{"id":"c0516bb8-0ba3-47ab-8ff5-c013f72631ac","arxiv_id":"2506.06332","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"The paper derives standard predictive coding update rules and claims a 99.92% CIFAR-10 accuracy that would beat the published leaderboard, but the claim is unverified and internally inconsistent.","lead":"This paper is an introduction to predictive coding networks, brain-inspired neural networks trained with local rules, and it reports a CIFAR-10 accuracy result it calls benchmark-smashing. The tutorial content is clear and standard, but the headline accuracy is a single unverified run with no error bars, so treat it as a claim to check, not a result to trust.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time inference may leak labels: §5.3 is ambiguous about whether the test loop uses Algorithm 2's target-dependent ε_sup; the 99.92% claim hinges on this.","rationale":"The PCN derivation in Sections 2–4 is internally consistent: the energy, inference, and learning rules follow from gradient descent on the stated loss, and the vectorized Algorithm 3 appears dimensionally correct. The tutorial content is a reasonable onboarding document. The problem is the empirical headline. The claimed 99.92% top-1 accuracy is far outside the credible range for a 3.58M-parameter PCN trained for 4 epochs without augmentation, but implausibility alone is not decisive. The decisive issue is that the test protocol is underspecified in exactly the place where it matters: §5.3 says inference is run 'exactly as in the base algorithm,' but the only supervised base algorithm defined earlier, Algorithm 2, consumes the target label and uses it to drive the top latent. Unless the notebook's test loop suppresses y and ε_sup, the test labels are part of the inference computation, which is leakage. This is a concrete mechanism, not a vague suspicion, and it is directly checkable from the published code. The reader's REJECT verdict stands: a central empirical claim that cannot be checked against its own described protocol, and whose only arbiter is unpublished test code, should not be accepted as a research result. Our concern sharpens the rejection rather than changing it.","tokens_in":15584,"tokens_out":6719,"duration_ms":67134,"concrete_test":"Open the accompanying GitHub notebook [20], locate the evaluation cell, and check whether y_batch is passed into the inference loop and whether ε_sup = ŷ − y is computed while test latents are updated. Re-run the frozen weights with the label input replaced by zeros (or with the supervised term removed); if top-1 accuracy drops materially below 99.92%, the published number is a label-leakage artifact. If the test code is absent, this single check already establishes that the central claim is unverifiable as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised result — 99.92% top-1 on CIFAR-10 after 4 epochs — is supported only by §5.3, which describes test inference as 'executed, exactly as in the base algorithm.' That phrase is ambiguous in a load-bearing way. The supervised base algorithm (Algorithm 2, §4.2) requires the target label y as an input: it computes ε_sup = ŷ − y and uses W_outᵀε_sup to update the top latent x^(3) during inference. If the test loop is implemented this way with y_batch supplied, then the correct label is used to move the inferred top latent toward that label before ŷ is read out — a direct test-label leakage. The paper does not print the test-time code, and the Section 5.3 text explicitly refers to 'input-label pair (x(0), y)' rather than only the input. The disclosure in §5.5 ('prevent data-leakage from the test set to validation') concerns validation, not test inference. Therefore the single most load-bearing assumption is that the notebook's evaluation phase omits y from the latent-inference loop. If instead y is absent, the claim still rests on one random-init run with no error bars, but the leakage mechanism is the concrete, checkable failure mode.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a tutorial introduction to predictive coding networks (PCNs). It defines a hierarchical energy function, derives the inference and learning update rules as gradient descent on that energy, presents unsupervised and supervised algorithms in pseudocode, and gives a PyTorch-oriented implementation. The application section reports a CIFAR-10 experiment: a 3,577,100-parameter PCN trained for 4 epochs with batch size 500, no data augmentation, and hyperparameters T_infer=50, T_learn=500, achieving 99.92% top-1 and 99.99% top-3 accuracy, which the paper claims tops the Papers With Code leaderboard at the time of writing.","tokens_in":15800,"tokens_out":9095,"duration_ms":85832,"significance":"The derivation of the PCN inference and learning rules from a single energy is clear, correct in substance, and useful for onboarding practitioners; the explicit algorithms and the disclosure that the experiment was virtually one-shot are strengths. However, the paper's central empirical claim—that a 3.58M-parameter PCN trained for 4 epochs without augmentation beats state-of-the-art CIFAR-10 accuracy—is extraordinary and, as presented, not supported by reproducible evidence. The test protocol is ambiguous about whether labels are used during test-time inference; no error bars or repeated runs are reported despite the paper's own statement that accuracies vary with random latent initialization; the leaderboard comparison is unverified; and the text contains a direct contradiction about whether the result is competitive. If the result is real and reproducible, it would be a major finding for the PCN community, but the current evidence does not establish it.","major_comments":[{"comment":"The test-time protocol is ambiguous in a load-bearing way. Section 5.3 says the inference loop is 'executed, exactly as in the base algorithm' for each input-label pair (x^(0), y). If 'base algorithm' refers to Algorithm 2 (the supervised algorithm), then the test loop computes epsilon_sup = y_hat - y and uses W_out^T epsilon_sup to update the top latent before reading out y_hat, which would be direct test-label leakage and would trivially inflate accuracy. If, instead, the test loop is meant to be the unsupervised algorithm with epsilon^(L)=0, the paper must state this explicitly and provide the test-time code. As written, the 99.92% claim in Table 1 is not verifiable.","section":"Section 5.3 and Algorithm 2"},{"comment":"The reported accuracies come from a single test pass, yet Section 5.3 concedes that predictions contain random noise from the randomly initialized latents and that accuracies vary from one test round to the next. A single realization cannot support a leaderboard-topping claim; report the mean and standard deviation over repeated test rounds with different latent initializations, and ideally over multiple training seeds. Without this, the result could be a lucky draw.","section":"Section 5.3 and Table 1"},{"comment":"The statement that 'neither the model architecture nor the hyperparameters were tuned for performance in any way or at any point' is contradicted by the preceding disclosure, which describes changing eta_infer from 0.1 to 0.05 and eta_learn from 0.001 to 0.005 after the first run because inference was unstable and learning was slow. That is hyperparameter adjustment based on observed behavior, even if not a systematic search. The 'virtually one-shot' characterization must be revised to acknowledge that the reported result came from a second configuration after an unsuccessful first run.","section":"Section 5.5"},{"comment":"The assertion that 99.92% top-1 accuracy 'comfortably tops the leaderboard on Papers With Code' and that the previous record was 99.5% set by ViT-H/14 is not substantiated by any citation, link, or benchmark protocol. A claim of this magnitude requires the exact leaderboard snapshot, the evaluation protocol, and the code used to produce Table 1. The same paragraph also says PCNs are 'generally not yet competitive with state-of-the-art deep learning methods in terms of... raw accuracy,' which directly contradicts the preceding leaderboard claim; these statements need to be reconciled.","section":"Section 5.5 and Table 1"}],"minor_comments":[{"comment":"The convergence statement says 'several works have established... that the alternating optimization procedure... converges to a local minimum' and cites reference [7], whose title is '(Non-)Convergence results for predictive coding networks.' The statement should be qualified to distinguish the settings in which convergence is guaranteed from those in which it is not.","section":"Section 2"},{"comment":"The text says the supervised extension 'entails simply clamping the top latent representation x^(L) to a predicted label y_hat,' but the described method adds a readout layer and does not clamp x^(L); this wording is misleading and should be corrected.","section":"Section 4.2"},{"comment":"The dataset name appears as 'CIF AR-10' in the abstract and once in Section 5; it should be 'CIFAR-10.'","section":"Abstract and Section 5"},{"comment":"The closing sentence 'Happy hunting!' is informal for a journal article and should be removed or replaced.","section":"Section 5.5"}],"recommendation":"major_revision","confidential_remarks":"The benchmark claim is the main barrier to publication. I recommend that the authors be required to supply the exact test-time code and repeated evaluation statistics before the paper is accepted. The tutorial content could be a reasonable expository contribution if the empirical section is either corrected or substantially toned down; with the 'benchmark-smashing' framing, the evidentiary standard should be high."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a decent tutorial on predictive coding networks, but the headline result—99.92% top-1 on CIFAR-10 after 4 epochs with no augmentation and a 3.58M-parameter network—is not credible, and the text itself contains a concrete mechanism for the implausibility: test-time inference may be using the true labels.\n\nWhat is genuinely good: Sections 2–4 give a clear, correct derivation of the standard PCN inference and learning rules from the energy function, and the locality discussion is helpful for newcomers. The PyTorch code in Section 5.4 is readable and would work as a starting template. That part of the paper is solid and worth having.\n\nThe problem is the empirical claim. It rests on one run, no error bars, no augmentation, and no code printed in the manuscript (only a GitHub link). The disclosure is self-contradictory: it calls the experiment \"virtually one-shot\" but then describes a first run with different hyperparameters and a second run that changed both rates. The same section admits PCNs are \"generally not yet competitive\" with SOTA. Those are red flags on their own.\n\nThe stress-test note points to a sharper, load-bearing ambiguity. Section 5.3 says test inference is executed \"exactly as in the base algorithm\" and refers to an \"input-label pair (x(0), y)\" for the test set. If the test loop follows Algorithm 2, the inference step updates the top latent using ε_sup = ŷ − y, i.e., it moves the latent toward the one-hot true label on the test set. That is direct label leakage. The paper never prints the test-time code, and the GitHub link is not a substitute for verification. Given the disclosure admits prediction noise from random latent initialization, a lucky run alone would not explain 99.92%.\n\nEven without leakage, a single random-init run with no error bars is insufficient to claim a leaderboard-topping result. The tutorial does not need this benchmark to be useful; the benchmark actively hurts it.\n\nFor whom: someone wanting a concise introduction to PCNs or a simple PyTorch example will get some value from the derivations and code, if they ignore the results section. For the research community, the paper as a whole is not a reliable contribution. I would not cite it, and I would accept it for peer review only if the editors saw value in flagging the leakage pathology; otherwise it deserves a desk reject. The right fix is to remove or clearly mark the CIFAR-10 claim as an unverified, likely flawed experiment.","headline":"A clean PCN tutorial wrapped around a CIFAR-10 claim that looks like test-label leakage; the benchmark should not be trusted.","tokens_in":16353,"tokens_out":3558,"would_cite":false,"duration_ms":35348,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A predictive coding network with 3.58 million parameters reaches 99.92 percent top-1 accuracy on CIFAR-10 after four epochs using only local, biology-inspired updates.","keywords":["predictive coding networks","local learning rules","inference learning","energy-based models","CIFAR-10","supervised learning","biologically plausible learning"],"falsifier":"Run the accompanying notebook's test phase many times with fresh random latent initializations and compare the distribution of top-1 accuracies; the central claim fails if the typical accuracy is far below 99.92 percent, or if accuracy drops sharply when test-time latents are all initialized to a fixed constant instead of random draws.","tokens_in":15323,"feed_emoji":"🧠","tokens_out":12306,"duration_ms":106139,"temperature":0.7,"pith_summary":"This paper is a tutorial that derives predictive coding networks from a single prediction-error energy and shows how the same update rules serve both unsupervised and supervised learning. Its central empirical claim is that a three-layer PCN with 3,577,100 parameters, trained for four epochs on CIFAR-10 with no data augmentation and no backpropagation, reaches 99.92 percent top-1 test accuracy, a result the paper says beats the previous 99.5 percent it cites. A sympathetic reader would care because the learning rule is local: each weight update uses only presynaptic activity, postsynaptic error, and a local activation derivative, which is the kind of rule biological circuits and low-power hardware can implement. The paper also releases the Python implementation and trained weights, making the claim directly checkable.","feed_headline":"Predictive coding net hits 99.92% on CIFAR-10 in 4 epochs","feed_subtitle":"A brain-inspired 3.6M-parameter network beats far larger models without augmentation or backprop.","key_machinery":"The load-bearing object is the predictive-coding energy $E = \\frac{1}{2}\\sum_{l=0}^{L-1}\\|\\varepsilon^{(l)}\\|^2$, where $\\varepsilon^{(l)}=x^{(l)}-\\hat{x}^{(l)}$ is the difference between a layer's activity and the prediction sent down from the layer above. All updates descend this energy: inference moves each latent $x^{(l)}$ by $-\\eta_{\\mathrm{infer}}\\big(\\varepsilon^{(l)} - W^{(l-1)\\top}(f'^{(l-1)}(a^{(l-1)})\\odot\\varepsilon^{(l-1)})\\big)$, and learning moves each weight by $+\\eta_{\\mathrm{learn}}\\big(f'^{(l)}(a^{(l)})\\odot\\varepsilon^{(l)}\\big)x^{(l+1)\\top}$. The shared factor $h^{(l)}=f'^{(l)}(a^{(l)})\\odot\\varepsilon^{(l)}$, a gain-modulated prediction error, is what makes both updates local, and the supervised extension simply injects $W_{\\mathrm{out}}^{\\top}\\varepsilon_{\\mathrm{sup}}$ at the top layer.","core_discovery":"On its own terms, the paper establishes that a predictive coding network can be trained end-to-end on a standard vision benchmark using only local, biologically plausible updates, and that this training generalizes well: after four epochs, top-1 accuracy is 99.92 percent and top-3 is 99.99 percent on CIFAR-10. The architecture has latent layers of width 1000, 500, and 10, with a linear readout from the top latent; weights are initialized once at the start, latents are reinitialized to small Gaussian noise for every sample, inference runs for 50 steps per sample at rate 0.05, and learning takes 500 weight steps per batch at rate 0.005. Because the energy landscape is deformed only by local, activity-based terms, the result indicates that global error backpropagation is not needed for strong supervised generalization in this setting.","pith_inferences":["Editorial extension: if the reported accuracy is stable across repeated test rounds and independent reimplementations, it implies that optimizing latent states during inference can substitute for much of the capacity and augmentation that feedforward vision models normally require.","Editorial extension: the paper's random test-time latent initialization makes predictions noisy; averaging over several latent draws, or initializing latents with an amortized bottom-up network as in the hybrid model the paper mentions, would be a natural next experiment and could make the headline number more robust.","Editorial extension: the local update rules are stated as extending to convolutional, recurrent, and graph layers, but the paper does not test those variants; checking whether the same two-timescale training holds in those architectures is a direct testable extension."],"forward_implications":["A PCN with 3.58 million parameters reaches 99.92 percent top-1 accuracy on CIFAR-10 after four epochs with no augmentation, which the paper identifies as higher than the previous 99.5 percent it cites.","The same code path handles unsupervised and supervised learning; the only change is the top-layer error signal, so the local-update machinery is not tied to label supervision.","Because inference runs for a variable number of steps, the network supports anytime computation: simple inputs settle quickly, and only surprising inputs need deeper inference.","Weight updates are neuron-local in the paper's sense, compatible with plasticity that does not require a global gradient tape."],"supporting_citations":[{"why":"shows predictive coding networks can approximate error backpropagation with local, activity-based synaptic updates, the theoretical basis for the training rules used here.","marker":"[31]"},{"why":"supplies the CIFAR-10 dataset, including its 50,000/10,000 split, on which the reported accuracies are measured.","marker":"[14]"},{"why":"contains the Python notebook and trained weights that back the reported 99.92 percent result.","marker":"[20]"},{"why":"introduces the predictive coding model of visual cortex whose top-down prediction and bottom-up error architecture the PCN instantiates.","marker":"[23]"},{"why":"demonstrates an exact implementation of backpropagation in predictive coding networks, anchoring the claim that local inference learning can carry global credit assignment.","marker":"[27]"}],"fun_headline_variants":["Predictive coding achieves 99.92% on CIFAR-10 in 4 epochs","Brain-inspired PCN scores 99.92% on CIFAR-10 without backprop","Local learning only: PCN reaches 99.92% on CIFAR-10","No backprop: PCN hits 99.92% on CIFAR-10 in 4 epochs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the test evaluation is clean: no label information leaks into the frozen-weight inference, and the single 99.92 percent run is representative rather than a lucky draw from the randomly initialized latents.","fun_headline_variants_meta":{"raw":{"variants":["Predictive coding achieves 99.92% on CIFAR-10 in 4 epochs","Brain-inspired PCN scores 99.92% on CIFAR-10 without backprop","Local learning only: PCN reaches 99.92% on CIFAR-10","No backprop: PCN hits 99.92% on CIFAR-10 in 4 epochs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1449,"prompt_tokens":802,"completion_tokens":647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":418,"completion_tokens_details":{"reasoning_tokens":546}},"tokens_in":418,"tokens_out":647,"duration_ms":5136,"temperature":1.0,"reasoning_tokens":546,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:06:39.086259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the accompanying notebook's test phase many times with fresh random latent initializations and compare the distribution of top-1 accuracies; the central claim fails if the typical accuracy is far below 99.92 percent, or if accuracy drops sharply when test-time latents are all initialized to a fixed constant instead of random draws.","supporting_citations":[{"cited_title":"Learning multiple layers of features from tiny images","cited_arxiv_id":null,"evidence_quote":"supplies the CIFAR-10 dataset, including its 50,000/10,000 split, on which the reported accuracies are measured."},{"cited_title":"An introduction to predictive coding networks for machine learning, 2025","cited_arxiv_id":null,"evidence_quote":"contains the Python notebook and trained weights that back the reported 99.92 percent result."},{"cited_title":"Can the brain do backpropagation? — exact implementation of backpropagation in predictive coding networks","cited_arxiv_id":null,"evidence_quote":"demonstrates an exact implementation of backpropagation in predictive coding networks, anchoring the claim that local inference learning can carry global credit assignment."}],"review_version":1}