{"id":"978ac82d-67f5-4b77-817c-fc3377158a9a","arxiv_id":"2412.19847","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ArSyD learns image representations where each generative factor is a separate hypervector, enabling property editing by vector exchange and dimension-agnostic disentanglement evaluation.","lead":"ArSyD is a new image model that stores each object attribute, such as color or shape, as its own hypervector, then combines these vectors so an image can be edited by swapping one attribute vector for another. The authors demonstrate controlled edits on synthetic datasets and early tests on faces, and they propose new metrics for comparing disentangled representations of different types.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'disentanglement by construction' claim is unsupported: the exchange loss (Eq. 5) does not prevent V*_i (Eq. 4) from encoding off-target factors, and the paper's own Figure 7 shows orientation remains entangled with shape.","rationale":"The reader's weakest assumption already identified the same point: the premise that attention outputs align one-to-one with factors is unproven and contradicted by Figure 7. I agree with that identification. The concern is load-bearing because the headline novelty of the paper is that disentanglement is achieved by construction; if V*_i can be factor-entangled, the architecture is just another weakly supervised encoder with an HDC-flavored latent, and the claimed guarantee disappears. The proposed probe experiment is the cleanest way to settle it: it directly measures whether V*_i is factor-pure, independent of decoder quality or the metric definitions. The DCM metric error noted by the reader is a separate correctness issue, but it is not the central representation claim; fixing the metric equation without addressing factor alignment would not rescue the 'by construction' statement. I therefore keep the reader's CONDITIONAL verdict rather than escalating, because the concern is substantial but could in principle be addressed by adding an explicit orthogonality/independence loss and re-evaluating, or by softening the claim to 'approximate disentanglement for some factors'.","tokens_in":20950,"tokens_out":6621,"duration_ms":62277,"concrete_test":"Train ArSyD on dSprites paired with the paper's settings, then probe the learned per-factor vectors: for each factor index i, train a linear classifier (using the same training split) to predict each ground-truth generative factor from V*_i alone. Under the alignment claim, V*_i should predict factor i well above chance and all other factors at chance. If V*_orientation (or any V*_i) predicts a non-target factor above chance by a statistically significant margin over 3 seeds, the V*_i are not factor-pure and the central claim fails. This directly tests the premise behind Eqs. (3)-(4) without relying on decoder quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that ArSyD achieves disentanglement 'by construction' (Abstract) and that each factor value vector V*_i (Eq. 4) is aligned with its intended generative factor. The load-bearing premise is that training only the attention weights (Eq. 3) over a frozen random codebook, with only reconstruction MSE (Eq. 5) as supervision, yields factor-pure vectors. This premise is not established and is arguably false. In each training pair, all factors except index p are identical between target and donor. The exchange loss therefore requires of V*_p only that it carry enough information to transfer factor p; it never penalizes V*_p for also encoding other factors, because those factors are constant in the pair and the decoder can ignore the redundant signal. The HDC binding in Eq. (1) makes the bound components G_i ⊛ V*_i quasi-orthogonal to one another, but it does not constrain V*_i itself to be independent of other factors. As a result, the mapping from image to V*_i can alias factors without increasing training loss. The paper's own qualitative results provide direct evidence: Section 4 and Figure 7 state that the 'Orientation' feature is 'strongly related' to the 'Shape' feature. Thus the abstract's 'by construction' claim overstates what the architecture guarantees; the central representation claim hinges on an unverified and partially contradicted alignment assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ArSyD, an architecture that learns 'symbolic disentangled representations' in which each generative factor is represented by a full hypervector, and the object representation is a superposition of factor-value vectors obtained by attention over a fixed random codebook. The model is trained with weak supervision: image pairs differ in exactly one factor, and the feature-exchange reconstruction loss (Eq. 5) is the only training signal. The paper also introduces two classifier-based metrics, DMM and DCM, intended to evaluate disentanglement and compactness for distributed latent representations. Experiments on dSprites, CLEVR1, CLEVR5 (with Slot Attention), and a CelebA proof-of-concept show qualitative controlled edits and quantitative comparisons against BetaVAE and FactorVAE on FID/IoU, DMM, and DCM.","tokens_in":21275,"tokens_out":5891,"duration_ms":52743,"significance":"If the central claims held, this would be a valuable contribution: it offers a genuinely different, VSA-based route to disentangled representations that supports interpretable vector-level editing and proposes model-agnostic metrics that do not assume localist coordinates. The paper is also carefully structured around explicit research questions, and the qualitative exchange results on dSprites and CLEVR are compelling proof-of-concept. However, the evaluation instrument DCM is implemented inconsistently with its definition, and the 'disentanglement by construction' claim is stronger than the training objective and the paper's own qualitative results support. The significance of the contribution therefore depends on the extent to which these issues can be resolved.","major_comments":[{"comment":"Eq. (8) defines DCM as (1/|C|) Σ_{c∈C} | Σ_{Gi∈G} [Cl(Ŝ'_c) ≠ ŷ]_{Gi} − 1 |, which is a per-unit row sum over generative factors. However, the text and Figure 4 state that DCM should use a per-factor column sum, counting how many latent units affect each generative factor. The metric's stated definition, 'whether each generative factor is encoded by a single latent unit,' requires a column sum. As written, DCM measures the average number of factors changed by a unit, not compactness across units for a factor. Consequently, the DCM values in Tables 3 and 4, including the conclusions that ArSyD is more compact than BetaVAE and FactorVAE, do not support the claims. Please correct Eq. (8) to aggregate over units for each factor, recompute the affected tables, and state clearly which operationalization is used.","section":"Section 3.4, Eq. (8), Figure 4"},{"comment":"The abstract and Section 3.1 claim that 'disentanglement is achieved by construction,' but this is not established by the training procedure. In each weakly supervised pair, all factors except the exchanged one are identical between target and donor, so the exchange loss in Eq. (5) only requires V*_p to carry enough information to transfer factor p; it never penalizes V*_p for also encoding other factors, because those factors are constant in the pair and the decoder can ignore redundant information. The HDC binding in Eq. (1) makes the bound terms quasi-orthogonal, but it does not constrain V*_i itself to be independent of the other factors. The paper's own Figure 7 and accompanying text say that the 'Orientation' feature is 'strongly related' to the 'Shape' feature, which is direct evidence that the factor-alignment assumption can fail. The claim should be weakened to 'encouraged by the weak supervision objective,' or the authors should provide a direct analysis (e.g., probing or decoding each V*_i in isolation, or a full pairwise intervention matrix) demonstrating that each V*_i is factor-pure.","section":"Abstract; Section 3.1, Eqs. (3)–(4); Section 3.2–3.3, Eq. (5); Section 4, Figure 7"},{"comment":"RQ8 concludes that the value vectors V*_i represent separate properties because decoding a single vector bound to a placeholder does not reconstruct a complete image. This conclusion does not follow from the evidence: a vector that encodes a mixture of several factors, but is alone insufficient to generate a full image, would produce qualitatively the same incomplete reconstructions. To support the factor-purity interpretation, the authors should test interventions directly, e.g., by modifying V*_i and measuring which downstream factor classifications change, or by training linear probes on V*_i to see whether they predict only the intended factor.","section":"Section 4, RQ8, Figures 13–14"}],"minor_comments":[{"comment":"The caption says 'DCN metric' in the last sentence; this should be 'DCM metric.'","section":"Figure 4"},{"comment":"The text after Eq. (5) says '˜Od – a reconstructed target object'; this should be '˜Ot – a reconstructed target object,' and the donor terms should be labeled consistently.","section":"Eq. (5)"},{"comment":"Section 3.7 states that dSprites paired was trained for 600 epochs, whereas the caption of Table 3 says 200 epochs for dSprites paired; please reconcile these numbers.","section":"Section 3.7 and Table 3"},{"comment":"The metric classifiers are described as six ResNet-34 models fine-tuned on the CLEVR1 paired dataset, but the paper also reports DMM and DCM for dSprites paired; please specify which classifiers are used for dSprites and CelebA, and report their per-factor accuracies on the reconstruction sets.","section":"Section 3.8"},{"comment":"The term 'unit' is introduced in the metrics section, but Figure 4 and the surrounding text use 'DCN' and 'DCM' inconsistently; please unify the notation.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The reader's report and my own reading agree on the two load-bearing issues: the DCM equation does not implement the stated compactness definition, and the 'by construction' claim is not supported by the exchange loss. Both are fixable within the manuscript's scope, which is why I am not recommending rejection. I would want the authors to recompute all DCM numbers after correcting the equation and to either provide a direct factor-purity analysis or substantially weaken the central claim. There is no indication of any issue with the validity of the qualitative editing results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: ArSyD is a credible, clearly written attempt to ground hypervector symbolic representations in images, and the qualitative edits on dSprites and CLEVR mostly work. The main evaluation instrument, DCM, has a real inconsistency, and the abstract's 'by construction' claim overstates what the training actually guarantees. Neither issue is fatal, but both need fixing.\n\nWhat is genuinely new: the specific combination of weakly supervised feature exchange, attention over a fixed random codebook, and HRR-style binding to produce factor hypervectors, plus the DMM/DCM evaluation scheme that is dimension-agnostic. The paper explains the architecture well, and the ablations are decent: seed stability, latent dimension sweep, multiple simultaneous exchanges, and the compositional generalization split. The Slot Attention extension for multi-object scenes is a useful proof of concept, and the qualitative edits show controlled behavior on most factors.\n\nThe soft spots are in the evaluation and the central claim. First, DCM as written in Eq. (8) sums over generative factors for each latent unit (a row sum), but the text and Figure 4 say compactness should be a column sum per factor. So the reported DCM numbers do not measure what the definition says. That is a concrete bug, not a stylistic quibble, and it weakens the comparison against BetaVAE and FactorVAE.\n\nSecond, 'disentanglement by construction' is not supported. The exchange loss only requires the exchanged vector to transfer the factor that differs between the pair; it never penalizes that vector for also encoding other factors, because those are constant in the pair. The paper's own Figure 7 shows orientation remaining entangled with shape, which is direct evidence that the premise fails in practice. The claim should be softened to say the architecture encourages disentanglement, or the authors should add an explicit independence objective.\n\nThird, no code or data is provided, and the classifier accuracies used for DMM/DCM are low for spatial factors (Tables 5 and 6), so the metric values carry a lot of noise. Reproducibility is therefore limited.\n\nOverall, the core idea is plausible and worth engaging. The paper is written honestly, and the limitations section acknowledges several issues, but the metric bug and the overclaim need to be addressed before the results can be taken at face value.\n\nRecommendation: send it to peer review with a request for major revision. The architecture and metrics are novel enough to deserve referee time, but I would not trust the quantitative results until the DCM equation is corrected and the code is released.","headline":"ArSyD is a promising HDC-based disentanglement architecture with real novelty, but the DCM metric is misimplemented and the 'by construction' claim overreaches.","tokens_in":21829,"tokens_out":2730,"would_cite":false,"duration_ms":25568,"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":"ArSyD stores each image property as a full-size hypervector so editing an object is a vector swap.","keywords":["disentangled representations","hyperdimensional computing","vector symbolic architectures","weakly supervised learning","image editing","slot attention","disentanglement metrics","compositional generalization"],"falsifier":"On the dSprites dataset, take two images that differ only in orientation, swap only the orientation value vector in the latent representation, decode both edited images, and classify their shape with a pretrained shape classifier. If the shape prediction changes when only orientation is exchanged, the orientation vector carries shape information and the factor-level disentanglement claim fails; the paper's own Figure 7 already shows such a shape distortion for orientation.","tokens_in":1384,"feed_emoji":"🧩","tokens_out":1543,"duration_ms":57065,"temperature":0.7,"pith_summary":"ArSyD aims to make image generation factors separately editable by representing each generative factor (shape, color, size, position, and so on) as its own full-dimensional hypervector, rather than as one coordinate of a latent vector. The representation of an object is the superposition of these factor-value hypervectors, so changing a property means exchanging one hypervector for another. The paper claims this disentanglement is achieved by construction: the encoder and attention modules are trained only to reconstruct images from weakly supervised pairs that differ in one factor, with no extra assumptions about latent distributions. If the claim holds, controlled image editing becomes a simple, interpretable vector swap, and the same representation can be used in scenes with several objects by combining it with slot-based object discovery.","feed_headline":"Swap one hypervector to edit a single object property","feed_subtitle":"New model stores each generative factor as a full-size vector and learns from paired images with plain reconstruction loss.","key_machinery":"The load-bearing mechanism is attention over a frozen random codebook (item memory), used as a bridge between the encoder's localist feature vector and a distributed hypervector representation. For each generative factor $i$, a projection of the encoder output is matched by softmax attention against fixed seed hypervectors from the codebook, and the weighted sum of those seeds becomes the factor's value vector $V_i^*$. These value vectors are bound to factor hypervectors and bundled into the object representation, following Holographic Reduced Representation operations. The same encoder runs on donor and target images, and the feature-exchange module swaps one value vector before decoding.","core_discovery":"On the paper's own terms, the central discovery is that a disentangled representation can be built as a sum of learned value hypervectors, one per generative factor, instead of a vector whose individual coordinates each encode a factor. Each value vector $V_i^*$ is produced by an attention mechanism that selects a weighted combination of fixed random seed vectors from an item memory, and the object representation is $O = \\sum_i G_i \\circledast V_i^*$. The model is trained by swapping the value vector of one factor between two images that differ only in that factor and reconstructing both images with an MSE loss. The authors argue that because each generative factor has its own vector and the value vectors are grounded in the image through attention, disentanglement holds by construction, and editing reduces to exchanging the corresponding vector.","pith_inferences":["Editorial inference: if the value hypervectors are truly independent symbols, then latent vector arithmetic (for instance, adding the vector for one object property and subtracting another) should produce coherent composite edits, a test the paper does not run.","Editorial inference: the entanglement of orientation with shape visible in the paper's Figure 7 suggests that adding a sparsity or orthogonality penalty on the attention weights would make the factor vectors cleaner; the paper itself does not propose such a penalty.","Editorial inference: because DMM and DCM measure disentanglement through classifiers on reconstructed images, better decoders or stronger classifiers could change the measured scores even if the latent representation is unchanged, so the metrics should be read as bounded by the reconstruction and classification pipeline."],"forward_implications":["Editing an image property becomes a vector substitution, so the same edit operation is well-defined regardless of where in the latent space a factor lives.","The model can be combined with Slot Attention to edit a single object inside a multi-object scene, not just isolated objects.","Because disentanglement is claimed by construction, the approach avoids distributional assumptions and the loss-tuning typical of beta-VAE-style methods.","The proposed DMM and DCM metrics allow comparisons between localist and distributed representations by measuring changes in pixel-space classifications.","Reconstruction from a single factor vector shows that individual value vectors carry property-specific information, which supports interpretable editing."],"supporting_citations":[{"why":"Supplies the hyperdimensional computing foundations: quasi-orthogonal seed vectors and the bundling and binding operations used to compose factor vectors.","marker":"[23]"},{"why":"Supplies Holographic Reduced Representations, the specific real-valued vector algebra whose circular-convolution binding and bundling ArSyD uses.","marker":"[54]"},{"why":"Supplies the attention mechanism that bridges encoder features and the fixed codebook vectors in Equations 3 and 4.","marker":"[36]"},{"why":"Provides the theoretical grounding that weak supervision with paired examples differing by one factor enables learning of disentangled representations.","marker":"[37]"},{"why":"Provides the weakly supervised disentanglement training recipe without compromises that the paired-image exchange procedure builds on.","marker":"[38]"},{"why":"Supplies Slot Attention, the object-discovery module that lets ArSyD edit one object in a multi-object scene.","marker":"[39]"},{"why":"Provides the dSprites dataset and its generative factors, used to train and test ArSyD for single-object images.","marker":"[40]"},{"why":"Provides the CLEVR dataset and its modified paired and multi-object splits used in the experiments.","marker":"[41]"},{"why":"Supplies the modularity and compactness notions of disentanglement that the DMM and DCM metrics formalize.","marker":"[66]"}],"fun_headline_variants":["Edit images by swapping hypervectors","Disentangled images via vector superposition","ArSyD: Disentanglement by construction","Swap a vector to change one image property","Weak supervision yields symbolic image control"],"cache_read_input_tokens":23936,"weakest_assumption_plain":"The paper's claim depends on the assumption that training only the attention weights, with nothing but reconstruction error on image pairs that differ in one factor, forces each selected hypervector to align with exactly one true generative factor and to stay independent of the others.","fun_headline_variants_meta":{"raw":{"variants":["Edit images by swapping hypervectors","Disentangled images via vector superposition","ArSyD: Disentanglement by construction","Swap a vector to change one image property","Weak supervision yields symbolic image control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1455,"prompt_tokens":971,"completion_tokens":484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":422}},"tokens_in":587,"tokens_out":484,"duration_ms":5038,"temperature":1.0,"reasoning_tokens":422,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:25:17.568409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the dSprites dataset, take two images that differ only in orientation, swap only the orientation value vector in the latent representation, decode both edited images, and classify their shape with a pretrained shape classifier. If the shape prediction changes when only orientation is exchanged, the orientation vector carries shape information and the factor-level disentanglement claim fails; the paper's own Figure 7 already shows such a shape distortion for orientation.","supporting_citations":[{"cited_title":"Hyperdimensional computing: An introduction to com- puting in distributed representation with high-dimensional random vec- tors","cited_arxiv_id":null,"evidence_quote":"Supplies the hyperdimensional computing foundations: quasi-orthogonal seed vectors and the bundling and binding operations used to compose factor vectors."},{"cited_title":"Holographic reduced representations: Convolution alge- bra for compositional distributed representations","cited_arxiv_id":null,"evidence_quote":"Supplies Holographic Reduced Representations, the specific real-valued vector algebra whose circular-convolution binding and bundling ArSyD uses."},{"cited_title":"Weakly-supervised disentan- glement without compromises","cited_arxiv_id":null,"evidence_quote":"Provides the weakly supervised disentanglement training recipe without compromises that the paired-image exchange procedure builds on."},{"cited_title":"dsprites: Disentanglement testing sprites dataset","cited_arxiv_id":null,"evidence_quote":"Provides the dSprites dataset and its generative factors, used to train and test ArSyD for single-object images."},{"cited_title":"Towards a defi- nition of disentangled representations, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the modularity and compactness notions of disentanglement that the DMM and DCM metrics formalize."}],"review_version":1}