{"id":"6189970d-01d0-4840-8004-9de0d7eb875d","arxiv_id":"2512.11000","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Wiring-pattern similarity alone lets a decoder recover which MNIST digit class an output neuron represents, with 100% accuracy in dropout-trained networks and 38% in standard networks.","lead":"This paper turns \"how ambiguous is a brain-like representation?\" into a number, by measuring how well a decoder can recover what a neuron represents from wiring patterns alone. For simple digit-classifying networks it shows dropout training creates much more decodable, less ambiguous wiring than ordinary training—even when both perform the same task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perfect decoding accuracy only establishes H(I|R,C)=0; label-convention-dependence undermines claim that relational structure alone determines content.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the decoding is conditioned on external labels/context (C), so the measured quantity is H(I|R,C) rather than H(I|R). This is the central soft spot. My proposed permuted-reference test directly demonstrates whether the recovered content depends on the arbitrary labeling of the reference, which is a concrete way to distinguish intrinsic disambiguation from convention-dependence. The paper's own Section 4.4 acknowledges this limitation, which already justifies the CONDITIONAL verdict. The empirical findings—especially the dropout-vs-standard dissociation and cross-architecture consistency—are interesting and valid as within-context results, but the strong philosophical claim that relational structure alone determines content is not supported. Therefore the verdict should remain CONDITIONAL: accept the experimental contribution conditionally on revising the theoretical interpretation to explicitly acknowledge that content is decoded relative to a known context C, and that H(I|R) is not achieved. The proposed test would settle whether the concern actually lands; if the decoded labels are invariant to reference permutations, the strong claim would be supported, and the verdict could be upgraded to ACCEPT.","tokens_in":8317,"tokens_out":11593,"duration_ms":128606,"concrete_test":"Construct two reference Gram matrices from the same 5 dropout-trained networks: G_ref as in the paper, and G_ref' obtained by applying a fixed random permutation π to the rows/columns of each reference network's Gram matrix before averaging. Then run geometric matching on the same 10 test dropout networks using both references. If the decoded class labels for each test network are related by π^{-1} (i.e., the output changes consistently with the reference permutation), this demonstrates that the recovered content is determined by the arbitrary labeling of the reference, not by the relational structure alone—supporting the H(I|R,C) rather than H(I|R) interpretation. If the decoded labels are invariant to the reference permutation, the strong claim gains support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that relational structure alone—without neuron ordering or training procedure—completely determines representational content (Section 3.1.2, Section 4.1) is not established. Both decoding methods are conditioned on an external label convention: geometric matching uses a reference Gram matrix built from 5 networks with known class ordering (Section 2.1.4, Eq. 2), and the learned decoder is trained on class labels. The measured quantity is therefore H(I|R,C), not H(I|R)—a limitation the authors themselves concede in Section 4.4. The high decoding accuracy demonstrates cross-seed consistency of Gram matrices within a fixed task and architecture, but it does not show that the relational structure alone fixes content. The reference's labeling is an arbitrary anchor: applying a global permutation to the class labels in the reference would yield different decoded content for the same test network. Hence the inference from finite decoding accuracy to 'content is structurally determined, not conventionally assigned' (Section 4.1) is a non-sequitur. Additionally, 100% accuracy on 10 validation networks is finite-sample evidence and does not by itself justify the claim H(I|R,C)=0 or ARS=1.0 as proof of intrinsic disambiguation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an information-theoretic measure of representational ambiguity, H(I|R), and tests whether the relational structure of a neural network's weights (cosine-similarity Gram matrices) can be decoded to identify which digit class an output neuron represents and which spatial position an input neuron represents. On MNIST-trained fully-connected networks, geometric matching achieves 100% accuracy for dropout-trained networks (on 10 validation networks), 38% for standard backpropagation, and 10% for untrained networks, despite similar task accuracies. For input neurons, a transformer decoder achieves R^2 = 0.844 (standard) and 0.695 (dropout) in predicting distance from image center. The authors interpret these results as showing that relational structure alone can unambiguously determine representational content, and connect this to narrow representationalism and IIT. The paper includes a theoretical framework, two experiments, ablations, and a discussion of limitations.","tokens_in":8607,"tokens_out":9527,"duration_ms":83052,"significance":"If established, the empirical findings would be of interest to computational neuroscience and representation analysis: they suggest that training paradigm can affect how consistently class identity is encoded in weight geometry, independently of task accuracy. The paper provides a clean operationalization of 'representational ambiguity' and a reproducible experimental pipeline, with code made available. The core observation—that dropout-trained networks exhibit substantially more consistent Gram matrices across seeds than standard backpropagation—is new and worth reporting. However, the central theoretical claim that relational structure alone determines content, independent of any convention, is not supported by the data as presented. The paper's own limitation section concedes that the measured quantity is H(I|R,C), not H(I|R), and the finite-sample evidence for H=0 is weak. The significance of the contribution therefore depends on whether the claims can be tempered to context-relative decodability.","major_comments":[{"comment":"The central claim that 'relational structure alone ... completely determines what each neuron represents' (§4.1) rests on the 100% geometric matching accuracy in Table 1. This accuracy is computed on only 10 validation networks. A perfect score on 10 samples cannot establish zero conditional entropy: with 10 trials, the one-sided 95% confidence lower bound for accuracy is about 0.69 (Clopper-Pearson). Consequently, the ARS = 1.0 in Table 2, which the text equates with H(I|R,C) = 0, is not justified by the data. Please report confidence intervals, increase the validation set, or explicitly treat the 100% figure as a finite-sample estimate and weaken the corresponding claim.","section":"§3.1.2, Table 1, §2.3"},{"comment":"Both decoding methods are conditioned on an external label convention. Geometric matching uses a reference Gram matrix built from 5 networks with known class ordering (Eq. 2); the learned decoder is supervised on class labels. If a global permutation were applied to the class labels of the reference, the optimal matching for the same test network would change accordingly, so the recovered 'content' is relative to the reference labeling. The paper acknowledges in §4.4 that it measures H(I|R,C), not H(I|R), but the rejoinder—dataset identity is decodable and cross-architecture transfer works—does not address label-convention dependence. Therefore the conclusion that content is 'structurally determined, not conventionally assigned' (§4.1) is a non-sequitur. To support the strong claim, the authors would need to demonstrate label-invariance, e.g., by showing that a decoder trained without la","section":"§2.1.4, §4.1, §4.4"},{"comment":"The ARS formula for regression is not scale-invariant as written. Eq. (6) sets Var(Y) = 1 implicitly through the denominator log(2πe). The target in Experiment 2, f(i,j) = Euclidean distance from the image center, has a variance determined by the pixel grid and is not 1. Unless the target is explicitly standardized to unit variance before computing R², the reported ARS values (0.419 and 0.654) are not valid entropy reductions and are likely overestimates. Please state whether Y was standardized; if not, derive the correct expression using the actual variance of Y. The same issue affects cross-task comparisons of ARS in Table 2.","section":"§2.3, Eq. (6), Table 2"}],"minor_comments":[{"comment":"The abstract says 'red square cannot alternatively encode the experience of a green triangle' while the introduction says 'green square.' Please harmonize.","section":"Abstract / §1"},{"comment":"The description of the transformer decoder is ambiguous: 'the only positional information is which row corresponds to the target neuron.' Clarify how the target row is indicated to the model (e.g., a special token, ordering, or masking).","section":"§2.1.4"},{"comment":"The cross-architecture transfer results are reported only as 'Figure 6 shows strong decoding performance.' Provide the numerical accuracies for the 3x3 matrix so the claim can be assessed.","section":"§3.1.5"},{"comment":"The spatial-position regression experiment does not explicitly state the train/validation split. The classification experiment uses 800/200 networks; specify whether the same split was used for the regression decoder.","section":"§3.2"},{"comment":"The claim that 'performance relative to random chance increases systematically with neuron count' conflates changes in absolute accuracy with changes in the chance baseline. Report absolute accuracies for all k values; the 10-neuron absolute accuracy (73.6%) is lower than the 5-neuron value (79.1%), so the 'systematic improvement' is driven mostly by the decrease in chance level.","section":"§3.1.4"},{"comment":"The 'universal decoder' paragraph is speculative. It is presented as evidence that the context dependence is practical rather than fundamental, but no proof or simulation is provided. Label it as a research direction.","section":"§4.4"},{"comment":"Fano's inequality bounds the conditional entropy in terms of the true error probability. Using the empirical accuracy as a plug-in estimate is fine, but the finite-sample uncertainty should be acknowledged where ARS values are interpreted.","section":"§2.3, Eq. (5)"},{"comment":"The choice of 2 training epochs is arbitrary and may underfit MNIST. A sentence justifying this choice or a sensitivity check would strengthen the paper.","section":"§2.1.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is thought-provoking but the gap between the empirical findings and the philosophical conclusions is large. The empirical core—that dropout training leads to more consistent cosine-similarity Gram matrices across seeds—could be a solid contribution if reframed as a context-relative phenomenon. The headline 100% accuracy comes from only 10 validation networks, and the label-convention dependence is a fundamental issue that the current rebuttal in §4.4 does not resolve. I would encourage the editor to consider whether the journal's readership is well served by the strong claims in the title and conclusion, or whether a more modest framing would be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2512.11000. The empirical core is real and worth a look: the authors show that for MNIST classifiers, the Gram matrix of output-layer weights, averaged across seeds, contains enough relational structure to identify which output neuron corresponds to which digit class, and that dropout training makes this dramatically more decodable than standard backprop (100% vs 38% geometric matching, chance 10%). That dissociation is new as far as I know. The spatial-position decoding (R^2 up to 0.844) is a reasonable second domain. The methods are mostly sound: they have untrained controls, an exhaustive permutation search for the output layer, and a local-versus-global ablation showing the full geometry matters. The paper is also unusually honest in §4.4, where it concedes the measured quantity is H(I|R,C), not H(I|R).\n\nThat last concession is the crux. The headline claim—that relational structure alone, without neuron ordering or training procedure, completely determines what each neuron represents—does not follow from the experiments. Both decoding methods rely on an externally supplied labeling convention: the reference Gram matrix in geometric matching is built from 5 networks with known class order, and the learned decoder is trained on class labels. A global permutation of labels in the reference would flip the decoded content for the same test network. The finite-sample issue is real too: 100% on 10 validation networks does not establish zero conditional entropy; at best it gives a lower bound that might be close to zero, but the paper treats it as exact. The authors try to argue the decoder 'discovers structure rather than imposes it,' but that argument is about generalization across network instances, not about independence from the labeling convention.\n\nThe ARS numbers derived from Fano's inequality are methodologically fine but should be presented as what they are: lower bounds under distributional assumptions, not estimates of the true entropy.\n\nThe paper is a serious empirical attempt with a clear and testable idea. It deserves peer review, but an editor should expect major revision: replace the perfect-accuracy overstatement with a proper confidence analysis, explicitly address the label-permutation dependence, and recalibrate the philosophical conclusions to match the H(I|R,C) evidence. The dropout-vs-standard dissociation alone is a solid contribution. I'd bring it to a reading group.","headline":"The empirical decoding results are real and interesting, but the inference to intrinsic unambiguous content is undercut by the paper's own H(I|R,C) concession.","tokens_in":9058,"tokens_out":2592,"would_cite":false,"duration_ms":26635,"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":"This paper establishes that representational content can be fully fixed by a network's relational geometry, and demonstrates 100% decoding in dropout-trained digit classifiers.","keywords":["representational ambiguity","conditional entropy","relational structure","Gram matrix","dropout training","neuron decoding","intentionality","consciousness"],"falsifier":"Scan a large set of dropout-trained MNIST networks seeded differently from the reference networks and check whether the true class permutation always yields the minimal Frobenius distance to the reference Gram matrix — one counterexample would break the zero-variance perfect result. A stronger test: train networks with class labels cyclically permuted, then apply the original decoder; if accuracy drops to chance, the relational geometry encodes a learned label convention rather than intrinsic class meaning.","tokens_in":8154,"feed_emoji":"🧠","tokens_out":5516,"duration_ms":54791,"temperature":0.7,"pith_summary":"The paper sets out to establish that a neural network's representational content can be fixed by the relational structure of its connectivity alone, in the same way conscious states are thought to be intrinsic rather than decoder-dependent. It formalizes ambiguity as conditional entropy H(I|R) and measures it by asking whether a neuron's class identity can be decoded from the Gram matrix of pairwise cosine similarities of its weights. For dropout-trained MNIST classifiers, geometric matching recovers which output neuron represents which digit with 100% accuracy and zero variance, versus 38% for standard backpropagation and 10% for untrained networks. The paper argues this shows relational structure can unambiguously determine content, and that ambiguity varies independently of task performance.","feed_headline":"Wiring alone identifies output neuron meaning with 100% accuracy","feed_subtitle":"In dropout-trained digit classifiers, every output neuron's role is pinned down by its relational geometry; in standard nets it stays near c","key_machinery":"The central object is the Gram matrix of row-wise (or column-wise) cosine similarities of a weight layer, which encodes the relational geometry among neurons invariant to scale and permutation. The paper's two decoding tools are: geometric matching, which searches all permutations of a test network's output neurons to minimize Frobenius distance to a reference Gram matrix averaged from five labeled networks; and a permutation-invariant transformer decoder with no positional encodings, trained on thousands of permuted networks. The theoretical link is an Ambiguity Reduction Score (ARS) derived from Fano's inequality, mapping decoding accuracy and R² to a lower bound on how much H(I|R,C) falls","core_discovery":"On its own terms, the paper claims that the pattern of pairwise cosine similarities between a neuron's weights and those of other neurons—the relational geometry—carries enough information to specify what that neuron represents, without knowing its position, training method, or even which architecture produced it. The central evidence: for dropout-trained feedforward MNIST classifiers, matching a test network's Gram matrix against a reference built from five labeled networks identifies the correct class permutation for all output neurons in every tested network (100%, zero standard deviation); standard backpropagation yields 38% and untrained networks 10%. A transformer decoder without posit","pith_inferences":["I would not take perfect decoding as proof of intrinsic, decoder-independent content: the decoder and reference Gram matrix are built with labels and from the same distribution, so the result may reflect a shared training-set convention rather than an intrinsic property of the representation (the paper concedes it measures H(I|R,C), not H(I|R)).","A testable extension: train networks with class labels permuted relative to a fixed reference; if geometric matching still aligns neurons perfectly, the geometry tracks label statistics rather than true class semantics, whereas if it fails, the geometry is conventional.","Applied to biological recordings, the framework suggests a concrete prediction: neural populations supporting conscious perception should have lower representational ambiguity (higher decodability of stimulus identity from population geometry) than populations processing the same stimuli unconsciously.","The layer-specific reversal (standard training better for input position, dropout better for output class) suggests ambiguity is not a single global property but decomposes by layer and representational domain; any theory linking consciousness to a single ambiguity value would need to specify which layer's geometry is relevant."],"forward_implications":["If correct, representational ambiguity is a measurable, optimizable dimension of neural systems, independent of task performance: dropout changed output decoding from near chance to perfect while classification accuracy stayed nearly constant.","Richer relational structures—more output neurons in the Gram matrix—improve decoding relative to chance monotonically (from 1.0x at two neurons to 7.36x at ten), implying relational complexity itself reduces ambiguity.","Because geometric matching transfers across hidden-layer architectures, the content encoded in relational geometry is not tied to a specific network shape.","Dataset identity can be read off connectivity (99.8% for dropout, 84.3% for standard training), so the context a decoder needs for disambiguation may itself be encoded relationally."],"fun_headline_variants":["Relational wiring alone IDs neuron meaning in dropout nets","Connectivity encodes what neurons represent, not just task","Information-theoretic measure of representational ambiguity","Spatial input position decoded from neural connectivity","Neural geometry uniquely identifies output neuron roles"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The result stands only if the consistency of Gram matrices across trained networks, as read by a decoder that was itself trained on labels, captures what the representation intrinsically means rather than a convention shared by the training set and the decoder.","fun_headline_variants_meta":{"raw":{"variants":["Relational wiring alone IDs neuron meaning in dropout nets","Connectivity encodes what neurons represent, not just task","Information-theoretic measure of representational ambiguity","Spatial input position decoded from neural connectivity","Neural geometry uniquely identifies output neuron roles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000385,"raw_usage":{"total_tokens":1887,"prompt_tokens":771,"completion_tokens":1116,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":1055}},"tokens_in":515,"tokens_out":1116,"duration_ms":11123,"temperature":1.0,"reasoning_tokens":1055,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:17:06.901612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan a large set of dropout-trained MNIST networks seeded differently from the reference networks and check whether the true class permutation always yields the minimal Frobenius distance to the reference Gram matrix — one counterexample would break the zero-variance perfect result. A stronger test: train networks with class labels cyclically permuted, then apply the original decoder; if accuracy drops to chance, the relational geometry encodes a learned label convention rather than intrinsic class meaning.","supporting_citations":[],"review_version":1}