{"id":"9835c0b1-9b6d-447a-8ea9-ef99f5268c00","arxiv_id":"2508.21255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Randomly reweighting a dataset and then optimizing a set of support points to match the weighted data produces diverse, interpretable sample sets at low cost, according to visual results on MNIST and CelebA.","lead":"This paper proposes a new way to generate sample images by finding 'support points', representative points that summarize a dataset, but with random weights so each run gives a different set. The method claims to be an interpretable and cheap alternative to deep generative models like GANs and diffusion models, though it only shows visual examples, not quantitative scores.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Euclidean energy distance (Eq. 6) is an unvalidated proxy for visual quality; the acknowledged blur in §5.1.1 and absence of FID/IS means 'high-quality' is unsupported.","rationale":"The reader's weakest assumption—that Euclidean energy distance is a valid proxy for visual quality—is also the most load-bearing point. The paper's headline contribution is empirical ('high-quality and diverse outputs at a fraction of the cost'), and every experiment optimizes Eq. (12) under that metric. The authors themselves acknowledge blur, and no quantitative quality or timing benchmarks are provided. A missing convergence proof for Eq. (13) is real but secondary, since the empirical results would still be interpretable if the algorithm runs to tolerance. The concern is addressable by adding perceptual metrics and timing comparisons, so the reader's CONDITIONAL verdict remains appropriate rather than outright rejection. I agree with the reader's identification of the weak assumption, hence 'agree'.","tokens_in":15351,"tokens_out":11138,"duration_ms":108948,"concrete_test":"Generate 5,000 weighted support points on the MNIST 6/9 and CelebA-HQ subsets by running the released code across many independent weight realizations. Compute FID and Inception Score using standard implementations and compare against the paper's GAN and DDPM baselines on the same test set. If the support-point FID is substantially worse (e.g., not within 20% of the GAN baseline) or human raters significantly prefer the baselines, the Euclidean energy-distance objective does not produce 'high-quality' generative outputs and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that weighted support points are a viable generative alternative producing 'high-quality and diverse outputs'—depends on the energy distance (Eq. 6) computed with Euclidean distances in raw pixel space being a meaningful visual-quality measure. This is insecure: Euclidean pixel distance correlates poorly with perceptual similarity, penalizes translations, and favors pixel-wise averages, which yields exactly the 'slightly blurred' outputs the authors concede in §5.1.1. The objective (Eq. 12) minimizes sums of Euclidean norms, so optimized points can lie off the data manifold as convex combinations of images. No FID, Inception Score, or human evaluation is reported; the 'high-quality' claim rests only on subjective inspection. A secondary issue is that the fixed-point update (Eq. 13) has no convergence or local-optimality proof, so the 'principled' part of the claim is also not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative modeling framework based on weighted support points. It replaces the target distribution in the energy-distance minimization (Eq. 6) with a random finite measure obtained by random subsetting plus symmetric-Dirichlet weights calibrated by a target coefficient of variation (Section 4.3). The optimization is performed with a CCP/fixed-point update (Eq. 13) and efficient caching (Appendix A). The authors prove the random measure is centered at the empirical distribution (Appendix C) and present qualitative experiments on MNIST digits 6/9 and downscaled CelebA-HQ, claiming high-quality diverse outputs at a fraction of the cost of GANs/DDPMs.","tokens_in":15630,"tokens_out":12117,"duration_ms":112259,"significance":"If the empirical claims were substantiated, the method would be a useful non-neural alternative for representative-set generation, with explicit interpretability and low computational overhead. The theoretical centering result is clean and the implementation appears reproducible. However, the current evidence is qualitative only: no quantitative distributional metrics (FID, IS, etc.) or timing comparisons are reported, and the acknowledged blur in Section 5.1.1 raises doubts about Euclidean pixel-space energy distance as a perceptual-quality proxy. The convergence of the proposed fixed-point iteration is not established. These gaps make the central significance claim premature.","major_comments":[{"comment":"The claim that the method produces 'high-quality and diverse outputs at a fraction of the computational cost' is not supported by any quantitative evaluation. No FID, Inception Score, precision/recall, or other distributional metric is reported for MNIST or CelebA-HQ; no wall-clock timing table compares the method with the GAN and DDPM baselines; and the baselines themselves are small custom architectures with unstated training/inference costs. Figures 1–6 plus the authors' own statement that outputs are 'slightly blurred' (Section 5.1.1) do not establish the central claim. Please add standard quantitative metrics and a timing comparison, and report run-to-run diversity quantitatively.","section":"§5 and Abstract/Conclusion"},{"comment":"The objective is the energy distance with Euclidean distances in raw pixel space. This is never validated as a proxy for visual quality or distributional representativeness in high-dimensional image spaces. Because the energy distance penalizes Euclidean deviations, the optimized points can be convex combinations rather than manifold samples, which is consistent with the acknowledged blur in Section 5.1.1. At minimum, the paper should report FID or a task-based evaluation to show that the optimized sets are competitive; without this, 'high-quality' is an unsupported qualitative judgment.","section":"§3.1, Eq. (6), and §5.1.1"},{"comment":"No convergence guarantee or descent property is proven for the fixed-point update (13). Section 3.2.1 asserts that the CCP 'converges to a local minimum' for the unweighted version, but no proof is given, and the weighted/randomized case adds further complications. Since the 'principled' part of the contribution relies on the algorithm actually reaching a useful optimum, please provide a convergence analysis or, failing that, empirical convergence diagnostics (objective decrease, sensitivity to initialization) for the reported settings.","section":"§4.2.1, Eq. (13), Algorithm 2"},{"comment":"The method is described as a 'generative modeling' alternative to GANs/DDPMs, but it does not learn a model from which an arbitrary number of new samples can be drawn; each run solves a fresh optimization problem to produce a fixed set of representative points. This is closer to representative-set generation or distribution compression than to generative modeling as commonly understood. The comparison to GANs/DDPMs should be framed accordingly, otherwise the scope of the claim is misleading.","section":"Title/§1/§6"}],"minor_comments":[{"comment":"The method is called the Convex-Concave Procedure in the text but Algorithm 1 is titled 'Cyclic Convex Procedure.' Please align the terminology.","section":"§3.2.1 / Algorithm 1"},{"comment":"The text says Eq. (13) is obtained by 'setting the total gradient to zero.' More precisely, it is a CCP fixed-point update in which the denominators of the gradient are evaluated at the previous iterate. Please state this clearly to avoid confusion.","section":"§4.2.1"},{"comment":"The text first says a single run produced 25 weighted support points, then says 'each row corresponds to a different run' in reference to Figure 6. Please clarify whether Figure 6 shows one run or multiple runs.","section":"§5.2"},{"comment":"The sentence 'Figure 2 shows no discernible pattern' is unclear; presumably it means the 50 points are diverse. Please rephrase.","section":"§5.1"},{"comment":"The conclusion states the construction 'remaining centered at the empirical distribution (expression (14))'. This centering property holds for the random measure, not for the optimized support points themselves. Please make this distinction explicit to avoid over-reading.","section":"§4.3 and §6"},{"comment":"The target CV (0.4), the retention rule θ∼Unif(0.7,0.9), and the 0.6 lower bound are hand-chosen. Since diversity claims rest on these, a sensitivity analysis (e.g., CV ∈ {0.2,0.4,0.6}) would strengthen the paper.","section":"§4.3.2 / §5"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core is sensible and honestly presented, but the experimental section is too underpowered for the claimed contribution. I would be willing to reconsider after the authors add quantitative metrics (FID or similar), a timing table, and convergence diagnostics. The novelty is incremental relative to Mak and Joseph (2018), but the random-measure weighting is a reasonable extension. The paper may fit a specialized statistics/ML journal rather than a broad venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is better than the packaging. Putting a random subset of the data with symmetric Dirichlet weights into the support-point objective is a simple, principled way to buy run-to-run diversity that deterministic support points do not have. The gradient and fixed-point update in Eq. (13) are correct, the centering property in Appendix C is proven, and the CV calibration gives a transparent knob for dispersion. That is a real extension of Mak and Joseph, not just a rehash. The R code and reproducibility materials are also a plus.\n\nThe soft spot is exactly what the stress-test flags: the empirical section does not support the abstract's claims. There are no FID/IS numbers, no timing comparisons, and the CelebA-HQ experiment uses 144x144 images and takes roughly 3 hours for 25 points. For a method pitched as 'a fraction of the computational cost' of GANs and DDPMs, that runtime is not obviously a win once you include the optimization. The authors do admit in Section 5.1.1 that outputs are 'slightly blurred due to the use of distance-based optimization.' That is not a small caveat—it is evidence that Euclidean distance in pixel space is a weak proxy for visual quality, which is the load-bearing assumption of the whole approach.\n\nThere is also no convergence analysis for the fixed-point update. The paper says the CCP iteration is repeated until convergence, but does not prove that the update in Eq. (13) converges to a local minimum, or even converges at all. That matters because the word 'principled' in the claim depends on it. The math is plausible, but it is not demonstrated.\n\nI would not desk-reject this. The construction is worth refereeing, and the empirical gap is fixable. The authors need to add standard quantitative metrics on MNIST and a properly scaled CelebA benchmark, report wall-clock times against a simple GAN and DDPM, and either prove convergence of the fixed-point iteration or provide strong empirical evidence that it reliably stabilizes. If they do that, the paper could be a useful contribution for people who want an interpretable, non-neural method for diverse representative sets. As it stands, the central claim is overreaching by a comfortable margin.","headline":"A clean, genuinely new randomization of support points with sound math, but the empirical case for 'high-quality, low-cost' generation is carried by eye rather than numbers and needs work before the hype lands.","tokens_in":16039,"tokens_out":1540,"would_cite":true,"duration_ms":17969,"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 random weighted support points—a non-neural, energy-distance optimizer—can generate diverse images at a fraction of the cost of GANs or diffusion models.","keywords":["GANs","DDPMs","Energy Distance","Dirichlet Process","Truncated Measure","support points","generative modeling"],"falsifier":"Give the same algorithm two datasets that are far apart in raw pixel space but perceptually close (e.g., shifted digit images) and check whether the weighted support points blur across the shift or preserve the mode; a perceptual-distance variant that visibly improves outputs would falsify the Euclidean proxy. Separately, run the optimizer from many random initializations and compare final objective values: high variance would indicate the unproved convergence claim is doing real work.","tokens_in":15308,"feed_emoji":"🖼️","tokens_out":9090,"duration_ms":85250,"temperature":0.7,"pith_summary":"This paper sets out to prove that a generative model can be built without any neural network: minimize the energy distance between a random weighted measure over a dataset and a small set of candidate points, and the resulting 'weighted support points' form diverse, interpretable generated samples. The randomness comes from a symmetric Dirichlet weighting on a random subset of reference data, calibrated by a coefficient of variation so that runs differ without collapsing onto a few atoms. The authors derive a fixed-point optimization based on the convex-concave procedure and show on MNIST and CelebA-HQ that outputs compare in quality and diversity with lightweight GANs and diffusion models while costing far less to train and sample. If correct, this means representative, structure-preserving generation is accessible to a transparent optimization routine rather than a black-box network.","feed_headline":"Weighted support points generate images without neural networks","feed_subtitle":"A Dirichlet-style randomness scheme turns data summarization into a cheap, interpretable generator.","key_machinery":"The machinery is the weighted empirical energy distance: a sum of weighted attractions from candidate points toward reference atoms minus a pairwise repulsion among candidates, with the random measure built by uniform subsampling plus symmetric Dirichlet weights whose concentration is calibrated through the coefficient-of-variation identity CV = sqrt((N0-1)/(kappa+1)). This objective is minimized by a convex-concave fixed-point update that alternates attraction toward weighted reference atoms and repulsion among support points. The randomness of the measure is what converts a deterministic summarization method into a generative sampler: each draw of subset and weights produces a different op","core_discovery":"The central claim is that randomizing the target measure in the support-point problem turns a data-compression tool into a generative model. The paper defines weighted support points as minimizers of a weighted empirical energy distance between a random measure F_N0—built by subsampling the dataset and assigning exchangeable symmetric-Dirichlet weights—and an n-point empirical candidate measure. Because the random measure is centered at the empirical distribution, every realization represents the original data; because the weights are dispersed in a controlled way, independent runs yield different configurations. The authors supply a CV-calibrated weighting scheme and an efficient convex-con","pith_inferences":["Beyond the paper: replacing the raw Euclidean pixel distance with a perceptual or learned metric inside the same objective is a natural, testable upgrade; if it works, the method could produce sharper outputs without changing the algorithm.","The interpolative behavior the authors highlight suggests the sweet spot is representative summarization—prototypes, coresets, balanced augmentation—rather than photorealistic novelty; a reader should expect structure preservation more than crisply invented detail.","Because convergence of the fixed-point update is asserted empirically, not proven, run-to-run comparisons of objective values from different random initializations could diagnose whether reported diversity comes from the random measure or from sensitivity to initialization.","The CV-calibrated compromise between Bayesian bootstrap and Dirichlet-process draws is itself a reusable statistical construction; finite-sample bounds connecting CV to the diversity of the resulting configurations would sharpen when to choose which regime."],"forward_implications":["If correct, a generative model can be trained by one convex-concave optimization instead of an adversarial game or a learned denoising chain, removing mode collapse and training instability from the failure modes.","Because generation requires only solving small weighted energy-distance problems, producing new sample sets is cheap; the paper reports about one minute per run for ten MNIST points and about three hours for twenty-five CelebA-HQ points on CPU.","The centeredness result—that the random measure is unbiased for the empirical distribution—gives a formal sense in which the generated configurations remain faithful to the data while diversity is explicitly user-tunable via the target CV.","Every generated point is a vector in data space, so outputs can be inspected, traced to the subset and weights that produced them, and reused directly as integration nodes or prototypes.","The distance-based formulation applies to any vectorized dataset without architecture design, making the method a drop-in generation scheme for moderate-dimensional structured data."],"supporting_citations":[{"why":"Defines support points and the convex-concave algorithm that the paper's weighted variant extends.","marker":"Mak and Joseph (2018)"},{"why":"Supplies the energy distance that serves as the objective for both unweighted and weighted support points.","marker":"Székely and Rizzo (2013)"},{"why":"The Bayesian bootstrap weighting scheme is one of the random-measure constructions the paper generalizes.","marker":"Rubin (1981)"},{"why":"Introduces the Dirichlet process whose finite-dimensional symmetric-Dirichlet weights inspire the random measure.","marker":"Ferguson (1973)"},{"why":"Provides the stick-breaking construction used to describe truncated Dirichlet-process weights.","marker":"Sethuraman (1994)"},{"why":"Defines the GAN baseline whose computational cost and training instability motivate the alternative.","marker":"Goodfellow et al. (2014)"},{"why":"Defines the DDPM baseline against which the method's cost and output quality are compared.","marker":"Ho et al. (2020)"}],"fun_headline_variants":["Randomized support points: generative model without neural nets","Weighted support points make cheap, interpretable image generators","Dirichlet-weighted support points: a new path to generative AI","Generative modeling via random support points, no deep learning required","Data summarization reimagined: random support points for generation"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that Euclidean distance between raw pixel vectors tracks visual similarity well enough that minimizing it yields representative, high-quality images—with the unproved secondary premise that the fixed-point iteration lands in a useful local optimum.","fun_headline_variants_meta":{"raw":{"variants":["Randomized support points: generative model without neural nets","Weighted support points make cheap, interpretable image generators","Dirichlet-weighted support points: a new path to generative AI","Generative modeling via random support points, no deep learning required","Data summarization reimagined: random support points for generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":996,"prompt_tokens":716,"completion_tokens":280,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":460,"tokens_out":280,"duration_ms":3189,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:27:06.177799+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the same algorithm two datasets that are far apart in raw pixel space but perceptually close (e.g., shifted digit images) and check whether the weighted support points blur across the shift or preserve the mode; a perceptual-distance variant that visibly improves outputs would falsify the Euclidean proxy. Separately, run the optimizer from many random initializations and compare final objective values: high variance would indicate the unproved convergence claim is doing real work.","supporting_citations":[{"cited_title":"and Joseph, V","cited_arxiv_id":null,"evidence_quote":"Defines support points and the convex-concave algorithm that the paper's weighted variant extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Bayesian bootstrap weighting scheme is one of the random-measure constructions the paper generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Dirichlet process whose finite-dimensional symmetric-Dirichlet weights inspire the random measure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the stick-breaking construction used to describe truncated Dirichlet-process weights."},{"cited_title":"J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y","cited_arxiv_id":null,"evidence_quote":"Defines the GAN baseline whose computational cost and training instability motivate the alternative."}],"review_version":1}