{"id":"58c43fca-62ab-4674-a0ce-757ca0b05a5b","arxiv_id":"2411.15292","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A regularity tangent vector, computable with constant training overhead, turns the loss gradient at a candidate point into an estimate of its influence on model complexity for active learning.","lead":"This paper proposes an active learning method that tracks a \"regularity tangent\" vector during training, giving a cheap score for which unlabeled points to label next. The score measures how much a point would change the model's complexity, but the paper does not yet test whether this improves learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core identity dL/ds = dR_s/dε is sound, but the active-learning claim rests entirely on the untested 'Hypothesis' in §3.1; the paper's own §3.3 defers all comparisons. A benchmark comparing the squared-loss-derivative query rule to random selection would settle whether the central claim holds.","rationale":"The reader's weakest assumption is precisely the untested heuristic, and I agree with that identification. The derivation of the regularity tangent and the identity dL/ds = dR_s/dε are correct under the stated smoothness and invertibility assumptions; I found no sign or algebra errors in Eqs. 64-70, Eqs. 112-125, or the LiSSA equivalence discussion. Thus the paper's contribution is a plausible and efficient influence-based score. However, the central claim that this score selects good active-learning queries requires the Section 3.1 Hypothesis, which is not supported: the paper explicitly postpones comparison, removes the body of Section 3.2, and provides only illustrative plots rather than a query-loop evaluation. A small benchmark using the paper's own linear model would directly test whether the heuristic beats random acquisition and whether the stated hypothesis has any empirical content. Given that no mathematical error is demonstrated but the central validation is missing, the reader's CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":26282,"tokens_out":6454,"duration_ms":66948,"concrete_test":"Implement the exact squared-loss-derivative query for the paper's polynomial/linear setting and 2-3 small regression datasets (e.g., UCI Diabetes, Energy, Boston): at each round, from the current labeled set compute s* (e.g., LOOCV), θ*=(X^TX+s*I)^{-1}X^Ty, and q(x)=16 σ̂² (x^T (X^TX+s*I)^{-1}θ*)² with σ̂² estimated from training residuals; acquire the unlabeled point maximizing q(x), retrain, and record test MSE versus number of labels. Compare with random acquisition, uncertainty sampling (predictive variance), and expected-model-change over at least 20 seeds, reporting mean and CI. If SLD does not beat random with non-overlapping CIs on most datasets, the central active-learning claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mathematical scaffolding survives scrutiny: Eqs. 68-70 follow from the implicit function theorem and symmetry of H, and SGDF (Eqs. 112-125) is a coherent forward-mode extension of SGD. What does not follow is the active-learning claim. Section 3.1 states a 'Hypothesis' — points whose predictions vary most with regularity s are those best able to stabilize the model — and every proposed query heuristic (SLD, and the removed §3.2 variants) inherits it. No derivation connects |dL(z,θ*)/ds| to expected reduction in generalization error, label informativeness, or any downstream objective. The paper's own evidence is limited to plots of score curves (§3.3) with no query loop, no test-set evaluation, and an explicit note that comparisons were postponed; §3.2's body is removed. Consequently the method could be selecting high-leverage or outlying candidates that change model complexity without improving predictions, and could underperform random acquisition. A second, secondary gap is that the SGDF dual iterate has no convergence proof, so the claimed constant-slowdown estimate of -H^{-1}ρ is not guaranteed even when θ converges. But the decisive gap is empirical: the central claim of an efficient active-learning method is unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an active-learning query heuristic for regularized regression based on a \"regularity tangent\" dθ*/ds = -H^{-1}ρ, where s is a scalar regularity hyperparameter and ρ = ∂²f/∂θ∂s. The central identity (Eqs. 68-70) shows that the loss derivative dL(z,θ*)/ds equals the influence of up-weighting a candidate point z on the regularizer derivative dR_s(θ*)/dε via Hessian symmetry. Squaring this scalar and averaging over plausible labels gives the squared loss derivative (SLD) query heuristic. The paper then proposes SGDF (Eqs. 117-125), a forward-mode automatic-differentiation extension of SGD that computes dθ*/ds alongside training with constant slowdown, and claims that SGDF generalizes the LiSSA algorithm. Additional sections sketch multi-user regularizer hierarchies, SGDF-based hyperparameter optimization, and second-order variants. The manuscript contains no query-loop experiments or test-set evaluations; Section 3.2, which was to describe additional heuristics, is removed, and Section 3.3 explicitly defers empirical comparisons.","tokens_in":26555,"tokens_out":5431,"duration_ms":51526,"significance":"If the main identity and the SGDF complexity claims hold, the paper contributes a cheap influence-type query score for large regression models and a useful formal connection between SGDF and LiSSA. The derivation of the identity is self-contained and follows from the implicit function theorem and symmetry of the Hessian, and the SGDF update is a coherent dual-number version of SGD; no constants are fitted to make the mathematics agree with data. However, the practical value of the SLD heuristic is entirely contingent on the unvalidated Hypothesis in Section 3.1, and the absence of any empirical evaluation means the central active-learning claim is not yet established.","major_comments":[{"comment":"The active-learning claim rests on the untested Hypothesis in Section 3.1: points whose model predictions vary most with regularity s are those best able to stabilize the model. No derivation connects |dL(z,θ*)/ds|² to expected reduction in generalization error, label informativeness, or any downstream objective, and Section 3.3 explicitly postpones comparisons with other query heuristics and evaluation on real data. Since the score could select high-leverage or outlier points without improving predictions, the manuscript needs at least a query-loop benchmark comparing SLD with random acquisition and standard uncertainty-based heuristics on the polynomial example and one or two real regression datasets; as written, the central claim is unvalidated.","section":"§3.1, §3.3"},{"comment":"The regularity tangent estimate requires the dual iterates ˙θ_t to converge to -H^{-1}ρ, but the text only says \"If it converges\" and gives no convergence conditions or proof. For stochastic updates with a step-size schedule and adaptive methods such as Adam, the claimed constant-slowdown estimate of -H^{-1}ρ is not guaranteed even when θ_t converges. The authors should either prove convergence under stated assumptions, such as convex quadratic losses with decaying step sizes, or provide empirical convergence diagnostics showing that ˙θ_t tracks the exact regularity tangent on the example problem.","section":"§3.5, Eqs. (114), (120), (125)"},{"comment":"The section describing additional regularity-tangent derived query heuristics has been removed from the draft, yet the abstract and conclusion refer to \"a number of ways\" of using the proposed quantities. Without the contents of Section 3.2, the set of proposed heuristics is incomplete and cannot be evaluated. The authors should either restore the section or explicitly state that only the SLD heuristic (together with the influence heuristics of Section 2.4.2) is proposed; the current pointer to a missing section is not a substitute.","section":"§3.2"}],"minor_comments":[{"comment":"In Eq. (61), the notation \"0 = df/dθ|_{θ∗=0}\" is confusing; the derivative should be evaluated at the optimum θ = θ*, not at θ* = 0.","section":"§2.4.2, Eq. (61)"},{"comment":"The last expression in Eq. (38) omits the argument θ* in the loss gradient; it should read -∂L/∂θ(z_test,θ*)⊤H^{-1}∂L/∂θ(z,θ*).","section":"§2.4, Eq. (38)"},{"comment":"The notation in Eq. (77) is ambiguous: power(x) is first described as a 6-element vector for scalar x and then as a 6-column matrix for vector x; using different symbols for the two cases would improve clarity.","section":"§3.3"},{"comment":"The proposal for joint stochastic optimization of θ and s is explicitly untested in the text and should be clearly labeled as speculative rather than as a demonstrated contribution.","section":"§3.7"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core is sound and worth publishing if the empirical gap is closed, but in its current form the paper does not substantiate the active-learning claim. If the author cannot add experiments, I would recommend rejection rather than acceptance; the observation that SGDF generalizes LiSSA may be publishable as a shorter methodological note, but the present manuscript claims a practical query-selection method that is never tested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Frederik Eaton's regularity tangent paper is a mixed bag: the math at the center is real, and the algorithmic observation is neat, but the paper sells itself as an active learning method and the evidence for that is absent.\n\nThe genuinely new thing is the identity dL(z,θ*)/ds = dR_s(θ*)/dε (Eqs. 68-70). It follows from the implicit function theorem and symmetry of the Hessian, and it gives a way to compute a data point's influence on model complexity with a single inner product after training. That's a real contribution. The SGDF algorithm — dual-number SGD that propagates dθ/ds while training — is also a reasonable forward-mode extension of SGD, and the observation that LiSSA is a special case of it (Section 3.6) is nice. The paper's literature review is honest and it's self-contained. Credit is due for the derivation.\n\nThe soft spots are exactly where the reader put them. The active learning claim rests entirely on the 'Hypothesis' in §3.1: points whose predictions vary most with regularity s are the ones best able to stabilize the model. No derivation connects |dL/ds| to expected reduction in generalization error or label informativeness. The paper's own §3.3 defers all comparisons, and §3.2's body is removed. There is no query loop, no test-set evaluation, no code. The plots are illustrations, not evidence. The SGDF dual iterate also has no convergence proof — the paper says 'if it converges' — so the constant-slowdown claim is conditional.\n\nNone of this is a demonstrated mathematical error; the identity is sound. But the central claim of an efficient active learning method is untested, and the paper openly admits it. A benchmark comparing the squared-loss-derivative query rule to random selection on a simple regression problem would settle the question, and it's missing.\n\nWho should read this? Anyone working on influence-based active learning or efficient Hessian-free influence computation will want to know about the identity and SGDF/LiSSA equivalence. But treat it as a theoretical note, not a method paper. It deserves a serious referee — the math is worth checking — but only if the referee insists on at least one experiment before acceptance. I'd send it to review, with the expectation of heavy revision or a demotion to a theory-only contribution.","headline":"Clever identity and a nice algorithmic observation, but the active learning payoff is entirely untested; worth a referee's time only if the empirical gap is filled.","tokens_in":27082,"tokens_out":2202,"would_cite":true,"duration_ms":19679,"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 single 'regularity tangent' vector, computed during training with constant slowdown, turns any candidate data point's loss gradient into an influence function on model complexity, and that squaring this score…","keywords":["active learning","influence functions","regularity tangent","regularized regression","stochastic gradient descent","forward-mode automatic differentiation","LiSSA","query heuristics"],"falsifier":"Falsify the recommendation by computing, on a regression benchmark, the squared-loss-derivative score for every unlabeled pool point, then adding labels in descending score order: if the cumulative held-out error after each query is not better than random label selection (or than a residual-based baseline), the Section 3.1 Hypothesis is false.","tokens_in":26020,"feed_emoji":"🎯","tokens_out":8450,"duration_ms":77952,"temperature":0.7,"pith_summary":"This paper proposes an active-learning rule for regression models whose query-time cost is essentially zero. The author introduces a 'regularity tangent'—the derivative of the optimal parameter vector with respect to the regularization strength—and shows it can be computed alongside the parameters during stochastic gradient descent with only a constant slowdown and one extra stored vector. Once training is finished, scoring a candidate unlabeled point is an inner product of this tangent with the point's loss gradient, so it costs the same as computing that gradient. The paper proves that this score is an influence function measuring the expected squared change in model complexity (the regularizer) that would result from adding the point, and it proposes several query heuristics built on that quantity. The practical payoff, if the paper's guiding hypothesis is right, is that large models could choose which labels to request without retraining or Hessian inversion.","feed_headline":"Active learning for regression at the cost of one inner product","feed_subtitle":"A tangent vector learned during training turns each unlabeled point's gradient into a label-worthiness score.","key_machinery":"The load-bearing object is the regularity tangent $\\frac{d\\theta^*}{ds} = -H^{-1}\\rho$, the derivative of the optimal parameter vector with respect to the regularity hyperparameter $s$, where $H$ is the Hessian of the training objective and $\\rho = \\frac{\\partial^2 R}{\\partial \\theta \\partial s}$ is the complexity gradient (for L2 regularization, $\\rho = 2\\theta^*$). The companion identity $\\frac{dL(z,\\theta^*)}{ds} = \\frac{dR_s(\\theta^*)}{d\\varepsilon}$ turns this tangent into an influence function on model complexity, so that the inner product with a point's loss gradient and the squaring of that scalar become the query score. The computational mechanism is SGDF: running SGD on dual numbers adds the update $\\dot{\\theta}_{t+1} \\leftarrow \\dot{\\theta}_t - \\eta_t(\\rho + H\\dot{\\theta}_t)$ to the ordinary parameter update, so $\\dot{\\theta}$ converges to $d\\theta^*/ds$ with the same step-size schedule as $\\theta$ and without Hessian inversion.","core_discovery":"The central claim is the identity at an optimum $\\theta^*$ of any smooth regularized objective: $\\frac{dL(z,\\theta^*)}{ds} = \\frac{dR_s(\\theta^*)}{d\\varepsilon}$, where the left side is the derivative of a point's loss with respect to the regularity hyperparameter $s$ and the right side is the derivative of the regularizer with respect to an infinitesimal up-weighting $\\varepsilon$ of that point. Because $\\frac{d\\theta^*}{ds} = -H^{-1}\\rho$, with $\\rho$ the complexity gradient of the regularizer, the left side is a single vector-inverse-Hessian-vector product $-\\sigma_z^\\top H^{-1}\\rho$. The paper takes the squared value of this scalar, averaged over plausible labels, as a query heuristic that prefers the point whose inclusion would produce the greatest expected squared change in model complexity. It then presents SGDF, a dual-number version of SGD whose updates carry $\\dot{\\theta} = d\\theta/ds$ alongside $\\theta$ and converge to $d\\theta^*/ds$ without ever forming $H^{-1}$; the LiSSA algorithm is shown to be a special case of SGDF when the regularizer update is applied at every stochastic step. An illustrative polynomial-regression experiment shows the resulting score coincides with the summed influence of all training points, a direct corollary of stationarity at the optimum.","pith_inferences":["A natural untested extension is to use the squared-loss-derivative score as the ranking function in batch or pool-based active learning, taking the top-$k$ scores each round; the paper's cost analysis suggests this would remain cheap.","The duality $dL/ds = dR/d\\varepsilon$ is more general than active learning: the same computation could be used as a data-curation diagnostic to identify which training points most affect model complexity, even when no new labels are being purchased.","Because SGDF inherits its step size from the underlying SGD schedule, it may resolve the tuning and poor-convergence issues reported for LiSSA on ill-conditioned problems; a direct comparison on a badly conditioned objective would test that.","The hierarchical regularizer models suggest a testable path to personalized large models: each user's tangent can couple sparse per-user gradients through the inverse Hessian, so the method's usefulness in federated settings could be measured without sharing raw data."],"forward_implications":["Once training is complete, scoring a candidate data point costs the same as computing its loss gradient, since only the inner product with the precomputed regularity tangent is needed.","Only one extra vector of the same dimension as the parameters is stored, so the method approximately doubles memory and adds a constant factor to training time.","The query heuristic selects points whose labels are expected to cause the largest squared change in model complexity, which is a different objective from maximizing prediction uncertainty or test-loss reduction.","SGDF produces the regularity tangent during ordinary training, making it possible to update the regularity hyperparameter $s$ stochastically from a held-out pool at the same time.","In a multi-user model with a shared base parameter vector, each user's regularity tangent inherits from the common tangent, so even a brand-new user can receive a non-random first query."],"supporting_citations":[{"why":"Defines the LiSSA algorithm for stochastic inverse-Hessian-vector products, which the paper proves is a special case of SGDF.","marker":"[1]"},{"why":"Provides the influence-function definitions and up-weighting formalism that the paper adapts to the regularizer and regularity tangent.","marker":"[2]"},{"why":"Classical influence-function monograph that frames the notion of parameter sensitivity the paper generalizes.","marker":"[9]"},{"why":"Earlier active-learning method using per-point influence functions, the computational baseline SGDF improves on.","marker":"[10]"},{"why":"Scaling-up influence functions; source of the reported tuning and convergence difficulties that SGDF's step-size inheritance may address.","marker":"[11]"},{"why":"Earlier influence-based active learning approach that motivates comparing against a reliability-oriented selector.","marker":"[12]"},{"why":"Prior 'expected model change' active learning heuristic, which the paper contrasts with its regularizer-based notion of model change.","marker":"[13]"},{"why":"Supplies the generalized cross-validation trace formula behind the paper's self-influence approximation $G_{\\mathrm{pert}}$.","marker":"[15]"}],"fun_headline_variants":["Active learning for regression: one inner product, full influence score","Curiosity via regularity tangents: active learning at zero extra cost","A tangent vector for active learning: same cost as loss gradient","Influence functions for free: active learning from a training tangent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the Section 3.1 Hypothesis that the data points most able to stabilize a model are precisely those whose predictions vary most when the regularity hyperparameter is varied; if that premise is false, the squared-loss-derivative score could be no better than random label selection.","fun_headline_variants_meta":{"raw":{"variants":["Active learning for regression: one inner product, full influence score","Curiosity via regularity tangents: active learning at zero extra cost","A tangent vector for active learning: same cost as loss gradient","Influence functions for free: active learning from a training tangent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000965,"raw_usage":{"total_tokens":4176,"prompt_tokens":1080,"completion_tokens":3096,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":3024}},"tokens_in":696,"tokens_out":3096,"duration_ms":21315,"temperature":1.0,"reasoning_tokens":3024,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:29:30.457019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Falsify the recommendation by computing, on a regression benchmark, the squared-loss-derivative score for every unlabeled pool point, then adding labels in descending score order: if the cumulative held-out error after each query is not better than random label selection (or than a residual-based baseline), the Section 3.1 Hypothesis is false.","supporting_citations":[{"cited_title":"Second-Order Stochastic Optimization for Machine Learning in Linear Time, 2017","cited_arxiv_id":null,"evidence_quote":"Defines the LiSSA algorithm for stochastic inverse-Hessian-vector products, which the paper proves is a special case of SGDF."},{"cited_title":"Understanding Black-box Predictions via Influence Functions, 2017","cited_arxiv_id":null,"evidence_quote":"Provides the influence-function definitions and up-weighting formalism that the paper adapts to the regularizer and regularity tangent."},{"cited_title":"Residuals and Influence in Regression, 1982","cited_arxiv_id":null,"evidence_quote":"Classical influence-function monograph that frames the notion of parameter sensitivity the paper generalizes."},{"cited_title":"Influence Selection for Active Learning, 2021","cited_arxiv_id":null,"evidence_quote":"Earlier active-learning method using per-point influence functions, the computational baseline SGDF improves on."},{"cited_title":"Scaling Up Influence Functions, 2021","cited_arxiv_id":null,"evidence_quote":"Scaling-up influence functions; source of the reported tuning and convergence difficulties that SGDF's step-size inheritance may address."},{"cited_title":"Reliable Active Learning via Influence Functions, 2023","cited_arxiv_id":null,"evidence_quote":"Earlier influence-based active learning approach that motivates comparing against a reliability-oriented selector."},{"cited_title":"Maximizing Expected Model Change for Active Learning in Regression, 2013","cited_arxiv_id":null,"evidence_quote":"Prior 'expected model change' active learning heuristic, which the paper contrasts with its regularizer-based notion of model change."},{"cited_title":"Generalized Cross-Validation as a Method for Choosing a Good Ridge Parameter, 1979","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized cross-validation trace formula behind the paper's self-influence approximation $G_{\\mathrm{pert}}$."}],"review_version":1}