{"id":"60c69d99-8a0f-4cbd-80cd-50db19ca0271","arxiv_id":"2509.00651","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"NICA, an iterative self-attention model styled after cellular automata, achieves lower mean RMSE than Mean, KNN, MICE, and GAIN on 14 of 15 tabular datasets with simulated MCAR missingness.","lead":"A new machine-learning model called NICA, inspired by Neural Cellular Automata, imputes missing table values by repeatedly applying self-attention to find similar rows and grow the missing entries. It reports lower imputation error than four common baselines on 15 public datasets, but modern deep-learning imputation baselines are not compared.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Negative missing-loss term in §4.1 is a negative squared error, so the objective rewards arbitrarily large imputed values; unless the released code contains an unstated bound or clip, the training is ill-posed and the SOTA claim is unsupported.","rationale":"The strongest claim is empirical: NICA outperforms state-of-the-art imputers in RMSE and downstream prediction. The reader correctly notes that the four baselines (Mean, KNN, MICE with default settings, GAIN) are too weak to establish 'state-of-the-art'. However, the more fundamental threat in the current manuscript is internal. In §4.1, L_missing is defined as the negative squared distance between the grown values and the zero-initialized values at missing cells. A negative quadratic term is unbounded below as a function of the imputed value; without an explicit bound, clipping, or carefully designed coupling through the known-cell losses, the optimizer is incentivized to make missing imputations arbitrarily large. The paper gives no such mechanism and yet reports low and stable RMSE. This tension means either the equations do not describe the implemented method, or the experiments reflect an undocumented early-stopping/regularization artifact. Either way, the central claim cannot be accepted as written. I retain the reader's CONDITIONAL verdict because the issue is checkable via the provided code, but the condition should now include verifying a well-posed training objective and re-running the benchmarks under that objective, in addition to strengthening the baselines.","tokens_in":13567,"tokens_out":12333,"duration_ms":164268,"concrete_test":"Download the released code from github.com/TrungTin98/NICA and instrument one training run (e.g., California) to log L_missing and the maximum absolute value of outputs at original missing cells for 1000 and 10000 iterations. If |v| grows monotonically and L_missing keeps decreasing, the objective is unbounded as written. If outputs stay bounded, identify the exact mechanism (e.g., clipping, normalization, stop-gradient, or a different L_missing) and re-run Table 2 with that mechanism reported. Additionally, replace L_missing with a bounded penalty, e.g., −min(v², τ) or a margin loss, and check whether the 14/15 RMSE result survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines the total loss as L_model = α1 L_recovering + α2 L_observed + L_missing, with L_missing = −E_{1−M}[ L(X_c^K, X̄_0) ], where L is squared difference and X̄_0 is the zero-imputed input. For each originally missing cell, this contributes −c·v² to the loss. Increasing |v| lowers the objective without any described bound; no clipping, bounded activation, or margin term is mentioned. The claimed stable, low-variance RMSE is therefore not explained by the stated objective: either the implementation uses a different/bounded missing loss, or the benchmarks reflect early stopping before the negative term drives missing outputs to large magnitudes. This is an internal correctness issue independent of the baseline set. It must be resolved before the abstract claim 'outperforms state-of-the-art' can be evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural Imputation Cellular Automata (NICA), a generative model for missing tabular data imputation. NICA represents a data table as a grid of cells, uses cosine self-attention to define neighbors, and iteratively updates the seed through K steps. The loss combines a recovering loss on corrupted observed cells, an observed loss on known cells, and a negative missing loss that penalizes small changes from the zero imputation. Experiments on 15 datasets with 40% MCAR missingness compare NICA against Mean, KNN, MICE, and GAIN; the paper reports RMSE and downstream prediction accuracy, claiming lower error in 14/15 datasets and 'state-of-the-art' performance.","tokens_in":13889,"tokens_out":5438,"duration_ms":58337,"significance":"The paper introduces a novel, conceptually interesting application of neural cellular automata to tabular imputation, offering a lightweight alternative to deep generative models. Strengths include a public code repository, evaluation across 15 datasets with means/standard deviations and multiple missing rates, and downstream task validation. If the loss issue is resolved and the baseline set is strengthened, NICA could be a useful new imputer. However, current evidence for the state-of-the-art claim is weakened by an unbounded negative loss term and a narrow baseline selection.","major_comments":[{"comment":"The missing loss L_missing = -E_{1-M}[L(X_c^K, X̄_0)] is a negative squared error between the imputed values and zero. Because L_observed and L_recovering apply only to observed/corrupted cells, the objective is unbounded below: it rewards arbitrarily large values at missing cells. No clipping, bounded activation, margin, or regularization is described. This makes the reported stable, low RMSE results internally inconsistent with the stated objective. The authors must clarify the actual loss (e.g., a bound/clip or a different formulation) and, if necessary, rerun the experiments; otherwise the central claim cannot be evaluated.","section":"§4.1, definition of L_missing"},{"comment":"The abstract and §5.2.1 claim 'state-of-the-art' performance, but the comparison includes only Mean, KNN, MICE with scikit-learn defaults, and GAIN. The paper's own literature review (§2.1) describes stronger generative imputers (MIWAE, HI-VAE, MIDA), and other self-attention imputation methods are not evaluated. Without these or otherwise tuned baselines, the reported 14/15 superiority may reflect baseline weakness rather than genuine SOTA performance. The authors should add stronger baselines or temper the claim.","section":"§5.1.2, Baselines"},{"comment":"The claim of 'outperforms' is based on mean RMSE, but no statistical significance tests are reported. In several datasets the standard deviations overlap (e.g., Yeast NICA 0.111±0.004 vs MICE 0.118±0.008; Breast NICA 0.070±0.002 vs MICE 0.072±0.005), and on Yacht NICA is not the lowest. The paper should include paired significance tests (e.g., paired t-test or Wilcoxon over the 10 runs) to support the superiority claim.","section":"§5.2.1, Table 2"}],"minor_comments":[{"comment":"Typographical errors ('architechture', 'insprired', 'apllied', 'infomation') and inconsistent notation ('L_recovering' vs 'L_recover') should be corrected.","section":"Throughout"},{"comment":"The notation for the expectation subscripts is not defined rigorously; e.g., E_{M_c} and E_{1-M} should be clarified as averaging over the corresponding cell subsets.","section":"§4.1"},{"comment":"The paper states 'missing data is standardized and 20% of the observed part is randomly corrupted for v=8 times.' Clarify whether the reported results are averages over the v versions or a single imputation; the relationship between training-time corruption and test-time missingness should be explicit.","section":"§5.1"},{"comment":"The clustering behavior is illustrated with a single t-SNE plot; a quantitative cluster metric would strengthen the claim.","section":"§5.2.1, Figure 2"},{"comment":"The line 'X(0)_c ← shuffle(X_c)' shuffles rows, but since rows are independent in the attention computation, this appears to have no effect; consider removing or explaining its purpose.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the negative missing loss: as written, the objective is unbounded and would drive missing imputations to large values, yet the reported RMSEs are stable. The authors likely intended a different loss (e.g., a margin penalty) or rely on early stopping; this needs to be resolved before publication. Additionally, the baseline set is quite minimal for a 'state-of-the-art' claim. I would recommend major revision but not rejection, as the core idea is interesting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper applies NCA-style iterative growth to tabular imputation with repeated cosine self-attention, and that combination is genuinely new. Second, the main objective has a serious flaw: the missing loss in Section 4.1 is a negative squared error, so it is unbounded below. As written, the model is rewarded for making missing values arbitrarily large. Unless the released code contains an unstated bound or clip, training is ill-posed, and the reported 14/15 RMSE win cannot be taken at face value.\n\nWhat the paper does well: the method is simple and practical—zero-impute, corrupt observed cells, iterate self-attention plus an FNN for K steps, and use three losses. It does not need a fully observed training set. The experiments are broad: 15 UCI/Kaggle datasets, 10 runs with standard deviations, missing rates from 0.1 to 0.8, and a downstream prediction task. NICA consistently beats the four baselines. The code is promised, which helps reproducibility.\n\nThe soft spots are real but secondary. The baseline set is thin for a 'state-of-the-art' claim: Mean and KNN are trivial, MICE uses default scikit-learn settings, and GAIN is the only generative competitor. The paper's own literature review cites MIWAE, HI-VAE, and MIDA, but none are benchmarked. All experiments are MCAR; no MAR or MNAR. There are no significance tests, so some margins could be noise. The post-imputation prediction results are more mixed, though still mostly favorable.\n\nIf the missing-loss sign is a typo and the code actually uses a bounded loss, then the method is an incremental but useful contribution to tabular preprocessing. If not, the empirical results are probably an artifact of early stopping. Either way, the authors should be asked to resolve the loss issue, add stronger baselines, and report statistical tests.\n\nThis paper is for researchers working on tabular imputation, especially applied settings. It deserves a serious referee rather than a desk reject—the idea is novel and the empirical scope is substantial—but it needs major revision before the claims are credible.","headline":"Novel NCA-style attention growth for tabular imputation, but the negative missing-loss term in §4.1 makes the objective unbounded below, so the SOTA claim is unsupported without a fix and stronger baselines.","tokens_in":14348,"tokens_out":4803,"would_cite":false,"duration_ms":55050,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural cellular automaton that grows missing values via self-attention claims state-of-the-art tabular imputation.","keywords":["missing data imputation","neural cellular automata","tabular data","self-attention","cosine attention","generative imputation","MCAR","NICA"],"falsifier":"Run the same 15 benchmarks at 40% MCAR with NICA compared against MIWAE and HI-VAE (or a well-tuned MICE with predictive mean matching). If NICA no longer achieves the lowest RMSE on most datasets, the state-of-the-art claim is falsified. A second check: introduce MNAR missingness where blankness depends on the missing values themselves; NICA's attention-based growing would need to handle the bias.","tokens_in":13491,"feed_emoji":"🧬","tokens_out":5107,"duration_ms":55593,"temperature":0.7,"pith_summary":"The paper introduces NICA, a model that imputes missing values in tabular data by treating rows as cells in a cellular automaton and replacing the usual local rule with a self-attention mechanism. The central claim is that after a fixed number of growing steps, the model regenerates unobserved entries and does so with lower imputation error than Mean, KNN, MICE, and GAIN on 14 of 15 benchmark datasets, and better downstream prediction in 11 of 15. A notable property is that training does not require a fully observed dataset, which matters for real-world settings where complete data is unavailable. If the claim holds, NICA offers a comparatively simple generative imputer that avoids adversarial or variational training.","feed_headline":"Cellular-automaton imputer wins 14 of 15 tabular tests","feed_subtitle":"NICA replaces local rules with attention, growing missing values from similar rows.","key_machinery":"The central machinery is the NICA block: CosineAttention(Q,K,V)=softmax(sqrt(d_k) * cosine(Q,K)) V, where cosine similarity replaces the dot product, no value projection is used, and the attention output is concatenated with the current state, passed through a two-layer FNN, and added residual-style. This block is applied K times (K=10 in experiments) to grow the data. The design is justified as analogous to NCA's fixed kernels while allowing each row to dynamically find neighbors in the whole dataset. A Dropout1d layer skips 10% of rows each step to mimic asynchronous growth.","core_discovery":"On the paper's own terms, the discovery is that the NCA paradigm transfers from images to tabular data if the convolutional neighborhood is replaced by a similarity-based attention neighborhood. NICA starts from a zero-filled, randomly corrupted version of the incomplete table, then iteratively applies a block of cosine self-attention, a feedforward network, and a residual update K times. Three loss terms steer the process: recovering corrupted observed cells, matching observed cells, and forcing missing cells away from their zero initialization. The grown matrix supplies the imputed values. Reported results: lowest RMSE in 14/15 datasets at 40% missingness, with 11–34% improvements over GAI","pith_inferences":["The claim of state-of-the-art is tied to the four chosen baselines; comparing against more recent generative imputers (e.g., MIWAE, HI-VAE, or MissForest) would stress-test the conclusion, and the paper itself names those methods as stronger generative alternatives.","Cosine attention with a single head and no value projection is an unusual simplification; testing whether multi-head or projected attention helps on wide or high-dimensional tables would clarify where the design boundary lies.","The method is evaluated only under MCAR; extending the same growing rule to MAR or MNAR patterns, where missingness depends on observed or missing values, would be a natural next test.","The K-step iterative refinement resembles a discrete diffusion process; a continuous-depth version could connect NICA to neural ODEs and allow adaptive step counts per dataset."],"forward_implications":["If NICA is as robust as reported, it becomes a strong default imputer for MCAR tabular missingness, especially when missing rates are high (up to 80%).","Because it trains directly on incomplete data, it can be applied in settings where no complete ground-truth table exists.","The growing-process view suggests a middle ground between single-pass imputers and deep generative models: no adversarial training, no variational bound, just repeated refinement.","The reported clustering of rows during growth points to a learned similarity structure that could be reused for outlier detection or data visualization."],"supporting_citations":[{"why":"Supplies the NCA growing rule that NICA adapts to tabular data.","marker":"[16]"},{"why":"Provides the attention mechanism that NICA modifies into cosine attention.","marker":"[36]"},{"why":"GAIN is the generative baseline and the main comparison for the state-of-the-art claim.","marker":"[40]"},{"why":"MICE is the chained-equations baseline used in the benchmarks.","marker":"[35]"},{"why":"KNNimputer is the nearest-neighbour baseline used in the benchmarks.","marker":"[34]"},{"why":"Supports that deep convolutional networks can represent cellular automata, motivating the NCA-based approach.","marker":"[5]"}],"fun_headline_variants":["Attention-based NCA imputer wins 14 of 15 tabular tests","Neural cellular automata impute missing data with attention","NICA beats state-of-the-art on missing tabular data","Similarity-neighborhood NCA outperforms imputation rivals","NCA paradigm adapted to tabular imputation succeeds"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the four baselines tested—Mean, KNN, MICE with default settings, and GAIN—are sufficient to define 'state-of-the-art' imputation; if stronger or carefully tuned baselines had been included, the reported margins could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Attention-based NCA imputer wins 14 of 15 tabular tests","Neural cellular automata impute missing data with attention","NICA beats state-of-the-art on missing tabular data","Similarity-neighborhood NCA outperforms imputation rivals","NCA paradigm adapted to tabular imputation succeeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1070,"prompt_tokens":670,"completion_tokens":400,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":414,"tokens_out":400,"duration_ms":4816,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:21:09.404363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 15 benchmarks at 40% MCAR with NICA compared against MIWAE and HI-VAE (or a well-tuned MICE with predictive mean matching). If NICA no longer achieves the lowest RMSE on most datasets, the state-of-the-art claim is falsified. A second check: introduce MNAR missingness where blankness depends on the missing values themselves; NICA's attention-based growing would need to handle the bias.","supporting_citations":[{"cited_title":"Growing neural cellular automata","cited_arxiv_id":null,"evidence_quote":"Supplies the NCA growing rule that NICA adapts to tabular data."},{"cited_title":"Gain: Missing data imputation using generative adversarial nets, in: International conference on machine learning, PMLR","cited_arxiv_id":null,"evidence_quote":"GAIN is the generative baseline and the main comparison for the state-of-the-art claim."},{"cited_title":"mice: Multivariate imputation by chained equations in r","cited_arxiv_id":null,"evidence_quote":"MICE is the chained-equations baseline used in the benchmarks."},{"cited_title":"Missing value estimation methods for dna microarrays","cited_arxiv_id":null,"evidence_quote":"KNNimputer is the nearest-neighbour baseline used in the benchmarks."},{"cited_title":"Cellular automata as convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Supports that deep convolutional networks can represent cellular automata, motivating the NCA-based approach."}],"review_version":1}