_title":"Exact and heuristic algorithms for semi-nonnegative matrix factorization","cited_arxiv_id":null,"evidence_quote":"This citation establishes that rank-1 semi-non-negative matrix factorization is NP-hard, motivating the paper's heuristic SVD-plus-threshold atom update."},{"cited_title":"Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning","cited_arxiv_id":null,"evidence_quote":"This citation documents the modality gap whose mean is subtracted from CLIP image embeddings before training."},{"cited_title":"The CLIP model is secretly an image-to-prompt converter","cited_arxiv_id":null,"evidence_quote":"This citation provides the image-to-prompt conversion used to visualize concept components through conditional generation."}],"review_version":1}