{"id":"83ff0b8b-1f02-4bac-b2ba-80a2d6a70a1d","arxiv_id":"1908.06754","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deterministic symbolic regression method grows a single expression tree by locally improving nodes, returning compact equations that the authors report to be competitive with a neural network on one dataset.","lead":"This paper introduces a deterministic algorithm that builds a compact math equation to fit data, one small change at a time, instead of using random evolutionary search. It promises faster, human-readable regression models, and the authors report accuracy comparable to a neural network on a housing price benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'guaranteed improvement' claim depends on an exact constant search, but Section 2.1's general case is a heuristic over N zero candidates, so the system can stop while an improving replacement exists.","rationale":"The reader's weakest assumption is the same one I would flag, and the conditional verdict already asks for code and broader validation. I see no reason to harden or relax the verdict. The per-change improvement itself is enforced by Eq. 26, so the concern is not that accepted moves fail; it is that the search can stop while an improving replacement exists. This undermines the guarantee reading of 'grown until it fits' and the claim that the method has a solid mathematical basis for the general case, but it does not make the empirical results invalid or the method uninteresting. A concrete completeness test would settle whether the heuristic is a real liability or merely a theoretical gap. I would keep the paper conditional pending that test, code release, and a held-out comparison.","tokens_in":28809,"tokens_out":7586,"duration_ms":82591,"concrete_test":"Build or minimally reimplement the Section 2 algorithm in Julia; at every node where constant search is attempted and the special cases (29)-(34) fail, also solve for all real roots of Eq. 28 (clearing denominators, excluding poles and S-forbidden values) using high-precision polynomial root-finding; evaluate Eq. 26 at every root and compare with the heuristic candidate. Run over random N=5-20 parameter sets with generic c_i,d_i and over the Boston folds. Record how often the heuristic reports no positive reduction while an exact improving constant exists. If even one such case occurs, the completeness premise of the central claim is false and the stop condition is premature.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1's general constant search is not an exact solver. When Equations 29-34 do not apply, the method evaluates only the N zero candidates z_i=b_i/a_i, discards values close to poles or forbidden by S, and picks the remaining candidate with lowest value of Eq. 27. The paper itself says the true minimum 'will not be any of these values' and only expects one to be close. No proof is given that if some constant k makes Eq. 26 positive, one of these zero candidates also does. Constant search is also the engine for constant-variable and constant-expression searches, so a miss can leave every search with no positive reduction and terminate the iteration. That makes the advertised properties—'each change ... guaranteed to improve,' 'grown until it fits the data,' and the resulting parity with an ANN—rest on an unproven completeness assumption. The Future Works section implicitly concedes this by listing exact minimization of Eq. 6 for vector c_i,d_i as open. The narrow statement that executed changes reduce MSE is true by construction, but the paper's stronger methodological claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a deterministic symbolic regression method that builds a single expression tree over {+,-,*,/}. Starting from the constant equal to the mean target, the algorithm searches at each node for a replacement subtree—a constant, a variable, a constant-variable expression, or a constant-expression expression—that reduces mean squared error (MSE). To make this search local, the paper propagates from the root a rational error expression (Eq. 6) with coefficient vectors a_i,b_i,c_i,d_i, using algebraic rules per operator (Eqs. 7-20), together with a set S of forbidden semantics intended to prevent division by zero in ancestors. The paper claims each executed modification is guaranteed to improve MSE and that the tree grows until it fits the data, and reports experiments on the Boston housing dataset, including a comparison to a multilayer perceptron, concluding performance is statistically indistinguishable. A worked example on Newton's law of gravitation is also given.","tokens_in":29114,"tokens_out":9552,"duration_ms":86249,"significance":"The algebraic core—propagating the MSE functional through arithmetic operators as rational functions and using positivity of Eq. 26 as the acceptance test—is a genuinely interesting and largely correct framework; the special-case constant solutions (Eqs. 29-34) are derived correctly, and the Newton's law example is a convincing demonstration that the method can recover known structure. However, the advertised guarantees are stronger than what is proved: the general-case constant search is a heuristic over finite zero candidates, and the empirical comparison selects hyperparameters on the test set. The paper is transparent about the former in Section 6, but this concession directly contradicts the 'guaranteed improvement' and 'grown until it fits' claims. With the claims appropriately scaled down and the evaluation redone with a validation protocol, the framework could be a useful addition to the symbolic regression literature.","major_comments":[{"comment":"The general-case constant search is a heuristic, not an exact solver. In the fallback case, the algorithm evaluates only the N zero candidates z_i=b_i/a_i, excludes values close to poles or forbidden by S, and selects the candidate with lowest Eq. 27. The paper itself notes that 'the minimum value of equation 6 will not be any of these values' and Section 6 lists exact minimization of Eq. 6 for general c_i,d_i as future work. No proof is given that if a constant k exists with positive reduction in Eq. 26, one of the z_i also yields positive reduction. Since every search (constant, constant-variable, constant-expression) and the constant-optimization pass ultimately depends on this constant search, a miss can terminate the entire iterative process prematurely. Consequently, the advertised properties in the Abstract and Section 1—'each change ... guaranteed to improve' and 'grown until it fits the data'—are not established. The narrow claim that executed changes reduce MSE is true by construction when Eq. 26 is positive, but the method's completeness as a search is an unproven heuristic.","section":"Section 2.1, Eqs. (26)-(28), pp. 10-12"},{"comment":"The hyperparameter comparison (minimum MSE reduction, maximum nodes) is made using the test-set MSE, and the 'best test' configurations in Tables 4 and 5 are chosen from the same test folds. The ANN baseline is likewise selected at its best test epoch (epoch 4347 in Table 6). There is no held-out validation set or nested cross-validation. Because model selection and evaluation are done on the same test data, the paired t-test reported on p. 25 does not support the generalization claim that the system 'is able to return test results as good as an ANN.' A validation-based selection procedure or nested cross-validation is necessary to make this claim.","section":"Section 4, Tables 4-5 and the ANN comparison, pp. 21-25"},{"comment":"The invariant that the propagated set S exactly captures all ancestor division-by-zero domains is only illustrated by examples, not proved. The rules involving Inf and NaN (e.g., 's_i=Inf/y_i=NaN when y_i=0' and 's_i=x_i/Inf=0') are stated informally. Since the guarantee that the tree 'will always be correct, with no divisions by 0' (p. 3) and the validity of every semantic check depend on this invariant, the authors should supply a formal inductive proof for the four operators, including the handling of Inf/NaN, or provide a counterexample and restrict the claims accordingly.","section":"Section 2, S-set propagation rules, pp. 8-10 and Tables 1-2"}],"minor_comments":[{"comment":"There is an inconsistency between the initial value of d_i: the text near Eq. 3 sets d_i=-1, while the paragraph introducing the S set (p. 8) sets d_i=1. Since the MSE is squared, both yield the same value at the root, but the sign matters in the propagation rules; please make this consistent.","section":"Section 2, root initialization"},{"comment":"The exponent of the constant selected in iteration 4 (3.932029293203675e19) appears inconsistent with the exponent 1e-19 in the iteration 3 result; please check and correct the table.","section":"Table 3"},{"comment":"The computation of z_i=b_i/a_i does not specify the behavior when a_i=0, nor how to handle c_i=0 when computing poles d_i/c_i; please define these cases explicitly.","section":"Section 2.1, zero-candidate step"},{"comment":"The manuscript states that the source code 'will be provided,' but no repository or link is included in this version; please include it to support reproducibility.","section":"Section 2.8, reproducibility"},{"comment":"The empirical evaluation uses a single dataset (Boston housing); the conclusion that the method is 'as good as an ANN' would be considerably more convincing with additional benchmark datasets.","section":"Section 4, empirical scope"},{"comment":"The expression x3− 2/(2·x3− 1) is ambiguous without parentheses; please insert parentheses to clarify the intended precedence.","section":"Section 2, Eq. (23)"}],"recommendation":"major_revision","confidential_remarks":"The paper is an interesting but overclaimed contribution. The propagation algebra (Eqs. 7-20) and the S-set bookkeeping idea are the strongest parts, and the Newton's law example demonstrates genuine potential. The main obstacles are the unproven completeness of the constant-search heuristic, which undercuts the 'guaranteed improvement' claim, and the test-set-based model selection in Section 4, which invalidates the ANN comparison as a generalization result. Both issues are fixable within the manuscript's scope by reframing the claims and redoing the evaluation with a proper validation protocol; I would not reject the paper on the current evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Joe,\n\nQuick take: this is a genuinely different deterministic symbolic regression algorithm, and the per-node propagation of the MSE equation through +, -, *, / is the real contribution. But the headline guarantee—that every step improves and the tree grows until it fits—is not fully established by the general constant search, and the experiments are too thin to support the \"as good as an ANN\" claim.\n\nWhat's new: the authors derive for each node a local rational MSE function (Eq. 6) and propagate the (a,b,c,d) coefficient vectors downward from the root, together with an S-set that tracks forbidden semantics that would create a division by zero anywhere above. The four arithmetic cases (Eqs. 7–20) check out under exact arithmetic. The S-set bookkeeping with Inf/NaN is more careful than GP's usual protected division. The Newton's-law example is a nice demonstration: the method indeed refines a constant through k/x, kx, etc., and returns the structure of Eq. 36. That is original and useful.\n\nThe soft spots are real, though. Section 2.1's general constant search is a heuristic: when the special cases don't apply, it tries N candidate zero values b_i/a_i, discarding poles and S-forbidden ones, and keeps the best. There is no proof that an improving constant exists whenever one of these candidates reduces MSE, and the paper itself concedes in Future Works that finding the exact minimum for vector c_i,d_i is open. So the stronger reading of \"each change performed to it is guaranteed to improve\"—namely, that the algorithm keeps improving whenever an improvement exists—is not established. The narrow reading (changes that are made improve the objective by construction) is true. That distinction matters for the method's reliability.\n\nThe experimental section is the other weak leg. Only Boston housing, one ANN baseline. Hyperparameters (grid of min-improvement and max node count) appear to be selected using test-set results, and the ANN comparison uses the test-best epoch. That's a favorable protocol for both methods, and the t-test is not convincing evidence of parity. The paper also promises code that I could not find on the arXiv page, and it never compares with GP, GSGP, or FFX even though the introduction sets the work against them.\n\nBottom line: the algebra is the asset. The paper deserves peer review because the propagation idea is novel and the derivation is checkable, but it needs major revision before acceptance: release the code, fix or weaken the guaranteed-improvement claim to what the heuristic can support, and run a held-out validation protocol with real SR baselines. I'd send it out, but I'd expect a tough referee report.","headline":"A sound algebraic core for deterministic symbolic regression, overclaimed by a heuristic constant search and an under-powered experiment.","tokens_in":29560,"tokens_out":3318,"would_cite":false,"duration_ms":35572,"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":"The paper claims that symbolic regression can be performed deterministically by growing a single expression tree whose every replacement step is guaranteed to lower the mean squared error, and that this returns test results as good as a…","keywords":["symbolic regression","deterministic algorithm","expression tree","semantic space","mean squared error","interpretable models","constant search","division by zero"],"falsifier":"Run the fallback constant search on a small synthetic node whose error terms are chosen so that the true minimizer of the equation lies far from every zero $b_i/a_i$ and every pole $d_i/c_i$; if a dense one-dimensional search finds an improving constant while all $N$ candidate zeros return negative reductions, then the claim that accepted changes are always improvements is false.","tokens_in":28617,"feed_emoji":"📈","tokens_out":9175,"duration_ms":87395,"temperature":0.7,"pith_summary":"This paper proposes a symbolic-regression method that replaces the usual random evolutionary search with a single expression tree that is grown step by step. The central claim is that each replacement of a subtree, whether by a constant, a variable, or a one-operator expression involving one of those, is guaranteed to reduce the mean squared error, so the construction never worsens and no population of candidates is needed. Because the process is deterministic, one run per configuration suffices, and because the user can cap the number of nodes, the returned model is a compact equation a human can inspect rather than a black box or a giant tree. The authors demonstrate the method on the law of universal gravitation, where it rediscovers the known formula, and on the Boston housing benchmark, where its test error is not significantly different from a neural network's in their experiments.","feed_headline":"A single expression tree can match neural-net regression accuracy","feed_subtitle":"No random population: every replacement step lowers the error, and the result is a compact, human-readable equation.","key_machinery":"The carrier of the argument is the assignment to every node of a scalar equation of the form $\\frac{1}{N}\\sum_{i=1}^N \\left(\\frac{a_i o_i - b_i}{c_i o_i - d_i}\\right)^2$, together with a set $S$ of forbidden output vectors inherited from divisions higher up the tree. In the $N$-dimensional semantic space, one coordinate per data pattern, so each model is a point, these equations turn improvement into geometry: for each node the set of strictly better subtrees is an ellipsoid (or a related shape), and the searches look for constants, variables, and simple one-operation branches inside it. The propagation rules for $+, -, \\times, \\div$ translate a parent equation into the two child equations and update $S$, so each candidate replacement can be evaluated locally without recomputing the whole tree. The constant search closes the loop analytically in the listed special cases and otherwise falls back to evaluating the zero candidates $z_i=b_i/a_i$.","core_discovery":"The paper's central claim is that symbolic regression can be posed as deterministic descent on a single expression tree. For any node, the contribution of that node to the global mean squared error has the rational form $\\frac{1}{N}\\sum_{i=1}^N \\left(\\frac{a_i o_i - b_i}{c_i o_i - d_i}\\right)^2$ with coefficient vectors $a_i,b_i,c_i,d_i$; the four arithmetic operations propagate these coefficients down the tree, and each propagation also carries a set $S$ of forbidden semantics that trace division-by-zero conditions to the leaves. Improving the tree means finding, at some node, a substitutable subtree whose semantic lies inside the current error ellipsoid of that node, which is exactly a positive reduction in that equation. The available replacements are constants, variables, one-operation constant-variable forms, and constant-plus-subtree forms; constants are refined by a closed-form minimum whenever the coefficients fall into the listed special cases, and otherwise by a candidate-zero heuristic. The authors conclude that the method can return test results as good as an artificial neural network, in low computational time, with a user-set bound on expression size.","pith_inferences":["The paper leaves implicit that its per-step guarantee is a descent property, not a global-optimality property; a natural extension is to run the same deterministic grower from several different initial constants or to stop on a validation set, and measure how often escape from local minima matters.","If the semantic-space geometry is as general as the root-equation derivation suggests, the same node-replacement test could be applied to any model that outputs one value per pattern, allowing the tree to assemble ensembles that mix closed-form expressions with opaque predictors.","The per-node ellipsoid test also yields a by-product the paper does not exploit: for every variable it reports how close that variable's semantic point is to the current improving region, which could be read as an interpretable feature-relevance score during construction."],"forward_implications":["If the central claim holds, symbolic regression becomes reproducible: the same dataset and parameters always return the same expression, so no averaging over random seeds is needed.","The recursion-local evaluation means a bounded-complexity tree can be grown efficiently, and the final model is an ordinary arithmetic expression that can be embedded in spreadsheets or any programming language without a machine-learning library.","Because the user can cap node count, the technique offers explicit control of overfitting by expression size, complementing or replacing implicit regularization.","The worked gravitation example implies the method can build constants from scratch, including very small or very large ones, rather than relying on random constant generation."],"supporting_citations":[{"why":"Supplies the neural-network baseline whose test performance the paper claims to match.","marker":"[1]"},{"why":"Introduces the geometric semantic space in which each model is a point and on which this method is built.","marker":"[12]"},{"why":"Documents the huge expression sizes produced by geometric semantic genetic programming, motivating the complexity limit the paper emphasizes.","marker":"[13]"},{"why":"Provides the law of universal gravitation used as the worked example that the method rediscovers.","marker":"[23]"},{"why":"Supplies the Boston housing dataset used in the 10-fold cross-validation experiments.","marker":"[24]"}],"fun_headline_variants":["Deterministic symbolic regression matches neural nets","No population, one tree: fast deterministic regression","One expression tree, no population, matches neural networks","Deterministic regression without genetic algorithms","One tree, no population: fast symbolic regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, in the general case, the fallback constant search, testing only the $N$ zero candidates $z_i=b_i/a_i$ after discarding values near poles, finds a constant that reduces the error whenever any constant reduces it, because if this heuristic misses, the algorithm stops prematurely and the guaranteed-improvement claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Deterministic symbolic regression matches neural nets","No population, one tree: fast deterministic regression","One expression tree, no population, matches neural networks","Deterministic regression without genetic algorithms","One tree, no population: fast symbolic regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3518,"prompt_tokens":885,"completion_tokens":2633,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2564}},"tokens_in":501,"tokens_out":2633,"duration_ms":21452,"temperature":1.0,"reasoning_tokens":2564,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:00:20.623201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fallback constant search on a small synthetic node whose error terms are chosen so that the true minimizer of the equation lies far from every zero $b_i/a_i$ and every pole $d_i/c_i$; if a dense one-dimensional search finds an improving constant while all $N$ candidate zeros return negative reductions, then the claim that accepted changes are always improvements is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the neural-network baseline whose test performance the paper claims to match."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the geometric semantic space in which each model is a point and on which this method is built."},{"cited_title":"Geometric semantic genetic programming is overkill","cited_arxiv_id":null,"evidence_quote":"Documents the huge expression sizes produced by geometric semantic genetic programming, motivating the complexity limit the paper emphasizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the law of universal gravitation used as the worked example that the method rediscovers."},{"cited_title":"Harrison and Daniel L","cited_arxiv_id":null,"evidence_quote":"Supplies the Boston housing dataset used in the 10-fold cross-validation experiments."}],"review_version":1}