{"id":"a6e0eead-cb49-4abb-9576-25d903c973da","arxiv_id":"2412.07980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TTVD is a new test-time adaptation loss built from Voronoi diagram geometries that improves error and expected calibration error on standard corruption benchmarks.","lead":"This paper proposes TTVD, a test-time adaptation method that uses Voronoi diagram structures, including cluster-induced Voronoi diagrams and power diagrams, to compute adaptation losses. The authors report improved accuracy and calibration over existing methods on CIFAR and ImageNet corruption benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CIPD loss as specified is internally inconsistent: Algorithms 2/3's β(−F) inverts the influence-maximizing partition defined in Eqs. 4/6, and (d²−v²)^−0.8 is undefined where d²<v², so Table 1 is not reproducible from the text.","rationale":"The reader's verdict (CONDITIONAL) centered on the unverified modeling premise of γ=−0.8 and τ=1. I found a sharper problem in the same equations that the reader did not name: the published method is internally inconsistent, which changes the verdict from 'fixable with experiments' to 'cannot be evaluated as written.' My reasoning: (a) Definition 3.2/3.4 and Algorithms 2/3 disagree on the sign of the softmax argument, so the entropy loss's gradient direction is opposite to the claimed alignment with Voronoi cells; this holds for both signs of γ, so it is not a tuning issue. (b) Eq. 6 is undefined for negative power distances with γ=−0.8 and singular on the boundaries that the noise-filtering story depends on. (c) The §3.3 noise-filtering mechanism (diagram subtraction, sample exclusion) never appears as an explicit step in Algorithm 3, so the 2.2% CIPD gain attributed to filtering cannot be traced. Each of these is a specification gap in the central construction, not a matter of benchmark choice. Credit where due: the geometric framing is genuinely interesting and the experiments are well-scoped under TTAB protocols; the VD→CIVD→CIPD ablation, the class-mean precision study (Table 4), and the batch-size/label-shift analyses are informative and would likely survive a re-run. The concern is specifically that the central loss, as printed, is not a well-defined adaptation algorithm. A one-line sign correction (β(+F)) plus a convention for negative power distances would likely repair the specification, and releasing code would settle everything. I therefore recommend UNVERDICTED rather than REJECT: the claim is currently neither confirmed nor refuted, and the paper's own text supplies the evidence that it cannot be checked as-is. If the authors release code and correct the equations, an ACCEPT or CONDITIONAL verdict would be appropriate after error bars are added to the headline improvements.","tokens_in":22153,"tokens_out":14504,"duration_ms":148368,"concrete_test":"Implement TTVD strictly from the text — Eqs. 3, 4, 6, Algorithms 1–3, γ=−0.8, τ=1, learning rates grid-searched as in Appendix D — and run the CIFAR-10-C Gaussian-noise stream with ResNet-26, comparing the error trajectory against Table 5's reported 27.4% for CIPD. If the literal implementation improves over no adaptation and reaches roughly 27.4%, the printed β(−F) must conceal an unspecified sign flip or re-normalization; if it degrades during adaptation or produces NaN whenever a feature lies inside a power cell (d²<v²), the published algorithm is not the evaluated algorithm. The test is decisive because it distinguishes 'the described method reproduces the headline numbers' from 'the text conceals the actual convention,' which is the only way the central claim can be checked without code.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The load-bearing concern is that the published specification of TTVD cannot be the algorithm that produced the reported results. Sign inconsistency: Definition 3.2/3.4 defines F(z,C_k) = −sign(γ) Σ_α d(μ_k^(α), z)^γ as an influence to be maximized for cell assignment (r(z) = arg max_k F). Under the announced γ = −0.8, F = +Σ_α d^−0.8, which is largest for the nearest cluster, exactly the intended semantics. Algorithms 2/3, however, compute the soft label as β(−F(z,C_k)+ε; τ), i.e., logits proportional to −Σ d^−0.8. As a feature point approaches its assigned site (d→0), its logit → −∞ and its soft-label probability → 0, while distant clusters receive roughly equal share; the entropy loss L_VD then rewards moving features away from the cluster that the partition rule assigns them to, the opposite of the mechanism claimed in §3.1–3.2. Eq. 3 (plain VD) is internally consistent because distance d is a cost, but negating F conflicts with F being an influence for any sign of γ (for γ>0, −F = +Σ d^γ also peaks at the farthest site). The equations and the algorithm blocks cannot both be right. Ill-defined exponent: Eq. 6 raises the signed power distance to γ=−0.8; {d²−v²}^−0.8 is non-real whenever d²<v² (inside the power circle) and singular at d=v, with no stated clipping, absolute-value, or sign handling, so the CIPD loss is undefined over a large portion of feature space. Unverifiable central claim: with no released code, Tables 1 and 5–8 cannot be tied to a self-consistent method, so the claimed SOTA error/ECE rests on a specification that, read literally, inverts the intended gradient direction and is ill-defined at the very boundaries the method claims to exploit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TTVD, a test-time adaptation method framed through computational geometry. It first observes that neighbor-based TTA methods correspond to a Voronoi Diagram, then generalizes to a Cluster-induced Voronoi Diagram (CIVD) whose influence function aggregates multiple sites per class, and further to a Cluster-induced Power Diagram (CIPD) with per-site weights for boundary flexibility and noisy-sample filtering. The method is evaluated with the TTAB toolkit on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R, reporting the lowest classification error and expected calibration error among the compared methods, plus additional experiments on batch size, label shift, and robustness of the class-mean estimates. The central claim is that the geometric formulation unifies self-supervision and entropy minimization while avoiding negative transfer, and that CIPD's flexible boundaries improve subsample selection.","tokens_in":22570,"tokens_out":7264,"duration_ms":68335,"significance":"If the claims held, TTVD would be a meaningful contribution to TTA: it is evaluated on an externally maintained benchmark (TTAB), it compares against a broad set of baselines, and it reports consistent improvements in both error and calibration across four benchmarks. The paper also shows robustness to reduced precision in the class means, which is practically useful. However, the significance is currently undermined by specifying inconsistencies that make the published algorithm not reproducible from the text: the sign flip in the soft-label computation, the undefined fractional power in Eq. (6), and the absence of the filtering mechanism that is central to the CIPD variant. These do not refute the geometric idea, but they prevent the reader from verifying the reported results and need to be corrected and supplemented with code or precise algorithmic details.","major_comments":[{"comment":"The soft-label computation in Algorithms 2 and 3 inverts the influence function defined in Definitions 3.2 and 3.4. With γ = −0.8, F(z,C_k) = Σ_α d(μ_k^(α),z)^{−0.8} because −sign(γ)=+1, so the cell assignment r(z)=arg max_k F selects the cluster with the smallest aggregate distance, as intended. However, the algorithms compute y~_k = β(−F(z,C_k)+ε; τ), so the logit for the assigned (nearest) cluster is the most negative and its soft-label probability tends to 0 as d→0. Consequently, the entropy loss L_VD pushes features away from the very site that the partition rule assigns them to, contradicting the mechanism described in §3.1–§3.2 and the sample analysis in Appendix A.1, which relies on F without the minus sign. For γ>0 the same inversion occurs because −sign(γ)Σd^γ is minimized at the nearest site. The equations and the algorithm blocks cannot both be correct; either the minus sign in Algorithms 2/3 is a typo (it should be +F) or the reported results do not come from the specified algorithm. This must be corrected and verified by released code.","section":"§3.2, Eq. (4), Eq. (6), Algorithms 2–3 (Appendix H)"},{"comment":"The Cluster-induced Power Diagram influence F(z,C_k) = −sign(γ) Σ_α {d(μ_k^(α),z)^2 − v_k^2}^γ is not well-defined for the announced γ = −0.8. The quantity d^2 − v^2 is negative for points inside the power circle of site k, and the fractional power of a negative number is complex; it is also singular at d^2 = v^2. The paper gives no clipping, absolute-value, sign-preserving power, or restriction on the domain, so the CIPD loss and its gradients are undefined over a substantial part of feature space. Because Table 2 credits CIPD with a further 2.2% improvement over CIVD, the evaluation of this core variant is not reproducible from the text. Please specify the exact definition used in practice (e.g., replacing the base by |d^2 − v^2| with sign, or clipping at a small positive value) and justify it geometrically.","section":"Definition 3.4, Eq. (6)"},{"comment":"The sample-filtering mechanism attributed to the Power Diagram is not implemented in the algorithm that is claimed to define TTVD. §3.3 describes subtracting the VD from the PD to detect noisy samples near boundaries, but Algorithm 3 only computes soft labels via Eq. (6) and applies the entropy loss; there is no step that excludes samples, no threshold, and no description of how the diagram subtraction is performed in feature space. Moreover, the PD weights v_k are never specified: Lemma 3.1 connects logistic-regression parameters to a power diagram, but it is unclear whether v_k are taken from the classifier biases, are learned, or are set to constants. Without these details, the 2.2% CIPD improvement over CIVD in Table 2 cannot be attributed to the proposed filtering, and the method cannot be re-implemented.","section":"§3.3, Algorithm 3, Table 2"},{"comment":"The influence exponent γ = −0.8 and temperature τ = 1 are fixed hyperparameters, but no sensitivity analysis, ablation, or derivation is provided. The behavior of the influence function depends critically on the sign and magnitude of γ; for instance, sign(γ) determines whether the influence is attractive or repulsive, and the power-law exponent determines how quickly distant sites are discounted. Since TTVD's claims rest on this choice and the value is not selected by a principled criterion, the paper should report performance over a range of γ (and τ) and justify the chosen value. This is particularly important because the reported gains over strong baselines are modest (0.7–1.6% error), and the optimal γ may vary across datasets.","section":"§4.2, Appendix D"},{"comment":"No error bars or repeated-run statistics are reported for any of the results. The claimed improvements are small on several benchmarks (e.g., 0.7% on CIFAR-100-C and ImageNet-R), and TTA results are sensitive to the order of the online stream, batch composition, and random seeds. Without confidence intervals or at least multiple seeds, the statistical significance of the main empirical claim cannot be assessed. Please report mean and standard deviation over at least three runs, or otherwise justify that the reported differences are reproducible.","section":"§4.2, Tables 1, 3, 5–12"}],"minor_comments":[{"comment":"There are numerous typographical and formatting issues: 'V oronoi' with a stray space appears repeatedly, 'inherit limitation' should be 'inherent limitation', 'iterpretability' should be 'interpretability', 'Internetional' should be 'International' (Appendix F), and 'Congugate' should be 'Conjugate' (Appendix F header). A copy-edit pass is needed.","section":"Throughout"},{"comment":"Lemma 3.1 is attributed to references [36,37] (DeepVoro and iVoro), but the logistic-regression power-diagram identity originates in the analysis of deep networks by Balestriero et al. (reference [5] in the bibliography). Please correct the citation.","section":"Lemma 3.1"},{"comment":"References [21] and [22] are the same publication (Influence-based Voronoi diagrams of clusters, Computational Geometry 96:101746, 2021) and are listed twice. In addition, two non-identical entries for the pyvoro software [38,45] appear. Please deduplicate.","section":"References"},{"comment":"The claim that CIVD 'avoids the negative transfer' is not directly supported by the experiments; no measurement or comparison of negative transfer is reported. Please either soften the claim or provide evidence.","section":"§3.2"},{"comment":"The caption of Figure 1 refers to an 'Augmented Voronoi Diagram (by self-supervision)' while the text calls this CIVD. Align the terminology between the figure and the main text.","section":"Figure 1, Appendix C"},{"comment":"The sample-level analysis in Appendix A.1 shows only three hand-picked examples. It would be strengthened by reporting aggregate statistics over the misclassified set (e.g., what fraction of CIVD corrections are explained by the rotated views agreeing with the ground-truth class).","section":"Appendix A.1"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency between Algorithms 2/3 and Eqs. (4)/(6) is the kind of issue that suggests the manuscript was not checked against a runnable implementation. Since no code is released, I recommend that the editor require the authors to release code and to confirm which objective was actually optimized. If the algorithm used +F, the text must be corrected; if −F was used, the method would be anti-correlated with the claimed mechanism and the results in Table 1 would be highly suspect. This is my primary reason for major_revision rather than accept. The paper's scope fits the journal, and the geometric framing is interesting, but the reproducibility concerns must be resolved before it can be recommended for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a genuinely interesting geometric angle on test-time adaptation, but the method as specified cannot produce the reported numbers. The sign convention in the algorithms contradicts the definition of the influence function, and the power-distance exponent is undefined over much of the feature space.\n\nWhat's new: framing neighbor-based TTA as Voronoi diagram partition, and proposing to use Cluster-induced Voronoi Diagrams (CIVD) and Power Diagrams (PD) to derive a loss. That's a fresh perspective in a crowded field. The experiments are run on the TTAB harness, which is good practice, and the ablation ladder (VD → CIVD → CIPD) shows a coherent story. The improvements in Table 1 are substantial if true.\n\nThe soft spots are not minor. Definition 3.2 defines F as an influence to be maximized for cell assignment (arg max), and Eq. 4 with γ=-0.8 makes F = Σ_d d^{-0.8}, so larger F means closer to the cluster. But Algorithms 2 and 3 compute the soft label as β(-F + ε; τ). That flips the sign: a feature point approaching its assigned site gets logits → -∞ and probability → 0. The entropy loss then rewards moving away from the cluster, the exact opposite of the claimed mechanism. Eq. 3 for the plain VD is consistent because distance is a cost, but negating F in the CIVD/CIPD cases inverts the intended gradient direction.\n\nSecond, Eq. 6 raises (d² - v²) to the power γ = -0.8. That is non-real whenever d² < v² (inside the power circle) and singular at d = v. No clipping, absolute value, or sign handling is stated. So the CIPD loss is undefined on a significant part of the feature space. The sample-filtering mechanism described in Sec 3.3 is also not implemented in Algorithm 3, which just uses Eq. 6. There are no error bars, and the key hyperparameter γ is chosen by grid search rather than derived. Combined with the absence of released code, I see no way to tie Tables 1 and 5–8 to a self-consistent method.\n\nIn short, the conceptual contribution is worth a conversation, but the implementation as written is not reproducible. I would not cite this in its current form. If the authors fix the sign, define a proper clipped or absolute-value power distance, spell out the filtering, and release code, the idea might be salvageable. But as submitted, I'd send it back with major revisions or reject.\n\nRecommendation: Send to peer review only if you expect the authors to fix these issues; otherwise desk reject. I'd keep an eye on a revision.","headline":"A fresh geometric take on TTA, but the algorithm as written has a sign inversion and an ill-defined exponent that make the headline results unreproducible.","tokens_in":23133,"tokens_out":5282,"would_cite":false,"duration_ms":47136,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that neighbor-based test-time adaptation is secretly a Voronoi diagram, and that upgrading it with cluster-induced influence and weighted cell boundaries yields the lowest errors and best calibration on four corruption…","keywords":["test-time adaptation","Voronoi diagram","cluster-induced Voronoi diagram","Power Diagram","distribution shift","entropy minimization","model calibration","noisy sample filtering"],"falsifier":"Run TTVD on CIFAR-10-C with $\\gamma$ swept over a grid (e.g., $-0.2$, $-0.5$, $-0.8$, $-1.5$) and with the CIVD influence function replaced by an ordinary distance average ($\\gamma = 1$), keeping all other settings fixed: if the error advantage over Tent and SAR concentrates at the single published value $\\gamma = -0.8$ and collapses at neighbouring values, or if the plain-distance version keeps the same gain, then the specific geometric influence term is not the cause of the reported improvement.","tokens_in":21973,"feed_emoji":"📐","tokens_out":8696,"duration_ms":76750,"temperature":0.7,"pith_summary":"Test-time adaptation fixes a model's drift when the test distribution shifts, but existing methods often rely on either self-supervision or entropy minimization, which can overfit or degrade over time. This paper argues that the neighbor-based family of methods is secretly a Voronoi diagram: the feature space is divided into cells around class prototypes, and adaptation pushes test features into the right cells. Building on that geometric reading, the paper constructs TTVD, which combines a cluster-induced Voronoi diagram (sites are clusters, so multiple rotations of a sample jointly vote) with a power diagram (cells carry weights, so noisy samples near boundaries can be filtered). Under the standardized TTAB evaluation, TTVD reports the lowest classification error and expected calibration error on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and ImageNet-R, and it keeps its edge at small batch sizes and under label shift. If the geometric reading is right, it gives a single lens for understanding and improving a broad class of adaptation algorithms.","feed_headline":"Voronoi geometry cuts test-time adaptation errors on four benchmarks","feed_subtitle":"Cluster-induced Voronoi cells plus power-diagram filtering beat entropy and self-supervision baselines on four corruption benchmarks.","key_machinery":"The load-bearing object is the Voronoi diagram used as an adaptation loss, not merely as a classifier. A feature $z$ receives soft label $\\beta(-d(z, \\mu_k)/\\tau)$ over the class sites $\\mu_k$, and the training signal is the entropy of that soft-label vector, so adaptation is equivalent to pulling $z$ into the cell of the correct prototype. The two refinements are defined by changing the distance: CIVD aggregates influences from a cluster of sites via $F(z, C_k) = -\\mathrm{sign}(\\gamma) \\sum_\\alpha d(\\mu_k^{(\\alpha)}, z)^\\gamma$, and the power diagram replaces $d^2$ with $d^2 - v_k^2$, adding a per-cell weight. The weighted boundaries of the power diagram are used for noisy-sample filtering by subtracting the PD partition from the VD partition, and because CIVD and PD are parallel structures, they compose into the cluster-induced power diagram (CIPD) that TTVD actually runs at test time.","core_discovery":"The paper's central claim is that any neighbor-based TTA method—one that compares test features to stored prototypes—is geometrically a Voronoi diagram, and that the diagram's two generalizations solve the two known failure modes of those methods. The first generalization, the cluster-induced Voronoi diagram (CIVD), replaces each single prototype by a cluster of prototypes obtained from self-supervised rotations, and defines the cell assignment by a joint influence function $F(z, C_k) = -\\mathrm{sign}(\\gamma) \\sum_\\alpha d(\\mu_k^{(\\alpha)}, z)^\\gamma$ with $\\gamma = -0.8$. This makes the soft label a pooled vote over all rotated views, which the paper shows corrects misclassifications that any single view would make. The second generalization, the power diagram (PD), gives every cell a weight so the boundary can shift; subtracting the PD from the plain Voronoi diagram isolates a region of high-gradient, unreliable samples, and those samples are excluded from adaptation. The paper reports that the combined method, TTVD, achieves the best error and expected calibration error in the TTAB benchmark on four datasets, and that the CIVD step contributes the largest single improvement (5.7% average error reduction on CIFAR-10-C) before the PD step adds another 2.2%.","pith_inferences":["The Voronoi reading suggests that any prototype-based TTA loss is a choice of influence kernel; replacing the power-law distance with a Gaussian kernel would produce a family of methods and could reveal whether the reported gains are specific to $\\gamma = -0.8$ or general to kernel-based cell assignment.","Since the power diagram is used only to filter samples, a natural variant would use the same weighted boundaries to down-weight rather than discard boundary samples, preserving gradient information from shifted but informative features.","The paper's geometric framing implies that adaptation methods can be compared by their cell geometry rather than only by accuracy; a testable extension is to predict which corruption types a method will fail on from the shape of its Voronoi cells in a 2D embedding.","The paper's own footnote warns that oracle-model-selection numbers overstate real-world performance; a fair replication should use the non-subscripted, label-free numbers when comparing to deployed settings."],"forward_implications":["A single Voronoi-style loss can carry both self-supervision and entropy minimization without tuning a trade-off weight, because the cluster-induced influence function combines the two objectives in one mechanism.","The power diagram's boundary subtraction filters unreliable samples without a separate entropy threshold, which the paper shows is error-prone on the entropy landscape.","TTVD keeps most of its advantage at batch sizes 32, 16, and 8 and under Dirichlet label shift with $\\alpha = 0.01$, meaning the method is usable in the non-i.i.d. online streams where TTA is normally deployed.","Adaptation curves on ImageNet-C noise show TTVD still descending at the last batch, while Tent and SAR plateau, which the paper interprets as a reduced tendency to overfit to the evolving test distribution.","The method is insensitive to the precision of the precomputed class means: using 10%, 5%, or 1% of ImageNet gives essentially identical error, so the offline site computation can be made cheap at scale."],"supporting_citations":[{"why":"Supplies the classical Voronoi diagram definition and properties that the paper identifies as the underlying structure of neighbor-based TTA methods.","marker":"[3]"},{"why":"Defines the power diagram (Laguerre–Voronoi) whose weighted cells are used for boundary shifting and noisy-sample filtering.","marker":"[2]"},{"why":"Defines the cluster-induced Voronoi diagram and the influence function in Eq. 4, the geometric core of CIVD.","marker":"[7,8,22]"},{"why":"Establishes that a logistic regression layer induces a power diagram, providing Lemma 3.1's link between deep networks and the geometric framework.","marker":"[5]"},{"why":"T3A, the neighbor-based TTA baseline whose prototype-adjustment behavior the Voronoi view is said to explain and which TTVD improves upon.","marker":"[24]"},{"why":"TAST, a nearest-neighbor self-training baseline that TTVD is compared against and outperforms in the reported experiments.","marker":"[25]"},{"why":"TTAB, the open benchmark whose evaluation settings the experiments adopt for fair comparison across methods.","marker":"[61]"},{"why":"Tent, the entropy-minimization baseline that TTVD is measured against both in tables and in adaptation curves.","marker":"[52]"}],"fun_headline_variants":["Voronoi geometry sharpens test-time adaptation on four benchmarks","Test-time adaptation gets a Voronoi geometry upgrade","Voronoi power diagram improves test-time adaptation","Geometric Voronoi view enhances test-time adaptation","TTVD: Voronoi cells refine test-time adaptation scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gains rest on the hand-chosen influence function $F(z, C_k) = -\\mathrm{sign}(\\gamma) \\sum_\\alpha d(\\mu_k^{(\\alpha)}, z)^\\gamma$ with $\\gamma = -0.8$ and temperature $\\tau = 1$; the paper gives no derivation of this choice, and if that soft-label signal is not a reliable training target across the full range of corruptions, the reported improvements would shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Voronoi geometry sharpens test-time adaptation on four benchmarks","Test-time adaptation gets a Voronoi geometry upgrade","Voronoi power diagram improves test-time adaptation","Geometric Voronoi view enhances test-time adaptation","TTVD: Voronoi cells refine test-time adaptation scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001118,"raw_usage":{"total_tokens":4752,"prompt_tokens":1145,"completion_tokens":3607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":761,"completion_tokens_details":{"reasoning_tokens":3528}},"tokens_in":761,"tokens_out":3607,"duration_ms":25446,"temperature":1.0,"reasoning_tokens":3528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:20:40.513412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TTVD on CIFAR-10-C with $\\gamma$ swept over a grid (e.g., $-0.2$, $-0.5$, $-0.8$, $-1.5$) and with the CIVD influence function replaced by an ordinary distance average ($\\gamma = 1$), keeping all other settings fixed: if the error advantage over Tent and SAR concentrates at the single published value $\\gamma = -0.8$ and collapses at neighbouring values, or if the plain-distance version keeps the same gain, then the specific geometric influence term is not the cause of the reported improvement.","supporting_citations":[{"cited_title":"V oronoi diagrams—a survey of a fundamental geometric data structure","cited_arxiv_id":null,"evidence_quote":"Supplies the classical Voronoi diagram definition and properties that the paper identifies as the underlying structure of neighbor-based TTA methods."},{"cited_title":"Aurenhammer","cited_arxiv_id":null,"evidence_quote":"Defines the power diagram (Laguerre–Voronoi) whose weighted cells are used for boundary shifting and noisy-sample filtering."},{"cited_title":"The geometry of deep networks: Power diagram subdivision","cited_arxiv_id":null,"evidence_quote":"Establishes that a logistic regression layer induces a power diagram, providing Lemma 3.1's link between deep networks and the geometric framework."},{"cited_title":"Test-time classifier adjustment module for model-agnostic domain generalization","cited_arxiv_id":null,"evidence_quote":"T3A, the neighbor-based TTA baseline whose prototype-adjustment behavior the Voronoi view is said to explain and which TTVD improves upon."},{"cited_title":"bike”, “bus","cited_arxiv_id":null,"evidence_quote":"TTAB, the open benchmark whose evaluation settings the experiments adopt for fair comparison across methods."},{"cited_title":"Tent: Fully test-time adaptation by entropy minimization","cited_arxiv_id":null,"evidence_quote":"Tent, the entropy-minimization baseline that TTVD is measured against both in tables and in adaptation curves."}],"review_version":1}