{"id":"de6f45e9-bb46-467f-956c-cd0df4fdc1fc","arxiv_id":"2512.08854","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For compositional generalization under polynomial-interaction generators, constraining a decoder is tractable while constraining the inverse encoder is generally infeasible, so inverting a generative decoder is a more data-efficient route.","lead":"This paper argues that, under a compositional latent-variable model, guaranteeing out-of-domain generalization is practical for generative models (learn a decoder, then invert it) but not for encoder-only models. It tests the idea on synthetic images and finds that decoder-inversion methods generalize better without extra real data.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 and Appendix A.2 do not establish that all encoder inductive biases are infeasible: pointwise derivative freedom and addition-closure universality do not cover the non-additively-closed encoder classes used in practice, so the central 'cannot' claim overreaches.","rationale":"The reader's verdict correctly identifies the paper's substantive but overreaching claim. I focused on the infeasibility step rather than the F_int assumption because the F_int scope is explicitly acknowledged in Sec. 7, whereas the infeasibility step is presented as a proven fact but is not fully supported. Theorem 3.2's pointwise freedom of Dg and D^2g at a single point does not rule out global architectural or regularizer constraints. The density results in Appendix A.2 require closure under addition and coordinate projections, which do not hold for the fixed neural network families used in the experiments; hence the argument leaves open the possibility of a well-designed encoder class with a useful inductive bias. The paper's own PUG-Object results (Fig. 6C) show non-generative encoders succeed when G_int has extra structure (n=0), indicating the asymmetry is not universal. A concrete derivation of higher-order constraints would test whether the pointwise freedom extends, or whether a usable constraint exists. If such a constraint exists, the central 'cannot' claim would need further tempering; if not, the infeasibility conclusion would gain support. The F_int premise remains a scope limitation, but it is the infeasibility gap that most directly threatens the headline claim as stated.","tokens_in":22956,"tokens_out":13666,"duration_ms":141174,"concrete_test":"Derive the third-derivative analogue of Eq. (A.2) (differentiate g∘f(z)=z three times) for f ∈ F_int with n=2, and determine whether, in the regime d_x ≥ d_z^3, the third-derivative tensor of g, projected onto the tangent space of the data manifold, satisfies any non-trivial algebraic identity that does not involve unobserved OOD geometry. If such an identity exists, enforce it as a regularizer on the from-scratch ViT+slot encoder using ID data only and measure OOD accuracy on PUG-Background; a substantial improvement would directly contradict Sec. 3.1's 'generally infeasible' conclusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that non-generative methods cannot achieve compositional generalization through inductive bias—rests on the assertion (Sec. 3.1, Takeaways) that constraining an encoder to G_int is generally infeasible. But Theorem 3.2 only shows that, for d_x ≥ d_z^3, the first- and second-order derivatives of a left inverse g at a single point x0 can be arbitrary (for almost every Jacobian). This rules out local pointwise derivative regularizers, but not architectural inductive biases that restrict the global function class, nor non-local regularizers (e.g., sparsity in a learned basis, invertibility, or smoothness penalties). The Appendix A.2 attempt to cover architecture constraints (Theorem A.8, Corollary A.11) assumes G_enc is closed under addition and, in the d_x > d_z case, under coordinate projections. Standard neural encoder families—including the ViT+slot-encoder used in Sec. 5—are not closed under either operation; the set of functions representable by a fixed architecture is not a vector space. Hence the density/universality results do not apply to the encoder classes of practical interest. The infeasibility conclusion is therefore a conjecture about unexamined architectures, and the paper's own Fig. 6C shows non-generative encoders can succeed on n=0 substructures within F_int. Even granting the F_int premise, the proof does not support the unconditional 'cannot be achieved.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether compositional generalization, as a hallmark of human data efficiency, can be guaranteed in non-generative (encoder-only) perceptual systems or only in generative systems that invert a learned decoder. Under a compositional latent-variable model with generators in the polynomial interaction class F_int (Brady et al., 2025), the authors argue that: (i) guaranteeing OOD generalization requires constraining the decoder to F_int or the encoder to its inverse class G_int; (ii) constraining a decoder is straightforward via architecture or derivative regularization (e.g., Eq. 3.2), whereas constraining an encoder is claimed to be generally infeasible because it would require knowledge of unobserved OOD manifold geometry; (iii) this asymmetry is supported by Theorem 3.2 showing local freedom of a left inverse's Jacobian and Hessian when d_x ≥ d_z^3, and by Appendix A.2 density results under closure assumptions on the encoder class; and (iv) experiments on PUG datasets show non-generative methods need large-scale pretraining or favorable structure (PUG-Object), while generative methods improve OOD accuracy via gradient search and replay. The central reading is that compositional generalization cannot be achieved in non-generative methods through inductive bias but can be achieved in generative methods by constraining and inverting a decoder.","tokens_in":23261,"tokens_out":4668,"duration_ms":52537,"significance":"The question addressed is important and timely: it connects a long-standing cognitive-science hypothesis about perception-as-inversion with modern object-centric representation learning. If the central asymmetry were fully established, the paper would provide a principled argument against encoder-only inductive biases for compositional generalization and would support generative approaches such as decoder inversion, gradient search, and replay. The paper has notable strengths: it formalizes the problem within the F_int/G_int framework; it gives concrete derivative-based regularizers for decoders; it provides an explicit theorem (Theorem 3.2) on local derivative freedom; and it tests the claims on a controlled, photorealistic benchmark with multiple pretrained encoders and both generative and non-generative training regimes. The empirical n=0 control (PUG-Object) is a useful falsification probe. However, the theoretical results as stated do not support the paper's strongest claim of general infeasibility of encoder inductive biases, and there is a mismatch between the theorem's dimension regime and the experiments. With substantial revision of the claims and additional theory or exp","major_comments":[{"comment":"The experimental setup does not satisfy the main theorem's dimension assumption. Theorem 3.2 requires d_x ≥ d_z^3, but the models use 3 slots of dimension 64, so d_z = 192, and images are 224×224×3, so d_x = 150,528. Since 192^3 = 7,077,888 > 150,528, the theorem does not apply to the experiments. The remark after Lemma A.5 suggests the expected threshold is d_x ≥ d_z(d_z+1)/2, which would be satisfied, but that is a remark, not a proved theorem. Please either prove the result under the weaker threshold and use it to ground the experiments, or clearly state that the experiments probe a regime beyond the proved range. As written, the empirical asymmetry is not covered by the formal result.","section":"Sec. 5 vs Theorem 3.2"},{"comment":"Theorem 3.2 establishes local freedom of Dg(x0) and D^2 g_l(x0) at a single point x0, for almost every Jacobian A, under d_x ≥ d_z^3. This rules out pointwise first- and second-order derivative regularizers applied independently at each point, but it does not rule out architectural inductive biases that constrain the global function class, nor non-local regularizers (e.g., invertibility penalties, sparsity in a learned basis, smoothness over neighborhoods, or multi-point constraints). The 'Takeaways' paragraph in Sec. 3.1 and the abstract's 'generally infeasible using regularization or architectural constraints' is therefore an extrapolation. The manuscript should either narrow the claim to local, pointwise, derivative-based inductive biases, or provide an additional theorem covering the broader classes of encoder constraints that are claimed to be infeasible.","section":"Sec. 3.1, Theorem 3.2 and Takeaways"},{"comment":"The architecture-level results assume properties that standard neural encoders do not satisfy. Theorem A.8 and Corollary A.11 require G_enc to be closed under addition and, for d_x > d_z, under coordinate projections. The function class of a fixed architecture such as the ViT+slot encoder used in Sec. 5 is not closed under addition: the sum of two such functions is generally not representable by the same fixed architecture. Coordinate projection closure is also not a natural property of attention-based encoders. Hence the density/universality results do not apply to the encoder classes of practical interest. The conclusion that 'it is challenging to find natural function classes G_enc' is plausible, but it is not proved for the architectures used in the paper. Please state the closure assumptions explicitly in the main text and explain which practical classes satisfy them, or reframe the","section":"Appendix A.2"},{"comment":"The central claim is conditional on the ground-truth generator belonging to F_int, the polynomial-interaction class in Eq. (2.7). The manuscript states 'we assume that ground-truth generators f belong to F_int' without independent justification, and Sec. 7 acknowledges that results may fail outside F_int. Since Eq. (2.5) is itself imported from prior work (Brady et al., 2025), the entire asymmetry between decoder and encoder biases is a conditional theorem. This is acceptable as a contribution, but the abstract and title ('Generation is Required...') present the conclusion as unconditional. Please state the F_int condition in the abstract, and discuss the scope of the claim. The n=0 case in Fig. 6C is especially relevant: within the paper's own framework, non-generative encoders achieve near-perfect OOD generalization on PUG-Object, so the blanket statement that non-generative methods ca","section":"Sec. 2, Eq. (2.7) and Sec. 7, Limitations"}],"minor_comments":[{"comment":"The title in the preprint header is typeset as 'GENERATION ISREQUIRED FORDATA-EFFICIENT PERCEPTION'; fix the spacing.","section":"Title"},{"comment":"The condition '|D_zk fi(z)| · |D_zl fi(z)| = 0' is written without norms or a formal definition of fi as a coordinate function. Please define the notation precisely or refer to the original definition in Brady et al. (2023).","section":"Sec. 3.1, n=0 paragraph"},{"comment":"The statement that replay 'cannot be applied' on PUG-Texture because slots capture objects/backgrounds and 'cannot be trivially recomposed' is unclear. It would help to explain why the slot recombination used for PUG-Background does not transfer to animal–texture composition, e.g., because texture is a property of the object slot rather than an independent slot.","section":"Sec. 5.2, PUG-Texture replay"},{"comment":"The name appears misspelled: 'Thaddäus Wiedmer' should likely be 'Thaddäus Wiedemer' (matching reference [Wiedemer et al., 2024]).","section":"Acknowledgments"},{"comment":"The sets Z_S, Z_ID, Z_OOD are used interchangeably with the latent space Z; please define Z as the image of the generator or clarify the domain/codomain conventions so that Eq. (2.4) is unambiguous.","section":"Sec. 2, notation"}],"recommendation":"major_revision","confidential_remarks":"The paper's conditional theory is interesting and the empirical study is substantial, but the current version overclaims the infeasibility of encoder inductive biases. The dimension mismatch between Theorem 3.2 and the experiments, the local-only nature of the theorem, and the restrictive closure assumptions in Appendix A.2 are all load-bearing for the central 'cannot' claim. I would like to see the main claims narrowed to the provable statements, or the theory extended to cover the practical encoder classes and the experimental regime. If the authors can do that, the paper would be a valuable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a genuine contribution with an overstated central claim. The novel part is the local structure analysis of the inverse class G_int — Lemma 3.1 and Theorem 3.2 show that in high ambient dimension, pointwise first and second derivatives of a left inverse are essentially unconstrained, while the decoder-side constraint (Eq. 3.1/3.2) is simple and axis-aligned. That asymmetry is real and worth writing down. The empirical comparison on PUG is consistent with the story: non-generative encoders often fail OOD unless heavily pretrained, and decoder search/replay helps. The n=0 control is a nice check, and the paper is honest enough to show non-generative methods succeeding there.\n\nWhat the paper does well: it states its assumptions openly, including the F_int premise, and the proofs in the appendix look coherent to me — standard calculus and linear algebra, not hand-waving. The empirical setup is a reasonable first test even if limited to synthetic data.\n\nWhere it overreaches: the jump from Theorem 3.2 to \"generally infeasible\" for encoder inductive biases. The theorem rules out local pointwise derivative regularizers; it does not rule out global architectural biases or non-local regularizers like smoothness or invertibility penalties. The Appendix A.2 argument for architectural constraints assumes G_enc is closed under addition and coordinate projections — a ViT plus slot encoder is neither. So the strong \"cannot be achieved\" language in the title and intro is not actually proven for practical encoder families. What the paper establishes is better stated as: no general principled way to enforce G_int on encoders is known, and the known local constraints are manifold-dependent. That weaker claim is still interesting. Also, the F_int assumption is the load-bearing premise, and the limitations section honestly flags that the results may fail outside it; the headline goes beyond what the paper itself is willing to defend.\n\nEmpirically, selecting the best configuration per model and omitting error bars makes the comparison weaker than it could be — that's fixable. The self-cited prior theorem (Brady et al. 2025) is the foundation, so independent scrutiny of that result would raise confidence, but self-citation by itself is not the issue.\n\nBottom line: this deserves a serious referee and likely publication after revision, with claims tempered to \"no principled inductive bias known\" rather than \"cannot be achieved,\" and with empirical reporting cleaned up. I'd bring it to reading group — it will generate good debate — and I'd cite it as the formal statement of the decoder/encoder asymmetry, not as a proof that generation is required.","headline":"Worth engaging seriously, but the paper proves less than its 'cannot be achieved' language claims; the decoder/encoder asymmetry is real and new, while the infeasibility theorem is narrower than advertised.","tokens_in":23826,"tokens_out":2492,"would_cite":true,"duration_ms":26783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Generation is required for data-efficient perception: only decoder-inversion guarantees compositional generalization.","keywords":["compositional generalization","out-of-domain generalization","generative models","encoder-only learning","inductive bias","identifiability","object-centric representations","data efficiency"],"falsifier":"Train an encoder-only model from scratch on PUG-Background (no large-scale pretraining, no decoder inversion, no added supervision) and reach near-perfect OOD accuracy on unseen animal–background pairs; the paper's position predicts this will fail except in the non-interacting n=0 case. Equivalently, exhibit any encoder inductive bias that provably keeps an encoder in G_int on OOD regions without access to X_OOD.","tokens_in":22763,"feed_emoji":"👁️","tokens_out":7875,"duration_ms":78785,"temperature":0.7,"pith_summary":"The paper tries to settle whether machine vision needs generative models to be data-efficient, and answers yes for compositional generalization. Assuming images are generated by a compositional function in the class F_int—slot-wise maps plus bounded polynomial interactions between slots—it shows that the inductive bias needed to generalize to unseen combinations of concepts is simple for a decoder to enforce but generally impossible for an encoder to enforce, because encoder constraints would require knowledge of the unseen image manifold. The theoretical core is that, when images live in a much higher-dimensional space than the latents, the local derivatives of the inverse generator are unconstrained: any Jacobian and Hessian can occur (Theorem 3.2), so no local regularizer or architecture can keep an encoder inside the good class. On photorealistic synthetic data, encoder-only models fail on out-of-domain combinations unless trained with large-scale pretraining, while generative models that constrain a decoder and invert it—by gradient search or generative replay—gain out-of-domain accuracy with no extra data. If the paper is right, the data efficiency of human perception points toward generative representations, and the scaling of non-generative models is buying generalization that structure could provide instead.","feed_headline":"Yes: data-efficient vision needs generation","feed_subtitle":"New theory and experiments show decoder inversion generalizes to unseen concept combinations without extra data.","key_machinery":"The load-bearing object is the function class F_int (Eq. 2.7): diffeomorphisms that write an image as a sum of slot-wise functions plus polynomial cross-slot interaction terms of degree at most n. Its defining property—that cross-slot derivative tensors of order n+1 vanish, e.g. D_{z_k}D_{z_l} f = 0 for n=1—can be imposed on a decoder by a block-diagonal architecture or the regularizer in Eq. 3.2, and this is what guarantees the decoder identifies the ground-truth generator out-of-domain. The inverse class G_int lacks such global structure: Theorem 3.2 shows that in ambient image dimension dx ≥ dz^3, the Jacobian and Hessian of an inverse at a point can be arbitrary (up to a measure-zero set","core_discovery":"The central claim is that under the compositional data-generating process used here, compositional generalization is achievable with principled inductive biases for generative methods and not for non-generative ones. A generative method learns a decoder that approximates the ground-truth generator; because the generator class F_int is characterized by block-diagonal derivative tensors (e.g., a zero cross-slot Hessian when interactions are first-order), the decoder can be constrained to F_int by architecture or a simple regularizer, and out-of-domain latents are then recovered by inverting this decoder. A non-generative method instead learns the inverse generator directly; the paper shows tha","pith_inferences":["Editorial extension: the entire asymmetry is conditional on the assumption that natural images are generated by F_int; if real-world generation involves higher-order or non-polynomial interactions outside this class, the feasibility gap between encoders and decoders could shrink or reverse—the paper acknowledges this limitation.","Editorial extension: Theorem 3.2 and the density result in the appendix suggest a concrete test for future work: search for encoder architectures that are expressive enough to invert F_int but are not closed under addition and not dense in continuous functions; if such a class exists, it might provide the principled encoder inductive bias the paper argues is impossible.","Editorial extension: the two inversion strategies are complementary—replay works when slots can be recombined into novel scenes, search works when they cannot—so a model that routes between them (or uses search to generate replay data) may scale the approach toward natural images, where slots are less well-separated."],"forward_implications":["If the central claim is correct, encoder-only (non-generative) vision models cannot be made to generalize compositionally by better inductive biases alone; improvements will continue to demand larger pretraining corpora or added supervision.","Generative methods that constrain a decoder to F_int and invert it can improve out-of-domain accuracy on unseen concept combinations without any additional data, relying only on gradient search and generative replay.","The n=0 special case (concepts that never interact, such as non-occluding objects) is the predicted exception: encoder-only methods should generalize OOD here, and the experiments confirm near-perfect OOD accuracy across all tested encoders.","The theory provides a formal grounding for the long-standing conjecture that generalization is easier in the causal direction (generating an image from latents) than in the anti-causal direction (inferring latents from an image).","Human perceptual data efficiency could be explained by the brain solving an inverse-generation problem rather than learning a direct feedforward encoder, which is why a child can recognize a penguin in a desert after seeing each in other settings."],"fun_headline_variants":["Decoder inversion is the key to data-efficient vision","Generative models generalize, encoders don't without pretraining","For compositional generalization, learn to generate","Why data-efficient vision needs generative learning","Encoders need data, decoders don't: new theory"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument assumes the ground-truth visual generator belongs to the function class F_int (slot-wise plus polynomial-interaction diffeomorphisms); if real images are generated by a broader class, the OOD identifiability guarantee and the decoder/encoder asymmetry can fail.","fun_headline_variants_meta":{"raw":{"variants":["Decoder inversion is the key to data-efficient vision","Generative models generalize, encoders don't without pretraining","For compositional generalization, learn to generate","Why data-efficient vision needs generative learning","Encoders need data, decoders don't: new theory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":2813,"prompt_tokens":734,"completion_tokens":2079,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2020}},"tokens_in":478,"tokens_out":2079,"duration_ms":15404,"temperature":1.0,"reasoning_tokens":2020,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:33:38.387306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an encoder-only model from scratch on PUG-Background (no large-scale pretraining, no decoder inversion, no added supervision) and reach near-perfect OOD accuracy on unseen animal–background pairs; the paper's position predicts this will fail except in the non-interacting n=0 case. Equivalently, exhibit any encoder inductive bias that provably keeps an encoder in G_int on OOD regions without access to X_OOD.","supporting_citations":[],"review_version":1}