{"id":"5da872ac-9522-461d-b07b-75bab34752b6","arxiv_id":"2505.23615","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A weighted sum over binary logic-rule activations lets Differentiable Logic Networks perform tabular regression with accuracy close to random forests at much lower inference cost.","lead":"This paper extends Differentiable Logic Networks, which learn binary logic rules for classification, to regression by summing rule outputs with learned weights. On 15 tabular benchmarks it matches random forest accuracy at a fraction of the inference cost, while keeping models human-readable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy parity may rest on untested binarization granularity: with 6–10 thresholds per continuous feature, the DLN output is piecewise constant, and Section 4.6 itself concedes capacity limits on complex targets.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the model's representational capacity is tied to the coarseness of the threshold binarization and the limited logic-gate depth. My independent reading of the architecture and results supports this. The SumLayer formula in Section 3.1.3 makes explicit that the prediction is a weighted sum of binary rule activations, so within any cell of the partition induced by the threshold predicates the prediction is constant. Six or ten thresholds per continuous feature is a small number, and the paper provides no direct evidence that this is enough for smooth or steep regression functions. The strongest internal evidence for the concern is the paper's own limitation statement in Section 4.6, which concedes capacity limits on complex decision boundaries, and the per-dataset results where MLP beats DLN by large margins. The absence of significance testing compounds the problem because the aggregate mean R2 difference is small. I do not think this concern overturns the paper's conditional acceptance: the architecture is clearly described, the ablation studies are useful, and the benchmarks are broad. But the central claim would be substantially stronger with a threshold-granularity experiment. Since the reader already returned CONDITIONAL, my recommendation is UNCHANGED rather than a new verdict.","tokens_in":19048,"tokens_out":10597,"duration_ms":114466,"concrete_test":"Run the same pipeline on Airfoil, Electrical, and Bike, plus a synthetic ramp target such as y = min(max(10x - 5, 0), 1) + epsilon, with 20 and 30 threshold neurons per continuous feature instead of 6/10, keeping all other HPO settings fixed. If the R2 gain on any dataset exceeds 0.02, or if the DLN-MLP gap closes by more than half on any of these tasks, the fixed binarization is the binding capacity constraint and the headline claim should be qualified. If R2 is flat within seed noise across threshold counts, the capacity concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central 'match or exceed' claim to hold beyond the 15 chosen benchmarks, the binarization granularity in Section 3.1.1 (six or ten threshold neurons per continuous feature) must be sufficient for continuous regression. Since the SumLayer is a linear combination of Boolean threshold features and their logic combinations, the trained predictor is piecewise constant in the original features. The paper tests this assumption only indirectly: there is no experiment varying the number of thresholds per feature, no synthetic target with known smoothness or steepness, and no paired significance test on the aggregate result. Section 4.6 explicitly states that DLN and other traditional methods 'lag behind the MLP on several datasets, indicating limited capacity when the target function has highly complex decision boundaries.' This aligns with the capacity concern: on Airfoil, Bike, Electrical, Housing, and MPG, the MLP beats DLN by nontrivial margins, while the aggregate mean R2 (0.810 vs 0.808) is close enough to be within noise. If increasing thresholds to 20 or 30 per feature materially improves those datasets, the reported parity is an artifact of the chosen threshold count rather than a property of the logic-rule architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends Differentiable Logic Networks (DLNs) from classification to tabular regression by replacing the classification output with a SumLayer that computes a continuous weighted sum of the binary activations of the last logic layer. It also unifies the previously two-phase training procedure into a single differentiable stage, introduces temperature annealing, and reuses subspace search, straight-through estimators, and input concatenation from the classification DLN. The method is evaluated on 15 public regression datasets against nine baselines, with 10 random seeds and 32/64/128 hyperparameter-search trials. The main claims are that regression DLNs match or exceed baseline accuracy while retaining interpretability and low inference cost, with a mean test R2 of 0.810, a geometric-mean inference cost 5.8x lower than random forest and 86x lower than MLP, and interpretability illustrated on three datasets.","tokens_in":19180,"tokens_out":4314,"duration_ms":43300,"significance":"If substantiated, the paper would broaden the applicability of logic-gate networks to a major task family (regression) while preserving the transparency and inference-efficiency advantages demonstrated for classification. The work is clearly presented and the evaluation is broad: 15 datasets, 10 seeds, and an HPO protocol with up to 128 trials, plus ablation studies for four design choices. The architecture description is detailed enough to reimplement, and the interpretability examples (Figures 7-9) are concrete and informative. However, the central accuracy claim currently rests on aggregate means and ranks without significance testing, the efficiency comparison uses an unvalidated float16 assumption for all models, and the representational-capacity assumption about binarization granularity is not tested. These gaps make the headline claims stronger than the evidence.","major_comments":[{"comment":"The claim that DLN 'matches or exceeds' baseline accuracy rests on the aggregate mean R2 (0.810 vs. RF 0.815 and MLP 0.808) and on average rank, but no significance test is reported across the 15 datasets. The mean differences are smaller than the per-dataset standard errors, and DLN is noticeably worse than MLP on several datasets (e.g., Airfoil 0.889 vs. 0.943, Electrical 0.936 vs. 0.968, Housing 0.766 vs. 0.796). Without a paired test across datasets (e.g., Wilcoxon signed-rank on per-dataset R2 or RMSE) or corrected confidence intervals, the aggregate parity claim is not statistically supported. Please add such tests, or explicitly soften the claim to 'comparable accuracy on these benchmarks.'","section":"§4.2, Table 3"},{"comment":"The efficiency comparison assumes float16 for floating-point arithmetic and int16 for integers for every model, and maps all high-level operations to gate-level OPs. However, training uses float32 for MLP and DLN and float64 for most traditional models, and no evidence is provided that every baseline retains accuracy at float16. Since the reported geometric-mean OP counts for SVR and MLP are 129M and 2.59M, respectively, versus 29.8K for DLN, the claimed margins and the Pareto-frontier conclusion in Figure 4 could change materially if float32 or float64 operation counts were used. Please justify the precision assumption per model, or report OP counts under the precisions actually used.","section":"§4.3, Table 5 and Figure 4"},{"comment":"The binarization granularity is a load-bearing representational assumption: with six or ten threshold neurons per continuous feature, the final predictor is piecewise constant in the continuous features, and Section 4.6 concedes that DLN 'lag[s] behind the MLP on several datasets, indicating limited capacity when the target function has highly complex decision boundaries.' The paper does not vary the number of thresholds per feature, does not test synthetic targets with known smoothness or steepness, and does not quantify how threshold count interacts with accuracy. Please add a sensitivity study over the number of thresholds per feature (e.g., 6, 10, 20, 30) and, ideally, a synthetic experiment where target complexity is controlled, so that the reported parity is not an artifact of a fixed discretization budget.","section":"§3.1.1 and §4.6, Table 6"}],"minor_comments":[{"comment":"The ThresholdLayer forward pass in Algorithm 1 is written as Sigmoid(si * (xi - bi)), omitting the division by temperature tau that appears in Eq. (1) and Table 1. Please make the algorithm consistent with the equations.","section":"Algorithm 1, line 7 vs. Eq. (1) and Table 1"},{"comment":"The Pareto-frontier plot would benefit from error bars or confidence bands on the R2 values, since each point is an average over 10 seeds, and the legend entries 'Interpretable' and 'Not interpretable' should be clarified in the caption.","section":"Figure 4"},{"comment":"The arrows (↑, ↓, →) used to indicate changes across HPO budgets are not explained in the caption; please define them or remove the arrows.","section":"Table 4"},{"comment":"For the 'No concat' ablation, the mean R2 drop is 0.802 to 0.782, but the per-dataset drop is much larger on Parkinson's (0.798 to 0.649); reporting the per-dataset changes alongside the mean would better characterize the effect.","section":"§4.5, Table 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript does not state whether code or trained models will be released; for an empirical architecture paper, a reproducibility artifact would strengthen the contribution. Also, the comparison omits modern interpretable regression baselines such as RuleFit and GAMs, which are direct competitors for the claimed transparency advantage; this could be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does what it says: it extends the authors' classification DLN to regression by swapping the output layer for a SumLayer that computes a weighted sum of binarized logic-rule activations. That is a real extension, not a restatement of [3]. The unified single-phase training and temperature scheduling are sensible refinements, and the ablation study actually tests them, which I appreciate. The evaluation is broad: 15 datasets, 10 seeds, 128 HPO trials, with standard baselines. The interpretability figures are genuinely readable; you can see the learned thresholds and rule weights. The code is not released and the hyperparameters for the DLN itself are only partially specified, which is a real weakness for reproducibility, but nothing here smells like result-shopping.\n\nWhere are the soft spots? First, the headline 'match or exceed' claim is weaker than it looks. The mean R2 is 0.810 vs 0.808 for MLP and 0.815 for RF. That is within noise; there is no significance testing, paired or otherwise, and the per-dataset margins are non-trivial in both directions. The paper's own Section 4.6 concedes DLN lags behind MLP on several datasets with complex decision boundaries. Second, the capacity assumption about binarization granularity is real: with six or ten thresholds per continuous feature, the predictor is piecewise constant. The stress-test note is right that there is no experiment varying threshold count, even though that is the obvious knob to turn. I would not call this a fatal flaw, because the empirical results cover a reasonable range of benchmark complexities, but it does mean the generality of the accuracy claim is unproven. Third, the OP-counting efficiency comparison assumes float16 for all models, which may flatter the logic-based models; the relative ranking is probably right, but the specific 5.8x and 86x figures are rough. Fourth, the omission of interpretable regression baselines the paper itself cites—GAMs, RuleFit—is a missed opportunity; the practical use case is regulated settings where interpretability matters, so comparing against other interpretable regressors would have been the stronger test.\n\nOverall, this is a solid empirical paper with honest limitations. The central claim—that regression DLNs are a viable cost-effective transparent alternative—holds up within the tested scope, but the accuracy parity is not as clean as the abstract suggests. The paper deserves a serious referee, primarily to push for code release, significance testing, and a threshold-count sensitivity analysis. I would bring it to a reading group focused on interpretable ML; it is a useful data point, not a breakthrough.","headline":"A competent, honest extension of the authors' classification DLN to regression, with broad benchmarks and a defensible central claim, but the accuracy parity is within noise on several datasets and the capacity/interpretability trade-off deserves a sharper test.","tokens_in":19792,"tokens_out":675,"would_cite":true,"duration_ms":7902,"reading_group":"yes","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 claims that a logic circuit whose final layer computes a weighted sum of binary rule activations can match random forests and multilayer perceptrons on tabular regression while staying interpretable and cheap to run.","keywords":["differentiable logic networks","tabular regression","interpretable machine learning","logic gate networks","efficient inference","rule extraction","continuous relaxation","SumLayer"],"falsifier":"Train the regression DLN on a one-dimensional dataset whose target is a high-frequency smooth function, such as $y=\\sin(20x)$ on $[0,1]$ with thresholds chosen from the data; if its $R^2$ stays far below a simple MLP's unless the number of thresholds per feature is raised far beyond ten, the capacity assumption fails.","tokens_in":18741,"feed_emoji":"⚡","tokens_out":6563,"duration_ms":58213,"temperature":0.7,"pith_summary":"The paper extends Differentiable Logic Networks (DLNs) from classification to tabular regression by replacing the final class-vote layer with a SumLayer that computes a weighted sum of binary logic-rule activations. The central claim is that this output change lets a network of two-input Boolean gates predict continuous targets with accuracy comparable to random forests and multilayer perceptrons, while keeping the model transparent and its inference cost orders of magnitude lower. On 15 public regression benchmarks the regression DLN ranks second overall in mean $R^2$ and sits on the Pareto frontier of accuracy versus operations under every hyperparameter budget tested. If true, this makes DLNs a practical option for regression settings where model transparency and deployment on resource-constrained hardware matter.","feed_headline":"Weighted logic rules match neural nets on regression","feed_subtitle":"A transparent sum of binary rule activations reaches top-tier accuracy with far fewer operations.","key_machinery":"The load-bearing mechanism is the SumLayer, which replaces the classification DLN's discrete class voting with a weighted sum $y=\\sum_j \\mathbb{1}\\{\\mathrm{Sigmoid}(s_j/\\tau)\\ge\\theta_{\\mathrm{sum-th}}\\}\\cdot c_j\\cdot x_j$ over the binary outputs $x_j$ of the final LogicLayer. Each $x_j$ is a two-input Boolean function of earlier binary signals, so the model is a logic circuit whose prediction is an intelligible weighted sum of rule firings. Around this output layer, the paper keeps the DLN machinery: a ThresholdLayer binarizes continuous inputs through learned biases and slopes, LogicLayer neurons choose among the 16 two-input Boolean functions and their two incoming connections by temperature-annealed softmax logits, and straight-through estimators keep the discrete forward pass differentiable. The learned rules are then simplified symbolically, shrinking the circuit and making the extracted rules easier to read.","core_discovery":"The authors' core discovery is that continuous regression does not require abandoning logic circuits: the activations of the final logic layer are binary rule evaluations, and a differentiable SumLayer that learns a real-valued coefficient per rule maps those binaries to a continuous prediction. At inference the connections are binarized by thresholding learnable gates, the coefficients stay in floating point, and the output is an explicit weighted sum of logic rules, so the whole model can be read as a small set of if-then rules. Across 15 tabular datasets, this architecture achieves a mean test $R^2$ of 0.810 with an average rank of 3.47, ahead of the MLP's 3.67 and behind random forest's 2.67, while its geometric-mean inference cost is 5.8 times lower than random forest and 86 times lower than the MLP. The paper also shows that a single end-to-end training phase with annealed temperatures outperforms the two-phase schedule used for classification DLNs.","pith_inferences":["The paper leaves implicit that the number of thresholds per feature is the main capacity dial; an adaptive threshold-count selection could extend the method to smoother targets.","A natural next step, not pursued here, is to attach the SumLayer to convolutional or temporal logic networks for continuous forecasting.","Because the output is linear in binary rules, classical sparsity penalties on the SumLayer coefficients could shrink the rule set further.","The Pareto analysis suggests that on strict inference budgets the DLN's advantage over random forest grows; a hardware benchmark on FPGAs or microcontrollers would quantify this."],"forward_implications":["If the central claim holds, transparent rule-based regression can reach the accuracy of black-box neural models on many tabular tasks without post-hoc explanation.","Because inference uses only binary logic gates and a short weighted sum, regression DLNs can be deployed where floating-point-heavy MLPs are too slow or too power-hungry.","The learned circuit is itself the explanation: each prediction is a readable weighted sum of binary conditions on the original features.","Unified single-phase training with temperature annealing means the same end-to-end gradient pipeline used for classification can be reused for regression, simplifying adoption.","Implicit feature selection in the learned rules can reduce the input features a deployed model needs to measure."],"supporting_citations":[{"why":"Establishes the classification DLN architecture, thresholding, subspace search, and straight-through estimator strategies that this paper adapts to regression.","marker":"[3]"},{"why":"Supplies the real-valued logic relaxation and the 16 two-input Boolean operators that every LogicLayer neuron chooses among.","marker":"[1]"},{"why":"Provides the straight-through estimator used to keep the discrete forward pass differentiable during training.","marker":"[44]"},{"why":"Motivates concatenating binarized inputs to intermediate LogicLayers, a shortcut whose removal substantially reduces accuracy in the ablations.","marker":"[45]"},{"why":"Supplies the symbolic simplification that shrinks trained logic expressions and improves interpretability.","marker":"[46]"}],"fun_headline_variants":["Logic neural nets match deep nets on regression","Weighted rule sums rival MLPs on tabular data","Interpretable logic networks keep pace on regression","Transparent DLNs match neural accuracy on tables","Binary logic rules deliver regression accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that turning each continuous feature into a handful of binary thresholds (six or ten) and combining them with two-input logic gates is expressive enough to approximate the target function as accurately as a neural network.","fun_headline_variants_meta":{"raw":{"variants":["Logic neural nets match deep nets on regression","Weighted rule sums rival MLPs on tabular data","Interpretable logic networks keep pace on regression","Transparent DLNs match neural accuracy on tables","Binary logic rules deliver regression accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000398,"raw_usage":{"total_tokens":2061,"prompt_tokens":906,"completion_tokens":1155,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":1087}},"tokens_in":522,"tokens_out":1155,"duration_ms":9214,"temperature":1.0,"reasoning_tokens":1087,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:41:42.705547+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the regression DLN on a one-dimensional dataset whose target is a high-frequency smooth function, such as $y=\\sin(20x)$ on $[0,1]$ with thresholds chosen from the data; if its $R^2$ stays far below a simple MLP's unless the number of thresholds per feature is raised far beyond ten, the capacity assumption fails.","supporting_citations":[{"cited_title":"Learning interpretable differentiable logic networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the classification DLN architecture, thresholding, subspace search, and straight-through estimator strategies that this paper adapts to regression."},{"cited_title":"Deep differentiable logic gate networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-valued logic relaxation and the 16 two-input Boolean operators that every LogicLayer neuron chooses among."},{"cited_title":"Wide & deep learning for recommender systems,","cited_arxiv_id":null,"evidence_quote":"Motivates concatenating binarized inputs to intermediate LogicLayers, a shortcut whose removal substantially reduces accuracy in the ablations."}],"review_version":1}