{"id":"421e2b95-a308-4675-bb03-d8e632cc29fa","arxiv_id":"2608.03249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Setting the entropic regularization of a balanced Sinkhorn plan to a constant times the mean closest-anchor cost of k-means anchors yields a cold-start active learning method that outperforms and unifies typicality-, coverage-, and diversity-based selectors.","lead":"Three major cold-start active learning methods, based on typicality, coverage, and diversity, turn out to be instances of one optimal-transport allocation program with a single smoothing knob. The paper derives a rule for setting that knob from unlabeled data alone, and the resulting method, called epsilon-AS, beats nine baselines in 25 of 26 settings across six image datasets.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Universal constant c=0.4 is not verified to lie in the theory's [t_minus,t_plus] window on any target dataset; the minimax-rate guarantee and the SOTA claim both depend on this unverified transfer.","rationale":"The reader's weakest assumption points to the unverified universality of c=0.4, and this is indeed the single most load-bearing concern for the paper's central claim. The theoretical path to the minimax rate (Theorem 3, final statement) explicitly requires c ∈ [t_-, t_+] under Assumption B.6; without that, the decoder-fidelity condition Q_P(S) ≤ κ_q m_P(b) is not guaranteed, so the upper bound in Theorem 2 does not apply. The empirical SOTA claim also depends on c being transferable, since all reported ε-AS results use c=0.4. The paper provides no evidence that 0.4 lies inside the certified window on any target dataset, and the calibration protocol (Section VI) uses only MNIST labels. The proposed test directly measures the quantities that Assumption B.6 controls—the anchor-to-query displacement and the resulting quantization error normalized by m_P(b)—and would settle whether the transfer assumption actually holds. The paper is otherwise careful: Theorem 1 is an exact finite-pool subsumption, Theorem 2 has a clean lower-bound argument, and the appendix includes logical-dependency tables and acknowledges the calibration limitation. Thus the reader's CONDITIONAL verdict is appropriate; my stress-test does not require moving the verdict.","tokens_in":42196,"tokens_out":4963,"duration_ms":52454,"concrete_test":"For each of the 22 transfer cells (5 target datasets × budgets), compute the decoded set S produced by ε-AS with c=0.4. Estimate m_P(b) from the pool (e.g., by the empirical closest-anchor cost bm(b) after k-means), and compute the normalized quantization ratio Q_P(S)/m_P(b) = (1/n Σ_i min_{s∈S} C(z_i,s)) / m_P(b). Also compute the anchor-to-query displacement term in Assumption B.6, D(t) = Σ_k p_k(bµ) C(bµ_k, s_k(t)) / m_P(b), at t=0.4. If Q_P(S)/m_P(b) exceeds a small universal constant (e.g., > 2(1+κ_d)) or D(0.4) grows with b, then c=0.4 is not in the certified window and the minimax guarantee does not apply to the deployed configuration. This test directly checks the load-bearing assumption without requiring labels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central rate claim (Theorem 3, final statement; Appendix D.4) holds only if the calibrated multiplier c lies in the per-distribution window [t_-, t_+] of Assumption B.6. The paper selects c=0.4 once by grid search on MNIST labels (Section VI, Calibration) and freezes it across all other datasets and budgets. No experiment checks whether 0.4 falls inside [t_-, t_+] for CIFAR-10, CIFAR-100, STL-10, Caltech-101, or ImageNet-1k. If the true optimal c varies with pool geometry and budget, then ε⋆ = 0.4·bm(b) is mis-scaled; the anchor-to-query displacement in Assumption B.6 is not controlled, the quantization-faithfulness condition Q_P(S) ≤ κ_q m_P(b) fails, and the minimax rate no longer follows. The empirical SOTA claim also relies on this same c transfer: the sensitivity analysis (Figure 3) only varies c on STL-10 and Caltech-101, not on the larger datasets, and the observed variation is within a small range. The paper itself lists 'eliminating the remaining global calibration' as future work, acknowledging that the universality of c is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a unified optimal-transport (OT) view of cold-start active learning (CSAL). It defines a transport-selection realization that exactly subsumes TypiClust, ProbCover, and ActiveFT (Theorem 1), establishes a task-agnostic minimax bound governed by the closest-anchor quantization cost mP(b) (Theorem 2), and derives a Sinkhorn-based selector, ε-AS, with data-adaptive temperature ε* = c·bm(b) (Theorem 3). Experiments on MNIST, CIFAR-10/100, STL-10, Caltech-101, and ImageNet-1k report the highest point estimate in 25 of 26 settings, a 1.29% average improvement over ActiveFT on ImageNet-1k, and a 2.3× selection-time speedup.","tokens_in":42467,"tokens_out":9275,"duration_ms":106945,"significance":"The unification is conceptually valuable: it reframes the choice among typicality/coverage/diversity methods as the choice of a single allocation scale, and the minimax connection to mP(b) gives a principled reason to adapt that scale to the unlabeled pool. The finite-pool subsumption identities are exact and carefully documented, and the appendix is unusually complete. However, the headline claims are not yet fully supported: a supporting inequality in Lemma D.5 is false as stated, the rate guarantee for ε-AS requires c to lie in an unspecified window [t−, t+] that is never verified for any target dataset, and the empirical superiority claim rests on three-seed point estimates with overlapping error bars. These issues are correctable, but they affect the paper's central claims.","major_comments":[{"comment":"The proof of the anchor-to-query transfer bound uses d(z,s_k)^2 ≤ d(z,µ_k)^2 + d(µ_k,s_k)^2. This is false in general Euclidean geometry (e.g., collinear points with both distances equal to 1 give left side 4 and right side 2). The correct inequality is d(z,s_k)^2 ≤ 2d(z,µ_k)^2 + 2d(µ_k,s_k)^2, which changes Lemma D.5 to QP(S) ≤ 4RP(µ) + 4Σ pk C(µk,sk), and the decoder-fidelity constants in Theorem 2 (κq > 2(1+κd)) become κq > 4(1+κd). The minimax rate survives, but the proof as written is incorrect and should be corrected.","section":"Appendix D.3, Lemma D.5"},{"comment":"The main-text Theorem 3 states that ε-AS attains the minimax rate 'with decoder fidelity,' omitting the appendix's condition that c must lie in the normalized-temperature window [t−, t+] of Assumption B.6. The experimental protocol selects c = 0.4 by grid search on MNIST labels (Section VI, Calibration) and never checks whether 0.4 lies in [t−, t+] on CIFAR-10, CIFAR-100, STL-10, Caltech-101, or ImageNet-1k. Since the canonical calibration game's ccan depends on dimensionless weights (A, λ, κ) that are not tied to data, the universality of c = 0.4 is not established. Please add a label-free diagnostic of the Assumption B.6 window (e.g., estimate supt QP(S_t bm)/mP(b) per dataset and budget) or weaken the claim to 'conditional on c in the theoretical window.'","section":"Theorem 3 and Assumption B.6 (Appendix D.4)"},{"comment":"The headline '25 of 26 settings' is based on three-seed point estimates with widely overlapping standard deviations (e.g., STL-10 b=200: ε-AS 97.48±0.88 vs. ActiveFT 97.03±1.06; CIFAR-100 b=100: 69.77±1.66 vs. ActiveFT 67.76±1.46). No significance test or paired bootstrap is reported, so 'consistently achieves state-of-the-art' is not statistically supported as written. In addition, the c-sensitivity analysis (Figure 3) is limited to STL-10 and Caltech-101; the stability of the transferred constant on CIFAR-10/100 and ImageNet-1k is not demonstrated.","section":"Section VI, Tables E.3–E.8"}],"minor_comments":[{"comment":"εmin is an absolute floor in the algorithm, but exact cost-scale equivariance requires the floor to be rescaled together with the cost. Please specify in Algorithm 1 that εmin should scale with bm(b), or state the asymptotic condition εmin = o(mP(b)) in the algorithm's input description.","section":"Algorithm 1, line 4"},{"comment":"The 'mixture consistency' axiom of Definition D.6 applies only to cost lists evaluated under a common, fixed anchor and cost system. This qualification appears only in Appendix D.4; it should be stated where the rule is introduced in the main text.","section":"Section V, Eq. (7)"},{"comment":"The number of seeds (three) is stated in the experimental setup but not in the figure caption or the main tables. Please add it to the caption/table headings for clarity.","section":"Figure 2 and Tables E.3–E.8"},{"comment":"The 'exact subsumption' claim is precise only for the evaluated specializations (e.g., deterministic no-repeat ProbCover and K=b TypiClust). Please state this qualification directly in the theorem statement, as it is currently only implicit in the appendix.","section":"Theorem 1 statement"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong conceptual contribution and unusually detailed appendices. The main risks are the false inequality in Lemma D.5 (fixable with a factor of 2, but currently invalid), the unverified placement of the calibrated constant inside the theoretical window, and the reliance on 3-seed point estimates for the empirical SOTA claim. I believe these are fixable within the manuscript's scope, so I recommend major revision rather than rejection. Editorially, the paper may benefit from a more cautious framing of 'task-agnostic' given that c = 0.4 is selected using MNIST labels."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is Theorem 1 plus Theorem 2. Recasting TypiClust, ProbCover, and ActiveFT as exact instances of one transport–selection program is a genuinely new and clarifying move, and the minimax bound tied to the closest-anchor cost is clean. The appendix deserves credit: full proofs, exact finite-pool identities, and a logical-dependency table. This is not a paper that hides its machinery.\n\nThe soft spots are real but narrower than a first read suggests. The empirically deployed c=0.4 is selected once on MNIST and frozen everywhere. The theory only guarantees the minimax rate for c inside an unspecified window [t_minus, t_plus] (Assumption B.6), and the paper never checks that 0.4 lies in that window on any target dataset. The sensitivity analysis varies c only on STL-10 and Caltech-101, not on CIFAR-100 or ImageNet-1k. If the optimal c shifts with pool geometry, epsilon* is mis-scaled and both the rate claim and the SOTA claim weaken. The authors acknowledge the remaining global calibration as future work, so this is not hidden, but it is load-bearing for the empirical universality claim.\n\nThe empirical support is thinner than the abstract implies: three seeds, overlapping standard deviations in several cells (e.g., STL-10 at b=200: 97.48±0.88 vs ActiveFT 97.03±1.06), and code not yet released. Twenty-five-out-of-twenty-six point estimates is suggestive, but the margin is often within noise. The circularity burden is low, as the reader says: the rate claim is independent of the fitted constant, and only the uniqueness theorem is partly axiomatic.\n\nWho gets value from this? Anyone working on cold-start active learning, data subset selection, or transport-based sampling. The unification is useful even if the universal-knob claim needs more evidence. It deserves a serious referee. I would send it to review with a request for more seeds, a check of the c window on the larger datasets, and code release. The central framework is solid enough to engage with; the SOTA claim just needs to be brought down a notch.","headline":"A genuinely useful unification of cold-start active learning with a clean minimax bound, but the universal constant c is an unverified load-bearing assumption and the empirical dominance rests on thin seeds.","tokens_in":43024,"tokens_out":1705,"would_cite":true,"duration_ms":21146,"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":"The paper claims that typicality-, coverage-, and diversity-based cold-start selectors are one transport program in disguise, and that one data-tuned temperature, $\\varepsilon^* = c\\hat{m}(b)$, makes the Sinkhorn selector minimax-rate-optim","keywords":["cold-start active learning","optimal transport","Sinkhorn scaling","entropic regularization","minimax rate","closest-anchor cost","intrinsic dimension","task-agnostic selection"],"falsifier":"Re-run the calibration grid ($c \\in \\{0.1, 0.2, \\dots, 0.9\\}$) on labels from a target dataset such as CIFAR-100 or ImageNet-1k: if the best-performing $c$ is far from $0.4$ there, the claim that one constant transfers across pool geometries fails. A direct check of the decoder-fidelity assumption is also possible: compute the probability-weighted anchor-to-query displacement $\\sum_k p_k(\\hat{\\mu})\\, C(\\hat{\\mu}_k, s_k(t)) / m_P(b)$ on each pool; if it exceeds the theory's $\\kappa_q$ bound at $c = 0.4$, the decoded subset is not quantization-faithful and the minimax-rate conclusion does not fo","tokens_in":42021,"feed_emoji":"🎯","tokens_out":12669,"duration_ms":120898,"temperature":0.7,"pith_summary":"Cold-start active learning must pick the first annotation batch with no labels and no trained model, forcing each existing method to commit to a fixed geometric bias — typicality, coverage, or diversity — that wins on some datasets and fails on others. The paper's claim is that those three routes are all the same allocation program viewed through different lenses, and that the only quantity needing adaptation is the entropic regularization strength $\\varepsilon$ of a balanced Sinkhorn coupling. The theory pins the right scale: $\\varepsilon$ should equal a dimensionless constant times the empirical mean closest-anchor cost, the unique scale-equivariant statistic, and with this choice the selector attains the minimax rate $\\Theta(b^{-2/d_{\\mathrm{int}}})$ — optimal label efficiency in the budget $b$ given the intrinsic dimension $d_{\\mathrm{int}}$ of the feature manifold. Empirically the one rule, $\\varepsilon$-AS, ranks first in 25 of 26 dataset-budget settings across six datasets and beats the strongest fixed-rule baseline on ImageNet-1k by 1.29% average accuracy while selecting 2.3$\\times$ faster. If true, cold-start selection needs no dataset-specific heuristic: unlabeled geometry alone sets the allocation scale.","feed_headline":"One data-tuned knob wins 25 of 26 cold-start label tests","feed_subtitle":"Typicality, coverage, and diversity collapse into one transport program whose only dial is set by the unlabeled pool.","key_machinery":"The central object is the balanced entropic transport program $J^R_\\varepsilon(\\theta,\\pi) = \\langle\\pi, C_\\theta\\rangle + \\varepsilon\\,\\mathrm{KL}(\\pi \\| a\\otimes u) + \\Omega^R(\\theta)$, together with a realization $R = (\\Theta^R, \\Pi^R, C^R, \\Omega^R, D^R)$ fixing the allowed representatives, allocation constraints, cost, support energy, and decoder. Theorem 1 shows this one program exactly subsumes TypiClust, ProbCover, and ActiveFT, converting a choice among methods into a choice of one scalar $\\varepsilon$. The carrying identity is the adaptive temperature $\\varepsilon^* = c\\,\\hat{m}(b)$, where $\\hat{m}(b)$ is the empirical mean closest-anchor cost; cost-scale equivariance of Sinkhorn p","core_discovery":"The paper's central claim is that three seemingly rival cold-start strategies — typicality (TypiClust), coverage (ProbCover), and diversity (ActiveFT) — are exact realizations of one transport-selection program, differing only in representatives, cost, and decoder. Its balanced entropic branch defines $\\varepsilon$-AS, whose only tunable quantity is the regularization strength $\\varepsilon$. The paper proves that $\\varepsilon^* = c\\,\\hat{m}(b)$, a dimensionless constant times the mean closest-anchor cost, is the unique scale-equivariant rule, and that it attains the task-agnostic minimax rate $\\Theta(b^{-2/d_{\\mathrm{int}}})$: label-optimal as budget $b$ grows, with the exponent set by the m","pith_inferences":["If $c$ is truly universal, the same frozen constant should transfer to other frozen feature extractors without recalibration, as long as the pool geometry has a well-defined intrinsic dimension — a direct test the paper does not run.","The axioms that single out the mean closest-anchor cost explicitly rule out robust statistics such as the median or trimmed mean; on pools with heavy-tailed closest-anchor costs, comparing $\\varepsilon^* = c\\hat{m}(b)$ against a median-based rule would probe whether the mean is load-bearing.","The unification relocates method design to the decoder: any new decoding rule inside the same balanced entropic branch inherits the proven minimax rate, which could yield cheaper or parallelizable selection schemes.","Extending the canonical two-route calibration game with a third route, such as a coverage-fidelity term, would produce a calibration surface $c(A, \\lambda, \\kappa)$ that could replace the single MNIST-fixed constant and make the one-knob claim fully data-driven."],"forward_implications":["A user no longer needs to choose among typicality, coverage, or diversity: $\\varepsilon$-AS derives the balance from the unlabeled pool and the budget alone.","Label efficiency is governed by the feature manifold's intrinsic dimension: minimax risk decays as $b^{-2/d_{\\mathrm{int}}}$, so pools with larger intrinsic dimension need larger budgets for the same target error.","The budget-dependent error floor is characterized exactly: no task-agnostic selector can beat $\\tfrac{L^2}{2} m_P(b)$, and the floor is attainable by any quantization-faithful selector.","The scale rule $\\varepsilon^* = c\\hat{m}(b)$ is the unique statistic consistent with the paper's axioms, so future selectors built inside the same framework inherit the same adaptive temperature.","On ImageNet-1k, $\\varepsilon$-AS beats the strongest fixed-rule baseline by 1.29% average accuracy while cutting selection time from 55.4 to 24.3 seconds (2.3$\\times$)."],"supporting_citations":[{"why":"Supplies TypiClust, the typicality-route selector whose objective and selection rule Theorem 1 exactly subsumes.","marker":"Hacohen, Dekel, and Weinshall 2022"},{"why":"Supplies ProbCover, the coverage-route selector subsumed as a truncated-cost greedy realization.","marker":"Yehuda et al. 2022"},{"why":"Supplies ActiveFT, the diversity-route objective subsumed as a semi-relaxed realization and the main empirical baseline to beat.","marker":"Xie et al. 2023"},{"why":"Supplies entropic regularization with Sinkhorn scaling, the solver all of $\\varepsilon$-AS's balanced couplings use.","marker":"Cuturi 2013"},{"why":"Supplies the frozen feature backbone whose normalized geometry defines the closest-anchor cost used in every experiment.","marker":"Oquab et al. 2024"},{"why":"Supplies the intrinsic-dimension estimator that motivates the scaling assumption $m_P(b) = \\Theta(b^{-2/d_{\\mathrm{int}}})$.","marker":"Levina and Bickel 2004"},{"why":"The computational-optimal-transport reference for the entropic program, cost geometry, and Sinkhorn theory the framework builds on.","marker":"Peyré and Cuturi 2019"},{"why":"Supplies near-linear-time Sinkhorn convergence guarantees invoked for the computational claims.","marker":"Altschuler, Niles-Weed, and Rigollet 2017"}],"fun_headline_variants":["One transport knob unifies cold-start active learning","Adaptive transport dial masters cold-start labeling","Single knob: typicality, coverage, diversity unified","Data-tuned transport wins cold-start label tests","One knob, set by data, outperforms all cold-start heuristics"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The theory certifies the minimax rate only for a calibrated constant $c$ inside an unspecified window $[t_-, t_+]$; the paper fixes $c = 0.4$ by one grid search on MNIST labels and never verifies that $0.4$ lies inside that window on any other dataset or budget.","fun_headline_variants_meta":{"raw":{"variants":["One transport knob unifies cold-start active learning","Adaptive transport dial masters cold-start labeling","Single knob: typicality, coverage, diversity unified","Data-tuned transport wins cold-start label tests","One knob, set by data, outperforms all cold-start heuristics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000549,"raw_usage":{"total_tokens":2493,"prompt_tokens":816,"completion_tokens":1677,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":1601}},"tokens_in":560,"tokens_out":1677,"duration_ms":16792,"temperature":1.0,"reasoning_tokens":1601,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:34:49.875355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the calibration grid ($c \\in \\{0.1, 0.2, \\dots, 0.9\\}$) on labels from a target dataset such as CIFAR-100 or ImageNet-1k: if the best-performing $c$ is far from $0.4$ there, the claim that one constant transfers across pool geometries fails. A direct check of the decoder-fidelity assumption is also possible: compute the probability-weighted anchor-to-query displacement $\\sum_k p_k(\\hat{\\mu})\\, C(\\hat{\\mu}_k, s_k(t)) / m_P(b)$ on each pool; if it exceeds the theory's $\\kappa_q$ bound at $c = 0.4$, the decoded subset is not quantization-faithful and the minimax-rate conclusion does not fo","supporting_citations":[{"cited_title":"Active Finetuning: Exploiting Annotation Budget in the Pretraining-Finetuning Paradigm","cited_arxiv_id":null,"evidence_quote":"Supplies ActiveFT, the diversity-route objective subsumed as a semi-relaxed realization and the main empirical baseline to beat."},{"cited_title":"Maximum Likelihood Estimation of Intrinsic Dimension","cited_arxiv_id":null,"evidence_quote":"Supplies the intrinsic-dimension estimator that motivates the scaling assumption $m_P(b) = \\Theta(b^{-2/d_{\\mathrm{int}}})$."}],"review_version":1}