{"id":"dda84dec-fa7c-47c9-af05-a5cdf8f4932d","arxiv_id":"2411.11162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RPN 2 adds interdependence functions to the Reconciled Polynomial Network and claims that CNN, RNN, GNN, and Transformer differ only in which interdependence function they use.","lead":"This paper extends a polynomial-based neural framework with new 'interdependence functions' that model relationships among data points and features, claiming to represent CNN, RNN, GNN, and Transformer as special cases. A smart generalist would read it because a single canonical framework for the major architecture families could reshape how AI models are designed, compared, and combined.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transformer unification in Section 8 is asserted, not demonstrated: exact representation of softmax attention, residual, and layer norm in the RPN 2 form is unverified, so the 'fundamental distinction is interdependence' claim is load-bearing on missing equivalence proofs.","rationale":"The paper's strongest contribution is a concrete, implemented library of interdependence functions with careful geometric and topological constructions, and the RPN 2 framework is a genuinely broad architectural umbrella. The central claim, however, is a representational universal: that dominant backbones are special cases differing mainly in ξ. The reader's weakest assumption captures the right risk—exact representability—and I agree it is the load-bearing point. My read of the available text is that Section 8 sketches constructions but does not supply formal equivalence proofs or numerical verification of exactness. The concrete test above would settle whether the Transformer mapping is exact or merely suggestive. If exact, the unification claim is far stronger and interesting; if not, the paper should be reframed as an approximate or common-structure framework, which is still useful but not the advertised unification. Since the verification is missing, the reader's CONDITIONAL verdict is appropriate, and I do not recommend changing it.","tokens_in":63902,"tokens_out":7599,"duration_ms":81104,"concrete_test":"Instantiate, using the released TINY BIG v0.2.0 code, an RPN 2 model with the interdependence functions and component choices prescribed in Section 8.4 for a Transformer, and a standard PyTorch nn.TransformerEncoder with identical weights (including multi-head Q/K/V projections, residual, layer norm, and MLP). Feed identical batches and compare hidden states after each block to 1e-5 absolute tolerance. If any block diverges, identify whether the unmodeled part is residual, norm, multi-head combination, or MLP, and determine whether adding it requires changing κ or ψ rather than only ξ. Passing this test would concretely validate the exact-unification claim; failing it would require the paper to relax the claim to approximate or to extend the canonical form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that CNN, RNN, GNN, and Transformer are unified by RPN 2 and that their differences 'primarily stem from' the interdependence function definition (Abstract; Section 8). This requires that the exact forward pass of each backbone be reproducible by composing the RPN 2 components—data transformation κ, interdependence ξ, parameter reconciliation ψ, remainder π—with only ξ varying across backbones. The available text provides constructive mappings (e.g., grid, chain, and graph matrices for CNN, RNN, and GNN) but does not prove that a standard Transformer encoder, with multi-head softmax attention, scaled dot products, residual connections, and layer normalization, can be written exactly in the form of Eq. (28) without placing the entire attention nonlinearity inside an opaque ξ or changing κ and ψ per backbone. If the Transformer mapping requires a backbone-specific κ (e.g., value projection) and a backbone-specific remainder (residual connection), then the claim that the distinction is 'primarily' the interdependence function is weakened: it becomes an assertion that all backbones contain some common nonlinear and linear components, which is not the advertised unification. The provided text does not include formal equivalence statements for Section 8; the burden falls on those derivations, and the issue is representational exactness, not empirical performance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces RPN 2, an extension of the Reconciled Polynomial Network. The model represents a function as g(x|w) = ⟨κξ(x), ψ(w)⟩ + π(x), where κξ composes a data transformation κ with new attribute- and instance-level interdependence functions ξ that produce interdependence matrices applied to the input batch. The paper claims that this canonical form unifies CNN, RNN, GNN, and Transformer backbones and that the differences among these backbones primarily stem from how interdependence functions are defined. It also contributes a large catalog of interdependence functions, data compression and fusion functions, complexity estimates, a software toolkit update, and experiments on image, language, time-series, and graph benchmarks.","tokens_in":64176,"tokens_out":3461,"duration_ms":38123,"significance":"If the exact unification claim held, this would be a valuable conceptual contribution with practical implications for architecture design, and the open-source TINY BIG v0.2.0 toolkit is a concrete asset. The taxonomy of interdependence functions and the constructive matrix mappings for grid, chain, and graph structures are presented in a clear and useful way. However, the Transformer equivalence is asserted rather than demonstrated in the reviewed text, and the empirical evidence needed to support the performance claims is not present in the material provided. The central advertised novelty is therefore not yet established.","major_comments":[{"comment":"The unification of Transformers is asserted but not demonstrated. A standard Transformer encoder combines multi-head scaled dot-product softmax attention, per-head value and output projections, residual connections, and layer normalization. The RPN 2 form in Eq. (28) contains an interdependence matrix ξ, a transformation κ, a parameter reconciliation ψ, and a remainder π, but no exact derivation is given that rewrites the full Transformer forward pass in this form without placing the softmax nonlinearity and row normalization inside an opaque ξ and without changing κ and π per backbone. If softmax is absorbed into ξ, then ξ is no longer merely an interdependence matrix, and the claim that backbones differ primarily in their interdependence functions needs a precise and formal statement. Section 8.4 must provide explicit equivalence statements for all components of the Transformer encoder, or the central claim must be weakened.","section":"Section 8.4, Eq. (28)"},{"comment":"The claim that 'the fundamental distinctions among these backbone models primarily stem from their diverse approaches to defining the interdependence functions' requires that the other components of the canonical form be held fixed or shown to be common. The paper does not establish this. The constructive mappings for CNN, RNN, and GNN use backbone-specific ingredients beyond ξ: patch geometry and compression for CNNs, temporal sharing and remainder terms for RNNs, and graph adjacency plus normalization for GNNs. If κ, ψ, or π vary between backbones, the distinctions are not primarily due to interdependence functions alone. A formal statement of what is fixed across backbones and what is allowed to vary is needed before the advertised conceptual conclusion is supported.","section":"Section 8, Eq. (24)"},{"comment":"The empirical claims in the abstract and introduction cannot be verified from the reviewed text. The abstract says RPN 2 'significantly outperforms' the previous RPN and that RPN 2-based models 'enhance existing backbones,' while Section 1 says RPN 2 'achieves performance comparable to leading models like CNNs, RNNs, and GCNs.' The reviewed portions do not include the experimental setup, result tables, baselines, standard deviations, or ablations needed to assess either claim. These materials must be included, and the apparent inconsistency between 'significantly outperforms' and 'comparable' should be resolved.","section":"Section 9 and Abstract"}],"minor_comments":[{"comment":"There is a grammatical error near the end of the section: 'the codomain dimensions m′ and b′ must are explicitly specified' should read 'must be explicitly specified.'","section":"Section 4.2"},{"comment":"Entries 7 and 8 in the data expansion list are both labeled 'Lucas Expansion'; one of them appears to be mislabeled and should be corrected.","section":"Figure 19"},{"comment":"The notation in Eq. (30) for the multi-head and multi-channel fusion is difficult to parse because the head and channel indices H, C are introduced without a fully explicit indexing convention; a cleaner definition of the fusion argument would improve readability.","section":"Eq. (30)"}],"recommendation":"major_revision","confidential_remarks":"The paper is partly a large catalogue of component functions, which is useful, but the advertised intellectual merit is the exact unification of Transformer with CNN, RNN, and GNN. The missing Transformer equivalence proof and the unverifiable empirical section are the main obstacles. The revision should either provide rigorous equivalence statements for Section 8.4 or explicitly reframe the contribution as an approximation/unification-up-to-components rather than exact unification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a real framework paper, not vapor, but the headline claim is stronger than the text supports. RPN 2 adds interdependence functions to the earlier RPN decomposition and gives a genuinely broad menu of structural interdependence matrices—grid patches with cuboid/cylinder/sphere packing, chain/adjacency/PageRank forms, bilinear and low-rank variants. The writing is concrete and the equations are mostly consistent. The toolkit exists and is public, which is real evidence.\n\nWhat is new: treating the interdependence function as a first-class architectural component and showing that CNN, RNN, and GNN can be written as instances of Eq. (28) with different ξ. That is a useful way to talk about architecture families, and the grid/chain/graph constructions are explicit enough to check.\n\nThe soft spot is exactly where the stress-test puts it. Section 8.4's Transformer unification is asserted, not derived. Standard multi-head softmax attention, residual connections, and layer norm are not shown to fit exactly into κ/ψ/π with only ξ varying. If the Transformer mapping needs a per-backbone κ or a remainder that carries the attention nonlinearity, the advertised conclusion—that backbones primarily differ in interdependence functions—is much weaker. This is a representational exactness issue, not an empirical one, and it is load-bearing. The paper also has a small internal contradiction: the abstract says RPN 2 significantly outperforms baselines, while the introduction says performance is comparable on multimodal benchmarks. That should be fixed.\n\nMinor: the framework is built by construction to make the divergence claim true, so the insight is more organizational than explanatory until the Transformer case is actually worked out. The empirical section is not fully described here; I would want to see dataset splits, exact baselines, and whether the code reproduces the headline numbers.\n\nWho is this for: anyone working on unified architecture theory or principled hybrids. It deserves a serious referee, but the referee should demand the missing equivalence proofs and a consistent empirical summary. I would cite it as a framework plus a public toolkit, not as a proven unification yet.","headline":"A real framework paper with a public toolkit, but the Transformer-unification claim is asserted rather than proven, and the empirical summary overreaches; worth refereeing with the equivalence proofs demanded.","tokens_in":64704,"tokens_out":2357,"would_cite":true,"duration_ms":25576,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that CNN, RNN, GNN, and Transformer architectures differ only in their interdependence functions within one shared RPN 2 model.","keywords":["Reconciled Polynomial Network","Interdependence Function","Unified Model Architecture","Multi-modal Learning","Function Learning","Data Interdependence","Backbone Models","Neural Architecture Unification"],"falsifier":"Take a standard transformer block with softmax attention, layer normalization, and a residual connection, and transcribe its forward pass using only functions in the paper's catalogue. If any operation requires a function outside that catalogue, or if the residual-plus-normalization interaction cannot be represented through the interdependence matrices, then the claimed unification is approximate rather than exact.","tokens_in":63659,"feed_emoji":"🔗","tokens_out":7317,"duration_ms":65108,"temperature":0.7,"pith_summary":"This paper proposes an extension of the Reconciled Polynomial Network (RPN) called RPN 2, whose central claim is that the dominant neural network families—CNNs, RNNs, GNNs, and Transformers—are all special cases of one canonical architecture once the model is allowed to model interdependence among data instances and attributes. The single new component, the interdependence function, computes a sparse matrix of pairwise relationships, and the paper argues that the real difference between backbones is just how that matrix is defined: local patches for convolution, chains for recurrence, graph adjacency for message passing, and bilinear or attention-style forms for transformers. If the claim is right, backbone design stops being a choice between competing model families and becomes a menu of interdependence functions that can be mixed and fused. The paper also reports experiments where RPN 2 models match or exceed standard backbones on image, language, time-series, and graph benchmarks.","feed_headline":"One formula reproduces CNN, RNN, GNN, and Transformer backbones","feed_subtitle":"If correct, architecture choice reduces to picking one tunable interdependence function inside a single model.","key_machinery":"The carrying object is the interdependence matrix: a typically sparse matrix $\\mathbf{A}$ computed from the input batch and optional structure, applied to the data before or after transformation as $\\mathbf{X}\\mathbf{A}_{\\xi_a}$ and $\\mathbf{A}_{\\xi_i}^{\\top}(\\cdot)$. The paper defines a library of interdependence functions—constant and identity, statistical and numerical kernels, parameterized bilinear forms, geometric grid patches (cuboid, cylinder, sphere) with packing schemes, chain-based unidirectional, bidirectional, and multi-hop functions, graph adjacency and PageRank functions, and hybrid fusions—and claims that these functions supply the locality, sequential, structural, and attention mechanisms that distinguish existing backbones. Everything else in the architecture, including data expansion and compression, parameter reconciliation, remainder, and fusion, is shared across families.","core_discovery":"RPN 2 retains the original RPN decomposition $g(\\mathbf{x}|w) = \\langle \\kappa_\\xi(\\mathbf{x}), \\psi(w)\\rangle + \\pi(\\mathbf{x})$, but replaces the plain data expansion with a data interdependent transformation $\\kappa_\\xi(\\mathbf{X}) = \\mathbf{A}_{\\xi_i}^{\\top}\\kappa(\\mathbf{X}\\mathbf{A}_{\\xi_a})$, where $\\mathbf{A}_{\\xi_a}$ and $\\mathbf{A}_{\\xi_i}$ are attribute and instance interdependence matrices returned by new interdependence functions. The paper's discovery claim is that every major backbone can be written in this form, and that the distinctions among CNN, RNN, GNN, and Transformer are explained by which interdependence function is used rather than by any deeper architectural difference. It substantiates this with explicit reconstruction recipes in Section 8 for each family, a catalogue of interdependence functions including grid patches, chains with multi-hop powers and matrix inverses or exponentials, graph adjacency and PageRank, and parameterized bilinear forms, plus a hybrid construction that reproduces attention-like behavior.","pith_inferences":["Beyond what the paper shows, the unification claim implies a concrete architecture-search agenda: treat the interdependence function class as the search space, with shared transformation and reconciliation components fixed, and measure whether such search finds backbones that beat hand-designed CNN and Transformer baselines.","The chain and graph functions that sum powers of an adjacency-like matrix connect RPN 2 to graph-diffusion and PageRank ideas; learning the hop weights instead of fixing them would be a natural next step the paper motivates but does not implement.","A shared RPN 2 model with per-modality interdependence functions could serve as a multimodal backbone without modality-specific heads; the paper discusses multimodal data but evaluates datasets one modality at a time.","The paper's exact-equality claim should be tested by reimplementing a standard transformer as RPN 2 and checking forward-pass equality up to floating-point precision; until that check is run, the statement that these backbones differ only in interdependence function is a representation claim rather than a verified equivalence."],"forward_implications":["Choosing the identity interdependence function for both instances and attributes recovers the original RPN and, with it, the models already unified by RPN, including probabilistic graphical models, kernel SVMs, MLPs, and KANs.","Using grid-patch interdependence reproduces CNN-style local receptive fields; varying patch shape and packing density gives a family of convolutional variants.","Using chain interdependence functions reproduces RNN-style sequential processing, with multi-hop powers and $(I-A)^{-1}$ or $\\exp(A)$ as computationally cheaper ways to propagate information along long sequences.","Using graph adjacency, PageRank, or hybrid bilinear-plus-graph interdependence yields GNN, GAT-like, and Transformer-like behavior, so attention can be understood as one interdependence function choice.","New backbones can be designed by selecting or fusing interdependence functions, and the reported experiments show such RPN 2 variants matching or outperforming standard backbones on image, language, time-series, and graph tasks."],"supporting_citations":[{"why":"Defines the original RPN three-component architecture that RPN 2 extends, and provides the parameter-reconciliation techniques reused to generate interdependence matrices.","marker":"[89]"},{"why":"Graph Attention Network is the precedent the paper uses to show that a hybrid bilinear-plus-graph interdependence function yields attention-like behavior.","marker":"[76]"},{"why":"Supplies the PageRank convergence-matrix form used as a graph interdependence function and the graph-BERT attention behavior that the hybrid graph function mirrors.","marker":"[91]"}],"fun_headline_variants":["RPN 2: One formula reproduces CNN, RNN, GNN, Transformer","RPN 2: Interdependence functions unify CNN, RNN, GNN, Transformer","RPN 2: One model, many backbones: CNN, RNN, GNN, Transformer","RPN 2: Architecture choice reduces to picking one interdependence function"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every layer of a real CNN, RNN, GNN, or Transformer, including nonlinearities, normalization, and skip connections, can be absorbed exactly into the RPN 2 composition with only the interdependence matrix varying.","fun_headline_variants_meta":{"raw":{"variants":["RPN 2: One formula reproduces CNN, RNN, GNN, Transformer","RPN 2: Interdependence functions unify CNN, RNN, GNN, Transformer","RPN 2: One model, many backbones: CNN, RNN, GNN, Transformer","RPN 2: Architecture choice reduces to picking one interdependence function"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001002,"raw_usage":{"total_tokens":4274,"prompt_tokens":1017,"completion_tokens":3257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":3162}},"tokens_in":633,"tokens_out":3257,"duration_ms":57628,"temperature":1.0,"reasoning_tokens":3162,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:50:41.172788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a standard transformer block with softmax attention, layer normalization, and a residual connection, and transcribe its forward pass using only functions in the paper's catalogue. If any operation requires a function outside that catalogue, or if the residual-plus-normalization interaction cannot be represented through the interdependence matrices, then the claimed unification is approximate rather than exact.","supporting_citations":[{"cited_title":"RPN: Reconciled Polynomial Network Towards Unifying PGMs, Kernel SVMs, MLP and KAN","cited_arxiv_id":"2407.04819","evidence_quote":"Defines the original RPN three-component architecture that RPN 2 extends, and provides the parameter-reconciliation techniques reused to generate interdependence matrices."}],"review_version":1}